Connect coding agents, AI SDKs, and LLM tools to the AntSeed buyer proxy. Use when configuring Claude Code, Codex, OpenCode, Pi, OpenClaw, Hermes, GenLayer Studio, Vercel AI SDK, LangChain, or raw HTTP to route inference through AntSeed at localhost:8377.
Connect coding agents, AI SDKs, and LLM tools to the AntSeed buyer proxy. Use when configuring Claude Code, Codex, OpenCode, Pi, OpenClaw, Hermes, GenLayer Studio, Vercel AI SDK, LangChain, or raw HTTP to route inference through AntSeed at localhost:8377.
AntSeed — Integration Skill
This file is the agent-readable companion to https://antseed.com/integrations.
It tells any AI agent (Claude, Codex, OpenClaw, Hermes, custom) exactly
how to wire its tool of choice up to the AntSeed peer-to-peer inference network.
What is AntSeed?
AntSeed is a peer-to-peer marketplace for AI inference. Buyers run a small
local daemon (the buyer proxy) that exposes an HTTP API at
http://localhost:8377 speaking the three caller-facing LLM API protocols:
Anthropic Messages, OpenAI Chat Completions, and OpenAI Responses. Legacy
OpenAI Completions is supported internally for adapter translation. The proxy
discovers providers on a DHT, routes the
request to a peer, translates between protocols when needed (via
@antseed/api-adapter), and settles in USDC on Base.
Important: AntSeed is for value-added AI services (specialized models, agents,
TEEs, fine-tunes, managed workflows), not raw resale of API keys or subscription
access. Providers must comply with upstream terms of service.
From the perspective of any tool, SDK, or agent, AntSeed is just a local
OpenAI/Anthropic-compatible endpoint — point a base_url at it and you are done.
Glossary (mental model)
Buyer proxy — the local server on localhost:8377 that accepts API calls
from your tools and forwards them to AntSeed peers. It is the only thing your
editor / agent / SDK ever talks to.
Peer — someone selling inference. Each peer has a peerId (40-char hex),
a display name, and a list of services. List with .
antseed network browse
Service — a single model id like claude-sonnet-4-6 or deepseek-v4-flash.
This is what you pass as model in your tool's config. Each service has its
own native protocol list and its own in / cachedIn / out pricing.
Protocols (per service) — the wire formats a service accepts natively,
advertised on each peer in providerServiceApiProtocols[provider].services[service].
Values are anthropic-messages,
openai-chat-completions, openai-responses, openai-completions. This is
the field to match your tool's wire format against. If your tool's wire format
is in this list, the request passes through untouched; if not, the api-adapter
translates on the fly.
Cached input pricing — services charge a separate, much lower rate
(typically 4–10×) for tokens that are reused across requests: system prompts,
tool schemas, prior conversation turns, long files you keep referencing. The CLI
exposes it as cachedInputUsdPerMillion. For long-running agents and chatbots,
this is often the dominant cost line.
Pin — telling the buyer proxy "route requests to this peer." In the
default manual flow, there is no peer auto-selection; you must choose a peer,
send a per-request pin header, or start the proxy with a router plugin that
performs selection. Common explicit routes:
Session pin: antseed buyer connection set --peer <peerId>. Persists in
~/.antseed/buyer.state.json and applies to every request until you change it.
Per-request header: x-antseed-pin-peer: <peerId> on each call. Overrides
the session pin for that request, and works without any session pin at all.
Model prefix: set model to <peerId>@<service>. The proxy uses the
prefix as the peer pin and forwards only <service> to the seller.
If both header and model-prefix pins are present, the header selects the peer;
the model prefix is still stripped before routing.
Until at least one of these is in effect, every request returns no_peer_pinned.
Universal setup (do this once)
Option A — AntStation desktop app (easiest)
Download from https://antseed.com — it ships the buyer proxy, a wallet, and a
peer browser in a GUI. While the app is open the proxy is reachable at
http://localhost:8377.
Option B — CLI (headless / servers / agents)
# 1. Install
npm install -g @antseed/cli
# 2. Identity (an EVM private key — 64 hex chars). Save this somewhere safe;# you will reuse it across machines and it controls your USDC deposits.export ANTSEED_IDENTITY_HEX=$(openssl rand -hex 32)
# SECURITY: never paste this key into chat, logs, GitHub issues, or a file# committed to git. It controls the buyer identity and access to deposits.# 3. Start the buyer proxy on :8377
antseed buyer start &
# 4. Browse the network and list every service (= model) each peer offers,# along with its native protocols and USD-per-1M-tokens pricing.# `service` is the model id you pass to your tool. `protocols` is the# wire format(s) the service accepts natively — match it against your# tool. `in` / `cachedIn` / `out` are fresh-input / cached-input / output.
antseed network browse --json --top 5 \
| jq '.peers | map({
peerId, name: .displayName,
services: [
(.providerServiceApiProtocols | to_entries[]) as $p
| ($p.value.services | to_entries[]) as $s
| {
service: $s.key,
protocols: $s.value,
in: (.providerPricing[$p.key].services[$s.key].inputUsdPerMillion // .providerPricing[$p.key].defaults.inputUsdPerMillion),
cachedIn: (.providerPricing[$p.key].services[$s.key].cachedInputUsdPerMillion // null),
out: (.providerPricing[$p.key].services[$s.key].outputUsdPerMillion // .providerPricing[$p.key].defaults.outputUsdPerMillion)
}
]
})'# 5. Inspect one peer in detail. Use `matchingServices[]` for pricing/tags and# `peer.providerServiceApiProtocols` for native protocol support.# `cachedIn` is typically 4–10× cheaper than `in` and often dominates the# cost line for long-running agents and chatbots — always include it when# comparing peers.
antseed network peer <peerId> --json \
| jq '{
peer: (.peer | { peerId, name: .displayName,
sessions: .onChainChannelCount,
ghosts: .onChainGhostCount }),
services: [
(.peer.providerServiceApiProtocols | to_entries[]) as $p
| ($p.value.services | to_entries[]) as $s
| (.matchingServices[] | select(.provider == $p.key and .service == $s.key)) as $m
| {
provider: $p.key,
service: $s.key,
protocols: $s.value,
in: $m.inputUsdPerMillion,
cachedIn: $m.cachedInputUsdPerMillion,
out: $m.outputUsdPerMillion,
tags: $m.tags
}
]
}'# 6. Pin a peer (session-wide). Until you do, every request returns# `no_peer_pinned` UNLESS the request includes an `x-antseed-pin-peer`# header (see Per-request peer selection below).
antseed buyer connection set --peer <peerId>
# 7. Verify the proxy advertises the services you expect
curl -s http://localhost:8377/v1/models | jq '.data[].id'# 8. (Optional) Deposit USDC on Base for paid services
antseed payments # opens portal at 127.0.0.1:3118?token=<hex> — connect a wallet, deposit USDC
Security notes for agents and deploys
Treat ANTSEED_IDENTITY_HEX / ~/.antseed/identity.key as a hot wallet key.
Never print it, paste it into chat, commit it, or copy it off the buyer host.
Keep the buyer proxy bound to 127.0.0.1 / localhost. Do not expose
:8377 directly to the public internet; use SSH tunnels or a private network
if another process must reach it remotely.
Start with small USDC deposits and conservative reserve caps for autonomous
agents. The funding wallet does not need to stay connected after depositing.
