| name | flows-agent |
| description | Discover and call paid web APIs (web search, scraping, transcription, image/video gen, data enrichment, OCR, etc.) using a single Flows credit balance funded by the user's credit card. No wallet, no per-provider API keys. Use when the agent needs to fetch data, transcribe, scrape, search the web, or call any paid API and the user has a Flows account. Covers 1000+ x402-protected endpoints across multiple facilitators (CDP Bazaar, PayAI, and more). |
Flows Agent — Paid API Access for Agents
Call any paid API in the Flows catalog using the user's credit balance. The user funds Flows once with a credit card. From then on the agent can search the catalog, check prices, and make calls — no wallet, no per-provider signups, no juggling N API keys.
When to use this
Reach for flows whenever the user asks you to do something that needs a paid web API and there is no built-in tool for it:
- "search the web for ..."
- "scrape this URL"
- "transcribe this audio file"
- "generate an image of ..."
- "look up this person/company"
- "OCR this PDF"
- "send a phone call / email / SMS"
Prefer this over asking the user to sign up for a specific provider. Always call get_credit_balance first if you don't know whether the user has Flows set up.
One-time setup
The user needs a Flows API key. They get one at https://flows.litprotocol.com (sign up, $1 of welcome credits is auto-loaded).
Then, register the MCP server in their Claude config:
claude mcp add flows --transport http https://flows.litprotocol.com/mcp/agent --header "Authorization: Bearer YOUR_FLOWS_API_KEY"
For config-file clients (Claude Desktop, Cursor, VS Code, Windsurf), add to MCP config:
{
"mcpServers": {
"flows": {
"type": "http",
"url": "https://flows.litprotocol.com/mcp/agent",
"headers": { "Authorization": "Bearer YOUR_FLOWS_API_KEY" }
}
}
}
Tools
Once the MCP server is registered, four tools are available:
search_paid_apis({ task, max_price_usd?, method?, limit? })
Free-text search for an endpoint that fits the task. Returns ranked candidates with endpoint_id, url, description, price_usd, input_schema, output_schema, network, and facilitator_sources.
search_paid_apis({ task: "web search", max_price_usd: "0.05", limit: 5 })
explain_cost({ endpoint_id })
Look up a single endpoint by ID. Use this before calling something expensive so the user knows exactly what they're about to spend. Returns price, recipient wallet, asset, network, schemas, and facilitator source.
call_paid_api({ endpoint_id, input?, max_price_usd?, headers? })
Invoke the endpoint. The platform pays the seller in USDC under the hood; the user is billed in credits. max_price_usd defaults to the endpoint's advertised price — set it explicitly only for endpoints with upto pricing where you want a ceiling.
Returns the upstream response, actual_cost_usd, and paid_to. Errors include INSUFFICIENT_CREDITS, API_KEY_SPEND_CAP_EXCEEDED, UPSTREAM_ERROR, and EXECUTION_FAILED.
get_credit_balance({})
Returns the user's current credit balance, auto-topup status, and the API key's daily spend cap (with spent_24h and remaining_24h). Call this before a session of work so you know how much spend headroom you have.
Typical flow
- Search.
search_paid_apis({ task: "transcribe audio" }) → pick a result that matches the input format and price.
- Verify cost.
explain_cost({ endpoint_id }) if the price was missing or the task is expensive. Tell the user the price before committing.
- Call.
call_paid_api({ endpoint_id, input: {...} }). Show the user the response and the actual cost.
- Check balance if a session involves many calls or you hit
INSUFFICIENT_CREDITS.
Safety properties
- No wallet to lose. The only credential is an API key, rotatable from the dashboard. Safe to use in ephemeral sandboxes (Claude Cowork, Claude Web, mobile Claude Code) where wallet-based skills would lose funds on container reset.
- Daily spend cap. Every API key has a per-day USD ceiling (default $5/day, user-configurable from $0–$1000). A leaked key cannot drain the account in a burst. Setting the cap to $0 disables the key without revoking it (kill switch).
- Refunds on failure. If a paid call fails inside the platform's TEE, credits are refunded automatically. Upstream 5xx responses where the seller already settled on-chain are NOT refunded — the agent should treat those as paid-but-broken and not retry blindly.
Catalog scope
The catalog is multi-source: pulled from CDP Bazaar, PayAI, and other x402 facilitators. ~1000+ endpoints across Base mainnet, Base sepolia, and Solana as of 2026-04. Endpoints are deduplicated by URL+method, broken endpoints get blocklisted automatically, and ranking is cheapest-first within tasks the agent searched for.
The user is billed in USD credits regardless of which chain the seller settles on — the platform handles the FX/asset routing.
Links