If a tool requires an API key, use a non-secret placeholder such as antseed;
the buyer proxy authenticates with the local identity key instead.
Endpoints exposed by the buyer proxy
Path
Wire format
Common callers
POST /v1/messages
Anthropic Messages
Claude Code, Anthropic SDKs, OpenClaw
POST /v1/chat/completions
OpenAI Chat Completions
Codex, Hermes, OpenAI SDKs, Vercel AI SDK, LangChain, most tools
POST /v1/responses
OpenAI Responses
Codex (newer builds), tools using the Responses API
All four protocols (including legacy openai-completions) are supported by
@antseed/api-adapter for translation, but only the three endpoints above are
exposed to callers. Translation is automatic: a request that arrives in one
format and is routed to a peer whose service advertises a different protocols
value is transformed both directions (request and streaming response).
No Authorization header is required by the buyer proxy. It authenticates and
pays peers using the local node's identity key and on-chain USDC deposits.
Per-request peer selection (no session pin needed)
Two ways to tell the proxy which peer to use:
Session pin — antseed buyer connection set --peer <peerId>. Persists in
~/.antseed/buyer.state.json (pinnedPeerId) and applies to every request
until you change it. Best for single-tenant setups (laptops, dedicated agents).
Per-request header — send x-antseed-pin-peer: <peerId> on each call.
Overrides the session pin for that one request. You do not need to call
antseed buyer connection set at all if every request includes this header
— the proxy will accept and route them. Best for scripts, schedulers, and
multi-tenant deployments that need to fan out to different peers per call.
x-antseed-provider: <providerName> — when a peer exposes the same service
through more than one seller-plugin (rare), force a specific one. Most tools
never need this.
Files the CLI creates in ~/.antseed/
Knowing what lives in ~/.antseed/ matters for backups, container deploys,
and debugging. The directory is created on first antseed buyer start (or
first run of any antseed command that needs it).
Path
Purpose
Survives restart?
Safe to delete?
identity.key
Raw 32-byte EVM private key for the buyer wallet. Fallback when ANTSEED_IDENTITY_HEX is not set.
yes
NO — deleting loses access to your USDC deposits. Back this up.
identity.enc
Encrypted copy of identity.key (when the desktop app sets a passphrase).
Live runtime state: pinnedPeerId, the discovered-peers cache (discoveredPeers), on-chain stats, the proxy pid and port. Re-built from the network on next start.
yes (the pin survives restart)
yes (you lose the pin and the cached peer list — next browse will repopulate)
metering.db
SQLite log of every request the proxy served (model, peer, tokens, USDC). Used by antseed buyer status and the payments portal.
yes
yes (you lose request history; settlement is unaffected)
payments/
Per-channel state used by the seller-side settlement flow (only relevant if you also run antseed seller).
yes
only if you do not run a seller
plugins/
Cache of downloaded provider plugins.
yes
yes (re-downloaded on next use)
chat/, projects/
Used by the desktop app for local chat history. Empty on a CLI-only setup.
yes
yes
config.json (commonly edited)
{"buyer":{"proxyPort":8377,// change if 8377 conflicts on the host"minPeerReputation":0,// optional: raise to filter lower-reputation peers"maxPricing":{// refuse to route to peers above this rate"defaults":{"inputUsdPerMillion":100,"outputUsdPerMillion":100}}},"payments":{"preferredMethod":"crypto","crypto":{"chainId":"base-mainnet"}// or "base-sepolia" for testnet},"network":{"bootstrapNodes":[]}// empty = use built-in defaults}
Edit, then restart the buyer proxy. Do not hardcode contract addresses —
the chain preset (base-mainnet / base-sepolia) resolves Deposits, Channels,
USDC, and the RPC URL automatically.
buyer.state.json (read-only in practice)
Top-level fields that an agent might want to inspect:
pinnedPeerId — the currently pinned peer (or null).
pid / port — the running proxy. If pid is non-null but the process is gone, antseed buyer start will detect the stale lockfile and clean up.
discoveredPeers — cached peer list from the last DHT browse. Refreshed on antseed network browse.
Below is every integration we currently document. Find your tool, copy the
config block, done. If your tool is not listed but accepts a custom OpenAI or
Anthropic base URL, follow the closest example (or look at the curl / Raw
HTTP entry — the contract is stable).
Claude Code
Anthropic's official CLI agent — launch through AntSeed with antseed claude.
TL;DR for agents: Prefer antseed claude --model <service-id>. It sets ANTHROPIC_BASE_URL and ANTHROPIC_API_KEY for Claude Code. Manual equivalent: set ANTHROPIC_BASE_URL=http://localhost:8377 and ANTHROPIC_API_KEY=antseed, then run claude --model <service-id>.
Claude Code is the official CLI coding agent from Anthropic. It speaks the Anthropic Messages API natively, so it slots into AntSeed through the antseed claude wrapper or by pointing ANTHROPIC_BASE_URL at your local proxy.
antseed claude resolves the active buyer proxy, sets the placeholder Anthropic API key for the child process, and forwards the rest of your Claude Code flags unchanged. Manual environment variables still work if you want to run claude directly.
No real Anthropic API key is needed — the AntSeed proxy authenticates each request with your local identity (ANTSEED_IDENTITY_HEX) and settles payments on-chain. The ANTHROPIC_API_KEY value is required by the Anthropic SDK only as a non-empty placeholder.
When Claude Code calls the Messages API, the proxy forwards the request to the peer you pinned in step 3 of the setup above. Whichever service ids that peer advertises (visible in antseed network peer <peerId>) become the valid --model values.
Install
Install Claude Code globally
npm install -g @anthropic-ai/claude-code
Verify it runs
claude --version
Example output:
1.4.2 (Claude Code)
Configure
antseed claude --model claude-sonnet-4-6
Recommended: the wrapper reads the active buyer proxy from buyer.state.json or config, sets ANTHROPIC_BASE_URL and ANTHROPIC_API_KEY for Claude Code, and forwards extra Claude args. Add --antseed-base-url http://host:port only when your proxy is somewhere else.
antseed claude --model <service-id> passes the value to Claude Code unchanged. The valid set is whatever your pinned peer advertises — see the discovery commands below.
These are the only ids that work with --model. To switch peers, run antseed network browse, then antseed buyer connection set --peer <peerId> and re-check this list.
Start a Claude Code session through the wrapper
antseed claude --model claude-sonnet-4-6
Manual equivalent after exporting the env vars above: claude --model claude-sonnet-4-6.
Troubleshooting
"invalid x-api-key" or 401 from Anthropic SDK — antseed claude sets ANTHROPIC_API_KEY=antseed for you. If you run claude directly, set the variable to any non-empty string; the proxy ignores the value.
Hangs forever on first message — No peer is pinned. Run antseed network browse to see peers, then antseed buyer connection set --peer <peerId>.
model_not_found for a model name you expected to work — The pinned peer doesn't advertise that service id. Check what it does offer with antseed network peer <peerId> (or curl http://localhost:8377/v1/models). Pin a different peer if needed.
Want to confirm a request actually went through AntSeed (not Anthropic direct) — After the request completes, run antseed buyer metering — you'll see the channel for the peer Claude Code routed to, with token counts and the USDC settled. antseed buyer status shows the snapshot (pinned peer, active-channel count, deposits).
How Claude Code talks to AntSeed
Wire format sent by Claude Code: Anthropic Messages (hits /v1/messages on the buyer proxy).
Best-fit services: any service whose protocols array contains anthropic-messages — that is what the peer advertises as natively-supported, so traffic passes through with zero translation overhead.
How to check a peer: run antseed network peer <peerId> --json and look at peer.providerServiceApiProtocols[provider].services[service] for each model. The browse command exposes the same field per peer.
When protocols differ: AntSeed's @antseed/api-adapter translates between Anthropic Messages and the service's native protocol on the fly. So a request from Claude Code can still reach a service that only advertises a different protocol — just with a small transform step.
Caveat: services whose only advertised protocol is openai-responses require streaming. If Claude Code sends a non-streaming request and the proxy routes it to one of those services, the call fails with HTTP 400: Stream must be set to true. Pick a service whose protocols includes anthropic-messages (or another non-responses protocol) to avoid this.
TL;DR for agents: Prefer antseed codex --model <service-id>. It injects the AntSeed Codex provider for one run using base_url=http://localhost:8377/v1 and wire_api="responses". Manual alternative: create user-level ~/.codex/antseed.config.toml with top-level model/model_provider plus [model_providers.antseed], then run codex --profile antseed.
Codex is OpenAI's terminal coding agent. Recent versions ignore OPENAI_BASE_URL and instead read provider config from Codex settings.
antseed codex supplies that provider config for one run with Codex -c overrides, points it at the active buyer proxy, sets the placeholder API key, and leaves your real CODEX_HOME untouched.
If you prefer a persistent manual setup, create ~/.codex/antseed.config.toml and launch Codex with codex --profile antseed; the wrapper is still the shortest path for one-off sessions.
Install
Install Codex globally
npm install -g @openai/codex
Verify it runs
codex --version
Configure
antseed codex --model claude-sonnet-4-6
Recommended: the wrapper resolves the proxy URL, injects an AntSeed model provider with wire_api = "responses", sets ANTSEED_API_KEY=antseed, and forwards extra Codex args. Put child flags after -- when they look like wrapper flags.
# Loaded by: codex --profile antseed# Set this to a service id returned by http://localhost:8377/v1/models# after pinning an AntSeed peer.model = "claude-sonnet-4-6"model_provider = "antseed"[model_providers.antseed]name = "AntSeed"base_url = "http://localhost:8377/v1"wire_api = "responses"
Manual profile only: this must be your user-level~/.codex/antseed.config.toml, then launch with codex --profile antseed. If your buyer proxy uses a non-default port, update base_url to match it. Project-local ./.codex/config.toml provider blocks are ignored by Codex.
GUI:
No real OpenAI key is needed. The AntSeed proxy authenticates with your local buyer identity; the wrapper and manual profile both point Codex at the local proxy instead of OpenAI.
Pass the peer service id to antseed codex --model <service-id>. For a manual profile, set top-level model = "<service-id>" in ~/.codex/antseed.config.toml or override with codex --profile antseed --model <service-id>.
The buyer dashboard at http://localhost:3118 is the authoritative real-time signal: a non-zero Reserved (channel opened) and/or a drop in Available (settled spend) after a real prompt confirms AntSeed served the request. The antseed buyer status CLI output is cached and may lag the dashboard — refresh the web view for confirmation. Do not rely on lsof -i | grep codex or ~/.codex/log/codex-tui.log: Codex keeps persistent TCP connections to Cloudflare/ChatGPT IPs (e.g. 172.64.0.0/13) for non-inference purposes (the cause was not isolated during testing), and the provider=OpenAI lines in the TUI log are not a reliable indicator that inference went to OpenAI — the on-chain numbers can show AntSeed served the request despite that log line.
Troubleshooting
OPENAI_BASE_URL / OPENAI_API_KEY are being ignored — Expected on recent Codex builds. Use antseed codex --model <service-id> so the wrapper injects the provider config for the current run, or use the manual ~/.codex/antseed.config.toml profile above.
How can I tell if Codex is actually routing through AntSeed? — Check the buyer dashboard at http://localhost:3118 (or antseed buyer status) after sending a test prompt. Reserved going from $0 to a non-zero value (a channel was opened) and/or Available dropping (spend settled) confirms AntSeed served the request. If both stay flat after a real prompt, the profile is not being applied. Do not trust lsof connections to Cloudflare IPs or provider=OpenAI lines in ~/.codex/log/codex-tui.log — neither is a reliable routing signal.
Codex prints Ignored unsupported project-local config keys … model_provider, model_providers — Provider settings must live in your user-level Codex profile file. For this manual flow, put the top-level model, model_provider, and [model_providers.antseed] block in ~/.codex/antseed.config.toml, then relaunch with codex --profile antseed. Codex silently rejects provider blocks in project-local ./.codex/config.toml and falls back to its default provider.
Hand-written Codex -c provider overrides behave inconsistently — Use antseed codex --model <service-id> so AntSeed supplies the complete provider block (base_url, wire_api, and model_provider) for the current run. If managing config yourself, keep the full provider/profile in user-level ~/.codex/antseed.config.toml.
Streaming stops after the first chunk with a manual profile — Use antseed codex, or set wire_api = "responses" in the manual [model_providers.antseed] block.
unknown profile: antseed — Codex caches profile config on launch. Make sure you saved ~/.codex/antseed.config.toml, then start a fresh codex --profile antseed session.
Hangs forever on first message — No peer is pinned. Run antseed network browse, then antseed buyer connection set --peer <peerId>.
How OpenAI Codex CLI talks to AntSeed
Wire format sent by OpenAI Codex CLI: OpenAI Chat Completions (hits /v1/chat/completions on the buyer proxy).
Best-fit services: any service whose protocols array contains openai-chat-completions — that is what the peer advertises as natively-supported, so traffic passes through with zero translation overhead.
How to check a peer: run antseed network peer <peerId> --json and look at peer.providerServiceApiProtocols[provider].services[service] for each model. The browse command exposes the same field per peer.
When protocols differ: AntSeed's @antseed/api-adapter translates between OpenAI Chat Completions and the service's native protocol on the fly. So a request from OpenAI Codex CLI can still reach a service that only advertises a different protocol — just with a small transform step.
Caveat: services whose only advertised protocol is openai-responses require streaming. If OpenAI Codex CLI sends a non-streaming request and the proxy routes it to one of those services, the call fails with HTTP 400: Stream must be set to true. Pick a service whose protocols includes openai-chat-completions (or another non-responses protocol) to avoid this.
TL;DR for agents: Prefer antseed opencode --model <service-id>. It creates a temporary OpenCode provider config using npm="@ai-sdk/openai-compatible", baseURL="http://localhost:8377/v1", apiKey="antseed", and one model entry. Manual alternative: put the same provider in opencode.json and run opencode.
OpenCode is an MIT-licensed terminal coding agent built on the Vercel AI SDK. It supports 75+ providers out of the box and lets you register custom ones via opencode.json.
antseed opencode creates that custom provider config in a temporary opencode.json, points OpenCode at it for the child process, and deletes it when the session exits. Manual project or global config still works if you want OpenCode to remember AntSeed outside the wrapper.
AntSeed plugs in as a custom provider using the @ai-sdk/openai-compatible adapter — the same one OpenCode recommends for any OpenAI-compatible endpoint (LM Studio, llama.cpp, Atomic Chat, etc.). No ANTHROPIC_BASE_URL: OpenCode reads provider config from JSON.
Each model you want to use must be listed under models. The id has to match what the buyer proxy returns from GET /v1/models — i.e. a service id advertised by your currently-pinned peer.
Install
Install OpenCode
npm install -g opencode-ai
Verify it runs
opencode --version
Configure
antseed opencode --model gpt-oss-120b
Recommended: the wrapper resolves the proxy URL, writes a temporary OpenCode config with one AntSeed model, sets OPENCODE_CONFIG for the child process, and forwards extra OpenCode args.
antseed opencode --model <service-id> generates a temporary config for that one id. In manual config, the keys under models must exactly match service ids returned by curl http://localhost:8377/v1/models.
Test it
Confirm the proxy lists the same ids your config references
Add or remove entries under models in opencode.json so they match this list.
Launch OpenCode through the wrapper
antseed opencode --model gpt-oss-120b
Extra OpenCode args are forwarded, so antseed opencode --model gpt-oss-120b run works too. Manual config equivalent: run opencode, then pick one of the AntSeed entries from /models.
Troubleshooting
AntSeed doesn't appear in /connect or /models — With antseed opencode, pass the service id via --model; the wrapper supplies a temporary config. With manual config, make sure opencode.json is in your project root (or ~/.config/opencode/opencode.json) and that the JSON is valid — a stray comma silently disables the whole provider.
Model is listed but every call returns model_not_found — The pinned peer doesn't advertise that service id. Run antseed network peer <peerId> to see what it actually offers, or pin a different peer.
OpenCode prompts for an API key — The proxy ignores auth, but the AI SDK sometimes asks anyway. Either skip the prompt (press enter on empty input) or set "apiKey": "antseed" inside options in opencode.json.
How OpenCode talks to AntSeed
Wire format sent by OpenCode: OpenAI Chat Completions (hits /v1/chat/completions on the buyer proxy).
Best-fit services: any service whose protocols array contains openai-chat-completions — that is what the peer advertises as natively-supported, so traffic passes through with zero translation overhead.
How to check a peer: run antseed network peer <peerId> --json and look at peer.providerServiceApiProtocols[provider].services[service] for each model. The browse command exposes the same field per peer.
When protocols differ: AntSeed's @antseed/api-adapter translates between OpenAI Chat Completions and the service's native protocol on the fly. So a request from OpenCode can still reach a service that only advertises a different protocol — just with a small transform step.
Caveat: services whose only advertised protocol is openai-responses require streaming. If OpenCode sends a non-streaming request and the proxy routes it to one of those services, the call fails with HTTP 400: Stream must be set to true. Pick a service whose protocols includes openai-chat-completions (or another non-responses protocol) to avoid this.
TL;DR for agents: Install Pi: npm install -g @mariozechner/pi-coding-agent. Install the AntSeed extension: pi install git:github.com/AntSeed/pi-antseed. Restart or /reload. The extension calls pi.registerProvider("antseed", { api: "openai-responses", baseUrl: "http://localhost:8377/v1" }) and auto-discovers models from the pinned peer via GET /v1/models. Switch with /model antseed/<service-id>. Override base URL with ANTSEED_BASE_URL env var; auth with ANTSEED_API_KEY.
What Pi is. Pi (@mariozechner/pi-coding-agent) is a minimal, hackable terminal coding agent by Mario Zechner — the same lineage as pi-mono. It ships with four default tools (read, write, edit, bash) and lets you extend everything else — commands, providers, themes, even the editor UI — through TypeScript extensions, skills, and prompt templates. No fork required.
What the AntSeed extension does.pi-antseed is a Pi extension that registers the local buyer proxy as a Pi provider named antseed. Once installed, every service your pinned peer advertises shows up under antseed/<id> in Pi's model picker (Ctrl+L or /model) — you switch with /model antseed/minimax-m2.7 just like any built-in.
Why an extension instead of env vars. Pi already speaks dozens of provider protocols natively. The extension calls pi.registerProvider("antseed", { api: "openai-responses", authHeader: true, baseUrl: "http://localhost:8377/v1" }) — Pi then handles auth headers, streaming, retries, and tool-calling. The Responses API path preserves reasoning items across turns for reasoning-capable models, while the extension still auto-refreshes the model list from GET /v1/models so the menu reflects what your pinned peer can serve.
Install
Install Pi itself (the coding agent CLI)
npm install -g @mariozechner/pi-coding-agent
Pi requires Node.js 20+. The binary is pi. Verify with pi --version. Without any extensions, Pi can already talk to Claude / GPT / Gemini / Groq / etc. via API key or OAuth — the AntSeed extension below is what teaches it to route through your local buyer proxy.
Install the AntSeed extension into Pi
pi install git:github.com/AntSeed/pi-antseed
Pi extensions install from a git URL or a local path. Alternatives: pi -e git:github.com/AntSeed/pi-antseed runs the extension once without installing, useful for trying it out. pi install ./pi-antseed works from a local clone.
Reload Pi so the new provider is picked up
/reload
Run this inside the Pi REPL (after typing pi to launch it). It re-scans extensions, skills, prompt templates, keybindings, and context files. A full restart works too.
No GUI config needed in the common case — the extension reads ANTSEED_BASE_URL (default http://localhost:8377/v1) and discovers models from the pinned peer automatically. Only set ANTSEED_API_KEY if you front the buyer proxy with your own auth layer, or ANTSEED_MODELS="id1,id2" to skip discovery and register a fixed list.
The extension auto-discovers from GET /v1/models after Pi loads, so anything your pinned peer advertises shows up under antseed/.... After re-pinning a different peer, run /reload in Pi to refresh the model list.
Test it
Launch Pi
pi
You'll see Pi's startup header, which lists loaded extensions. Look for antseed (or pi-antseed) in that list — if it's there, the extension loaded successfully.
Open the model picker and pick an AntSeed-routed model
/model
Or press Ctrl+L. The picker is fuzzy-searchable; type "antseed" to filter. You should see entries like antseed/claude-sonnet-4-6, antseed/deepseek-v4-flash, etc. — one for each service your pinned peer advertises.
Or switch directly via slash command
/model antseed/minimax-m2.7
Replace minimax-m2.7 with any id from curl http://localhost:8377/v1/models. After this, every prompt routes through AntSeed → your pinned peer → the model.
Troubleshooting
pi: command not found after install — Your global npm bin is not on PATH. Run npm prefix -g to find it, then add <that-path>/bin to PATH in your shell rc. Or use a Node version manager (nvm, fnm, volta) which handles this automatically.
antseed doesn't appear in the model picker (/model or Ctrl+L) — The extension didn't load. Re-run pi install git:github.com/AntSeed/pi-antseed, restart Pi, and watch the startup header — it lists every loaded extension and surfaces load errors there.
Picker only shows a few hard-coded antseed/... ids, not what my peer offers — Pi started before the buyer proxy was up, so the extension fell back to its built-in seed list. Make sure antseed buyer start is running and a peer is pinned, then run /reload inside Pi to refresh the model list.
Empty /v1/models from the proxy — No peer is connected. Run antseed network browse to see options, then antseed buyer connection set --peer <peerId>. Or launch the proxy with antseed buyer start --router <name> for automatic peer selection.
5xx from the proxy mid-conversation — Usually means the pinned peer doesn't offer the model you asked for, or has just gone offline. Re-pin via antseed buyer connection set --peer <peerId> and /reload in Pi.
Want to use a custom buyer proxy URL (remote host, custom port) — Set ANTSEED_BASE_URL=http://your-host:8377/v1 in the shell that launches pi. The extension reads this on startup. If your proxy is fronted by auth, also set ANTSEED_API_KEY=<token>.
How Pi talks to AntSeed
Wire format sent by Pi: OpenAI Responses (hits /v1/responses on the buyer proxy).
Best-fit services: any service whose protocols array contains openai-responses — that is what the peer advertises as natively-supported, so traffic passes through with zero translation overhead.
How to check a peer: run antseed network peer <peerId> --json and look at peer.providerServiceApiProtocols[provider].services[service] for each model. The browse command exposes the same field per peer.
When protocols differ: AntSeed's @antseed/api-adapter translates between OpenAI Responses and the service's native protocol on the fly. So a request from Pi can still reach a service that only advertises a different protocol — just with a small transform step.
TL;DR for agents: Edit ~/.openclaw/openclaw.json: under models.providers, add an antseed entry with baseUrl=http://127.0.0.1:8377, api="anthropic-messages", apiKey="antseed-p2p", and a models[] array whose id values match service ids from GET /v1/models. Optionally openclaw config set agents.defaults.model.primary "antseed/<id>". Reload with openclaw config reload.
What OpenClaw is. OpenClaw is an open-source agent runtime for autonomous, long-running tasks (research, coding, web automation). It loads its provider catalog from ~/.openclaw/openclaw.json — each entry is an HTTP endpoint plus a wire protocol (anthropic-messages, openai-chat, etc.) and a list of models.
How AntSeed plugs in. Add a provider entry called antseed that points at http://127.0.0.1:8377 with api: "anthropic-messages". Each model id you list under that provider must be a service id your pinned peer advertises — OpenClaw will surface them in its model picker as antseed/<service-id>.
Why a config entry instead of env vars. OpenClaw runs many providers in parallel (one per task, sometimes one per agent). A single base-URL override would force every agent through AntSeed; a named provider lets you mix AntSeed with hosted Anthropic, OpenAI, or local models on a per-agent basis.
Install
Install OpenClaw
npm install -g openclaw
Verify with openclaw --version. The config file lives at ~/.openclaw/openclaw.json and is created on first launch.
Each id under models[] must match a service id from curl http://127.0.0.1:8377/v1/models. apiKey is required by OpenClaw's validator but ignored by the proxy — any non-empty string works. The "antseed-p2p" value is just convention.
Test it
Confirm the proxy advertises the service ids you put in config
If a model id you listed in openclaw.json doesn't appear here, your pinned peer doesn't serve it. Pin a different peer or remove the entry.
Reload OpenClaw and check the provider list
openclaw config reload && openclaw providers list
Or restart OpenClaw. You should see antseed with the model count you configured.
Run an agent against AntSeed
openclaw run "Summarize the README in this repo" --model antseed/claude-sonnet-4-6
Troubleshooting
provider "antseed" not found when launching an agent — JSON parse error in openclaw.json, or you put the entry in the wrong nesting level. The provider must live under models.providers.antseed. Run openclaw config validate to surface parse errors.
OpenClaw lists antseed/<id> but every call returns 404 model_not_found — The pinned peer doesn't advertise that service id. Run antseed network peer <peerId> to see what it actually offers, or pin a different peer with antseed buyer connection set --peer <peerId>.
Streaming errors on long-running agents — AntSeed supports SSE streaming. If you see truncated responses, check that no proxy in front of OpenClaw is buffering (Cloudflare, nginx). The buyer proxy itself does not buffer.
Agent stalls on first request after a deploy — AntSeed opens a payment channel on the first request to a new peer (one on-chain transaction, ~5–15s on Base). Subsequent requests reuse the channel. Pre-warm by running a quick curl before launching the agent.
How OpenClaw talks to AntSeed
Wire format sent by OpenClaw: Anthropic Messages (hits /v1/messages on the buyer proxy).
Best-fit services: any service whose protocols array contains anthropic-messages — that is what the peer advertises as natively-supported, so traffic passes through with zero translation overhead.
How to check a peer: run antseed network peer <peerId> --json and look at peer.providerServiceApiProtocols[provider].services[service] for each model. The browse command exposes the same field per peer.
When protocols differ: AntSeed's @antseed/api-adapter translates between Anthropic Messages and the service's native protocol on the fly. So a request from OpenClaw can still reach a service that only advertises a different protocol — just with a small transform step.
Caveat: services whose only advertised protocol is openai-responses require streaming. If OpenClaw sends a non-streaming request and the proxy routes it to one of those services, the call fails with HTTP 400: Stream must be set to true. Pick a service whose protocols includes anthropic-messages (or another non-responses protocol) to avoid this.
TL;DR for agents: Edit ~/.hermes/config.yaml: add a custom_providers entry named antseed with base_url=http://127.0.0.1:8377/v1, api_mode=chat_completions, api_key="antseed-p2p", and a models: list whose ids match service ids from GET /v1/models. Set model.default to one of those ids and model.provider: antseed. Pin auxiliary.title_generation.model and auxiliary.compression.model to a chat_completions model to avoid streaming errors against openai-responses peers.
What Hermes is. Hermes is the agent framework from Nous Research (successor to OpenClaw's lineage). It's designed for autonomous, multi-step workflows — research agents, coding agents, swarms — and reads its model catalog from ~/.hermes/config.yaml.
How AntSeed plugs in. Add an entry under custom_providers with base_url: http://127.0.0.1:8377/v1, api_mode: chat_completions, and a list of models. Each model id must be a service id your pinned peer advertises. Then point model.default at the one you want as primary.
One Hermes-specific gotcha. Some peers serve GPT-style models via the openai-responses protocol, which requires streaming. Hermes' auxiliary calls (title generation, context compression) are non-streaming and will fail against those models with HTTP 400: Stream must be set to true. Pin auxiliary slots to a chat_completions model (config example below).
Install
Install or build Hermes
# Follow Nous Research setup at https://github.com/NousResearch/hermes-agent
Hermes is typically run as a long-lived process (often under systemd on a server). The config file ~/.hermes/config.yaml is read at startup — changes require a restart.
Configure
model:default:claude-sonnet-4-6provider:antseedcustom_providers:-name:antseedbase_url:http://127.0.0.1:8377/v1api_key:antseed-p2papi_mode:chat_completionsmodels:-claude-sonnet-4-6-claude-opus-4-7-deepseek-v4-flash-gpt-oss-120b-minimax-m2.7# Pin auxiliary calls to a chat_completions model so non-streaming# requests (title generation, compression) don't break against# openai-responses peers.auxiliary:title_generation:provider:antseedmodel:minimax-m2.7compression:provider:antseedmodel:minimax-m2.7
Only ids listed under models: show up in Hermes' picker — mirror it against curl http://127.0.0.1:8377/v1/models so you don't advertise models no peer serves. model.provider: antseed pins the default to this custom provider.
Test it
Confirm the proxy advertises the same ids your config references
Or whatever supervisor you use. Then check the journal: sudo journalctl -u hermes --no-pager -n 30.
After the first request, confirm a channel opened and is being metered
antseed buyer status
antseed buyer metering
status shows Active channels: 1 once the first request settles (~5–15s on Base — one on-chain tx to open the channel). metering shows the per-peer token + USDC totals for each channel. To poll: watch -n 1 antseed buyer metering.
Troubleshooting
HTTP 400: Stream must be set to true from auxiliary calls — You're routing through a peer that serves the model via openai-responses (which requires streaming), but Hermes' auxiliaries are non-streaming. Pin the auxiliary.* slots to a chat_completions model (see the config block above). Confirm a model's protocol with antseed network peer <peerId> — look for protocols: openai-chat-completions vs openai-responses.
Hermes loads the provider but every call returns no_peer_pinned — In the default manual flow AntSeed does not auto-select a peer — pin one with antseed buyer connection set --peer <peerId>, send x-antseed-pin-peer per request, or start the buyer with a router plugin. The session pin survives buyer-proxy restarts (it's persisted to ~/.antseed/buyer.state.json).
Hermes runs on a remote host and can't reach 127.0.0.1:8377 — Either run the buyer proxy on the same host as Hermes (recommended — keeps the hot signing key local), or expose the proxy via SSH tunnel: ssh -N -L 127.0.0.1:8377:127.0.0.1:8377 user@hermes-host. Do not bind the buyer proxy to a public interface.
Want to swap the routed model without restarting AntSeed — Edit model.default (and models: if needed) in config.yaml, re-pin a peer that serves it (antseed buyer connection set --peer <peerId>), then sudo systemctl restart hermes. The buyer proxy stays up; no contract calls.
How Hermes talks to AntSeed
Wire format sent by Hermes: OpenAI Chat Completions (hits /v1/chat/completions on the buyer proxy).
Best-fit services: any service whose protocols array contains openai-chat-completions — that is what the peer advertises as natively-supported, so traffic passes through with zero translation overhead.
How to check a peer: run antseed network peer <peerId> --json and look at peer.providerServiceApiProtocols[provider].services[service] for each model. The browse command exposes the same field per peer.
When protocols differ: AntSeed's @antseed/api-adapter translates between OpenAI Chat Completions and the service's native protocol on the fly. So a request from Hermes can still reach a service that only advertises a different protocol — just with a small transform step.
Caveat: services whose only advertised protocol is openai-responses require streaming. If Hermes sends a non-streaming request and the proxy routes it to one of those services, the call fails with HTTP 400: Stream must be set to true. Pick a service whose protocols includes openai-chat-completions (or another non-responses protocol) to avoid this.
TL;DR for agents: In GenLayer Studio: drop one JSON file per model into backend/node/create_nodes/default_providers/ with provider: "antseed", plugin: "openai-compatible", model: "<service-id>", and plugin_config.api_url: "http://host.docker.internal:8377" (NO /v1 suffix — the plugin appends it). Add "antseed" to the provider enum and an if/then rule to BOTH backend/.../providers_schema.json and frontend/.../providers_schema.json. Set ANTSEED_API_KEY=antseed in .env. Restart with genlayer up --reset. The user must run AntStation or antseed buyer start and pin a peer that serves the listed model ids.
What GenLayer Studio is. Studio runs Intelligent Contract validators that consult LLMs to reach consensus. Each validator is configured with a provider entry that has a provider name, a plugin (one of openai-compatible / anthropic / google / ollama / custom), a model id, and a plugin_config with api_url and api_key_env_var.
How AntSeed plugs in. Drop one JSON file per model into backend/node/create_nodes/default_providers/ with plugin: "openai-compatible" and api_url: "http://host.docker.internal:8377". Studio's openai-compatible plugin appends /v1/chat/completions automatically, so the buyer proxy receives a standard OpenAI Chat request and routes it to your pinned peer. Mirror the existing LibertAI entry (PR #1526) — it is the closest analogue: an openai-compatible host with a hosted base URL replaced by your local proxy.
Why host.docker.internal, not localhost. Studio's backend runs in Docker via genlayer up. From inside the container, localhost means the container itself, not your host machine — it cannot reach the AntSeed buyer proxy on the host. Mac/Windows Docker exposes the host as host.docker.internal; on Linux you must add extra_hosts: ["host.docker.internal:host-gateway"] to the backend service in docker-compose.yml or run with --network=host.
On Linux only — make host.docker.internal resolve from inside the backend container
# docker-compose.yml — patch the backend (jsonrpc) serviceservices:jsonrpc:extra_hosts:-"host.docker.internal:host-gateway"
Mac and Windows Docker Desktop already expose the host as host.docker.internal automatically — skip this step on those platforms. Restart with genlayer up --reset after editing.
# AntSeed authenticates with your local identity key, not this value.# Studio's openai-compatible plugin still requires the env var to be set.
ANTSEED_API_KEY=antseed
// In each schema, add "antseed" to the provider enum's examples…"provider":{"type":"string","examples":["ollama","openrouter","libertai","antseed", …]},// …and add an if/then block locking provider:antseed to plugin:openai-compatible{"if":{"properties":{"provider":{"const":"antseed"}}},"then":{"properties":{"plugin":{"const":"openai-compatible"}}}}
Both schema files must be kept in sync — the backend uses one for validation, the frontend uses the other for the UI dropdown. This is exactly what PR #1526 did for LibertAI.
Each provider JSON file pins exactly one model. Studio enumerates these into the validator-creation UI; pick services you know your pinned peer offers (check antseed network peer <peerId> --json | jq '.matchingServices[].service'). To expose more models later, drop in more antseed_<model>.json files — no schema edit needed.
Test it
Restart Studio so it re-scans default_providers/
genlayer up --reset
get_default_providers() in backend/node/create_nodes/providers.py reads every *.json in that folder once on boot, validates against providers_schema.json, and caches the result. Schema-validation errors abort startup with the offending file path — watch the logs.
In the Studio UI, create a new validator with provider "antseed"
You should see your antseed_*.json model ids in the dropdown. Save and trigger a contract that calls genlayer.eq_principle.prompt(…) — the request hits http://host.docker.internal:8377/v1/chat/completions on the AntSeed proxy and is forwarded to your pinned peer.
Confirm the validator call hit AntSeed
antseed buyer metering
Each validator call adds tokens + USDC to the channel for the peer you pinned. Run after a Studio request to see the totals update. To poll live: watch -n 1 antseed buyer metering.
Troubleshooting
Error validating file … antseed_*.json on genlayer up — The schema rejected your provider JSON. Most common cause: missing the if/then rule for provider:antseed, so it falls through with the wrong plugin. Add the rule to bothbackend/.../providers_schema.json and frontend/.../providers_schema.json. Run genlayer up --reset after editing.
Validator hangs, then errors with Connection refused to host.docker.internal:8377 — The backend container can't see your host. On Linux, add extra_hosts: ["host.docker.internal:host-gateway"] under the backend service in docker-compose.yml (see install step 2). On Mac/Windows, confirm Docker Desktop is running and the AntSeed proxy is up: curl http://host.docker.internal:8377/v1/models from inside the container with docker compose exec jsonrpc curl ….
Validator returns no_peer_pinned — No peer is pinned in the buyer proxy. Run antseed network browse, pick one, then antseed buyer connection set --peer <peerId>. Alternatively, send a per-request x-antseed-pin-peer header by extending the openai-compatible plugin — not currently exposed in the standard schema, so session pin is the path of least resistance.
404 model_not_found from a validator using e.g. claude-sonnet-4-6 — Your pinned peer doesn't advertise that service id. Run antseed network peer <peerId> --json | jq '.matchingServices[].service' to see what it does serve. Either pin a different peer or remove that antseed_<model>.json file.
First call after a restart takes 5–15 seconds — AntSeed opens a payment channel on the first request to a new peer (one Base-mainnet transaction). Subsequent calls reuse the channel. Pre-warm with curl -s http://localhost:8377/v1/chat/completions -d '{"model":"<id>","messages":[{"role":"user","content":"hi"}]}' before triggering Studio.
Caveats
AntSeed is a local daemon, not a hosted endpoint. Every Studio operator must run AntStation or antseed buyer start on their own machine and fund their wallet — there is no central account.
Free services exist on the AntSeed network (in: 0, out: 0), but using paid ones requires a USDC deposit on Base. AntStation guides users through this on first launch; the CLI exposes it as antseed payments.
How GenLayer Studio talks to AntSeed
Wire format sent by GenLayer Studio: OpenAI Chat Completions (hits /v1/chat/completions on the buyer proxy).
Best-fit services: any service whose protocols array contains openai-chat-completions — that is what the peer advertises as natively-supported, so traffic passes through with zero translation overhead.
How to check a peer: run antseed network peer <peerId> --json and look at peer.providerServiceApiProtocols[provider].services[service] for each model. The browse command exposes the same field per peer.
When protocols differ: AntSeed's @antseed/api-adapter translates between OpenAI Chat Completions and the service's native protocol on the fly. So a request from GenLayer Studio can still reach a service that only advertises a different protocol — just with a small transform step.
Caveat: services whose only advertised protocol is openai-responses require streaming. If GenLayer Studio sends a non-streaming request and the proxy routes it to one of those services, the call fails with HTTP 400: Stream must be set to true. Pick a service whose protocols includes openai-chat-completions (or another non-responses protocol) to avoid this.
TL;DR for agents: createOpenAICompatible({ name: 'antseed', baseURL: 'http://localhost:8377/v1', apiKey: 'antseed' }), then antseed('') as the model. Use @ai-sdk/openai-compatible (NOT @ai-sdk/openai). Service ids come from GET http://localhost:8377/v1/models. Per-request peer override: pass headers: { 'x-antseed-pin-peer': '' } in generateText/streamText.
What the AI SDK is. Vercel's ai package is a provider-agnostic TypeScript toolkit for building LLM apps and agents. You pick a provider (a small adapter package), instantiate a model from it, and pass that model into one of the framework's primitives: generateText, streamText, generateObject, or streamObject. The AI SDK handles tool-calling, structured output, message history, and streaming for you.
How AntSeed plugs in. AntSeed is OpenAI-Chat-compatible at http://localhost:8377/v1, so the right adapter is @ai-sdk/openai-compatible (not @ai-sdk/openai). The official OpenAI provider is locked to OpenAI's API surface and quietly drops third-party fields; the openai-compatible provider is the one Vercel's own docs recommend for proxies, gateways, and any non-OpenAI server that speaks Chat Completions. You point it at the AntSeed proxy with baseURL and pass any non-empty apiKey placeholder — the proxy authenticates with your local identity key, not with this header.
Which model ids work. The first argument to the provider call is the AntSeed service id (e.g. claude-sonnet-4-6, deepseek-v4-flash). It must match a service your pinned peer advertises — confirm with curl http://localhost:8377/v1/models.
Prerequisites
Node.js 18 or newer
Install
Install the SDK and the openai-compatible provider
npm install ai @ai-sdk/openai-compatible zod
zod is only needed if you call generateObject / streamObject. Skip it for plain text generation.
Configure
// antseed.ts — a single provider instance you can import everywhereimport { createOpenAICompatible } from'@ai-sdk/openai-compatible';
exportconst antseed = createOpenAICompatible({
name: 'antseed',
baseURL: 'http://localhost:8377/v1',
apiKey: 'antseed', // any non-empty string — proxy ignores this headerincludeUsage: true, // surface token counts in streaming responses too
});
// stream.tsimport { streamText } from'ai';
import { antseed } from'./antseed';
const result = streamText({
model: antseed('claude-sonnet-4-6'), // an AntSeed service idprompt: 'Why is the sky blue?',
});
forawait (const chunk of result.textStream) {
process.stdout.write(chunk);
}
console.log('\nusage:', await result.usage);
// structured.ts — generateObject works the same wayimport { generateObject } from'ai';
import { z } from'zod';
import { antseed } from'./antseed';
const { object } = awaitgenerateObject({
model: antseed('claude-sonnet-4-6'),
schema: z.object({
title: z.string(),
bullets: z.array(z.string()).min(3).max(5),
}),
prompt: 'Summarize the AntSeed buyer-proxy README as a slide.',
});
console.log(object);
The string you pass to antseed('<id>') is forwarded verbatim as model in the OpenAI Chat request. Run curl -s http://localhost:8377/v1/models | jq '.data[].id' to see exactly what your pinned peer offers.
Test it
Run a smoke test with tsx
npx tsx stream.ts
Example output:
The sky is blue because shorter (blue) wavelengths of sunlight
scatter much more than longer (red) wavelengths in Earth's atmosphere…
usage: { promptTokens: 14, completionTokens: 78, totalTokens: 92 }
If you see 404 model_not_found, the pinned peer does not advertise the id you passed. If you see no_peer_pinned, run antseed buyer connection set --peer <peerId> first — or send the per-request header (next step).
Per-request peer override (no session pin needed)
// Use `headers` to fan out to different peers per call.const result = streamText({
model: antseed('claude-sonnet-4-6'),
prompt: 'hi',
headers: {
'x-antseed-pin-peer': 'cccccccccccccccccccccccccccccccccccccccc',
},
});
Useful when one Node process serves many tenants and you want each request routed to a different peer. The header overrides the session pin for that single call.
Troubleshooting
TypeScript complains that antseed has no call signature — You imported from @ai-sdk/openai instead of @ai-sdk/openai-compatible. Switch the package — the SDK's official OpenAI provider is locked to OpenAI's service ids and rejects unknown ones.
generateObject returns malformed JSON — The AI SDK is strict about JSON Schema support. Pass supportsStructuredOutputs: true to createOpenAICompatible only if your pinned peer's service supports OpenAI-style structured outputs natively. If unsure, leave it off — the SDK falls back to tool-call-based JSON which works everywhere.
includeUsage is set but result.usage is undefined — Some upstream providers behind AntSeed do not emit usage on streamed responses. Try generateText instead of streamText for definitive token counts; otherwise run antseed buyer metering for the authoritative per-channel token + USDC totals AntSeed itself measured.
Browser/edge runtime fails with fetch errors — The AntSeed proxy listens on 127.0.0.1:8377, which is not reachable from a browser tab on a deployed site. The AI SDK is designed to run on the server (Route Handlers, Server Actions, edge functions on your own machine, or a Node process); don't call it from a client component when the model is AntSeed.
How Vercel AI SDK talks to AntSeed
Wire format sent by Vercel AI SDK: OpenAI Chat Completions (hits /v1/chat/completions on the buyer proxy).
Best-fit services: any service whose protocols array contains openai-chat-completions — that is what the peer advertises as natively-supported, so traffic passes through with zero translation overhead.
How to check a peer: run antseed network peer <peerId> --json and look at peer.providerServiceApiProtocols[provider].services[service] for each model. The browse command exposes the same field per peer.
When protocols differ: AntSeed's @antseed/api-adapter translates between OpenAI Chat Completions and the service's native protocol on the fly. So a request from Vercel AI SDK can still reach a service that only advertises a different protocol — just with a small transform step.
Caveat: services whose only advertised protocol is openai-responses require streaming. If Vercel AI SDK sends a non-streaming request and the proxy routes it to one of those services, the call fails with HTTP 400: Stream must be set to true. Pick a service whose protocols includes openai-chat-completions (or another non-responses protocol) to avoid this.
TL;DR for agents: ChatOpenAI(model='', base_url='http://localhost:8377/v1', api_key='antseed') from langchain-openai. Drops into LCEL, create_react_agent, RAG, with_structured_output. Per-request peer override: extra_headers={'x-antseed-pin-peer': ''}. Service ids come from GET http://localhost:8377/v1/models. Reasoning traces (reasoning_content, etc.) are NOT preserved by ChatOpenAI — use the Responses endpoint for those.
What LangChain is. LangChain is the Python framework for composing LLMs with tools, retrievers, memory, and agents. The chat-model interface is BaseChatModel; ChatOpenAI from langchain-openai is a concrete subclass that talks the OpenAI Chat Completions wire format.
How AntSeed plugs in. Pass base_url="http://localhost:8377/v1" and any non-empty api_key to ChatOpenAI. Once you have an instance, every primitive that accepts a chat model — LCEL pipes (prompt | llm | parser), tool-calling agents, create_react_agent, LangGraph nodes, RAG chains, structured-output binding via with_structured_output — will route through AntSeed without any further changes.
One thing to know. LangChain's ChatOpenAI is OpenAI-strict by design: it will not preserve non-standard response fields like reasoning_content, reasoning, or reasoning_details that some third-party servers emit. For chat, tool-calling, and structured output this is fine. If you specifically need a model's reasoning traces, consider using the AntSeed buyer proxy with the OpenAI Responses endpoint (/v1/responses) via a different provider package, or use a model that returns reasoning inline.
Prerequisites
Python 3.10 or newer
Install
Install LangChain and the OpenAI integration
pip install -U langchain langchain-openai
Configure
# antseed_llm.py — import this once, reuse everywhere.from langchain_openai import ChatOpenAI
antseed = ChatOpenAI(
model="claude-sonnet-4-6", # an AntSeed service id
base_url="http://localhost:8377/v1",
api_key="antseed", # any non-empty string
temperature=0.7,
# max_completion_tokens=2048, # uncomment for hard caps
)
print(antseed.invoke("Hello").content)
# pipeline.py — LCEL chain. Identical to OpenAI; the swap is invisible.from langchain_core.prompts import ChatPromptTemplate
from langchain_core.output_parsers import StrOutputParser
from antseed_llm import antseed
prompt = ChatPromptTemplate.from_messages([
("system", "You are a concise technical writer."),
("human", "Explain {topic} in one paragraph."),
])
chain = prompt | antseed | StrOutputParser()
print(chain.invoke({"topic": "payment channels"}))
# tools.py — tool-calling agent. Works because AntSeed forwards OpenAI tool calls verbatim.from langchain_core.tools import tool
from langgraph.prebuilt import create_react_agent
from antseed_llm import antseed
@tooldefget_weather(city: str) -> str:
"""Return the current weather for a city."""returnf"It's 22°C and sunny in {city}."
agent = create_react_agent(antseed, [get_weather])
result = agent.invoke({
"messages": [("user", "What's the weather in Lisbon?")]
})
print(result["messages"][-1].content)
Pick services whose protocols array includes openai-chat-completions (most do natively; the rest are translated automatically by @antseed/api-adapter). Tool calling and structured output rely on the service supporting OpenAI-style function-call syntax — confirm with a quick smoke test before building large agents.
Test it
Run the basic example
python antseed_llm.py
Example output:
Hello! How can I help you today?
Per-request peer override (no session pin needed)
# extra_headers is forwarded as-is to the proxy.from langchain_openai import ChatOpenAI
llm = ChatOpenAI(
model="claude-sonnet-4-6",
base_url="http://localhost:8377/v1",
api_key="antseed",
extra_headers={
"x-antseed-pin-peer": "cccccccccccccccccccccccccccccccccccccccc",
},
)
print(llm.invoke("hi").content)
Use this when a single Python process needs to fan out to different peers per call (multi-tenant, scheduled jobs, A/B tests across peers).
Verify it actually went through AntSeed
antseed buyer metering
buyer metering reads the local SQLite log and prints per-channel token + USDC totals. After your python call, the channel for the peer you pinned should show non-zero input/output tokens. (buyer status is a snapshot view — it shows the active-channel count but not per-call usage.)
Troubleshooting
openai.NotFoundError: 404 … model_not_found — The pinned peer does not advertise the id you passed. Confirm with curl http://localhost:8377/v1/models | jq and either pin a different peer or change the model= argument.
openai.APIConnectionError: Connection refused — The buyer proxy is not running. Start it with antseed buyer start (or open AntStation desktop). Confirm curl http://localhost:8377/v1/models works before retrying from Python.
with_structured_output returns the right schema but empty fields — Either the model behind the pinned peer does not support OpenAI tool-call syntax, or you used method="json_mode" against a service that does not honor it. Try method="function_calling" (the default), and prefer services tagged coding or tools in antseed network peer <peerId> --json.
Streaming with stream=True truncates mid-response — A buffering proxy (nginx, Cloudflare) sits between your code and the buyer proxy. The AntSeed proxy itself does not buffer SSE. Either bypass the intermediate proxy or set its buffering off (proxy_buffering off; in nginx).
Reasoning traces missing on a model you know emits them — See the third paragraph above: langchain-openai does not preserve non-standard response fields. For first-class reasoning support, route the request through the OpenAI Responses endpoint (POST /v1/responses on the proxy) using a Responses-aware client, or pick a model that puts reasoning inline in content.
How LangChain (Python) talks to AntSeed
Wire format sent by LangChain (Python): OpenAI Chat Completions (hits /v1/chat/completions on the buyer proxy).
Best-fit services: any service whose protocols array contains openai-chat-completions — that is what the peer advertises as natively-supported, so traffic passes through with zero translation overhead.
How to check a peer: run antseed network peer <peerId> --json and look at peer.providerServiceApiProtocols[provider].services[service] for each model. The browse command exposes the same field per peer.
When protocols differ: AntSeed's @antseed/api-adapter translates between OpenAI Chat Completions and the service's native protocol on the fly. So a request from LangChain (Python) can still reach a service that only advertises a different protocol — just with a small transform step.
Caveat: services whose only advertised protocol is openai-responses require streaming. If LangChain (Python) sends a non-streaming request and the proxy routes it to one of those services, the call fails with HTTP 400: Stream must be set to true. Pick a service whose protocols includes openai-chat-completions (or another non-responses protocol) to avoid this.
curl / raw HTTP can send any of AntSeed's supported wire formats. Match the request format against each service's protocols array (advertised per service in providerServiceApiProtocols) — when it matches, the request passes through untouched; when it does not, @antseed/api-adapter translates on the fly.
