acpx

name

acpx

description

Use acpx as a headless ACP CLI for agent-to-agent communication, always inside an isolated SubAgent. Use when running coding agents through acpx, managing persistent ACP sessions, queueing prompts, consuming structured agent output from scripts, comparing the same prompt across multiple agents, or composing multi-agent workflows with defineFlow/decision/decisionEdge. Never invoke the claude adapter (nested-instance blacklist).

acpx

What acpx is

acpx is a headless, scriptable CLI client for the Agent Client Protocol (ACP). It is built for agent-to-agent communication over the command line and avoids PTY scraping.

Core capabilities:

• Persistent multi-turn sessions per repo/cwd
• One-shot execution mode (exec)
• Multi-agent comparison (compare)
• Named parallel sessions (-s/--session)
• Idempotent session creation (sessions ensure)
• Session retention controls (sessions prune with age filters and history cleanup)
• Portable session export/import for moving records and history across machines
• Queue-aware prompt submission with optional fire-and-forget (--no-wait)
• Cooperative cancel command (cancel) for in-flight turns
• Graceful cancellation via ACP session/cancel on interrupt
• Session control methods (set-mode, set <key> <value>)
• Agent reconnect/resume after dead subprocess detection
• Prompt input via stdin or --file
• Config files with global+project merge and config show|init
• Session-scoped MCP servers from an external file (--mcp-config)
• Session metadata/history inspection (sessions show, sessions history)
• Local agent process checks via status
• Stable ACP client methods for filesystem and terminal requests
• Stable ACP authenticate handshake via env/config credentials
• Structured streaming output (text, json, quiet) with optional --suppress-reads
• Built-in agent registry plus raw --agent escape hatch
• Claude system prompt override via --system-prompt / --append-system-prompt
• Optional terminal capability disable via --no-terminal for review-only flows
• Optional filesystem capability disable via --no-fs for agent-native file operations
• Tool whitelist (--allowed-tools), turn cap (--max-turns), retry on transient failures (--prompt-retries)
• Multi-agent flows via acpx flow run and the acpx/flows authoring API (defineFlow, decision, decisionEdge, acp, action, compute, checkpoint)

Install

npm i -g acpx

For normal session reuse, prefer a global install over npx.

Operating protocol (when loaded inside Claude Code)

CRITICAL: when this skill is loaded inside Claude Code, the following protocol governs every invocation. These rules override any upstream default or example in this document.

1. Run inside a dedicated, clean SubAgent

Every acpx-driven task MUST execute inside an isolated SubAgent launched for that task (e.g. via the Task/Agent tool), with its own fresh context. Do not run acpx work inline in the main conversation thread. Rationale: agent-to-agent calls are long, noisy, and token-heavy; isolating them keeps the parent context clean and lets the SubAgent return only the distilled result.

2. Never call the claude adapter

The claude built-in is blacklisted (see the registry rule below). Calling acpx claude ... spawns a nested Claude instance — forbidden. Route to a non-Claude agent instead.

3. Execution sequence

For each user request, follow this order inside the SubAgent:

1. Discover available agents first. Probe which ACP adapters are actually installed on this machine before choosing one. Run:

   acpx --help                                  # lists built-in agent names
   command -v codex gemini cursor copilot 2>/dev/null   # which CLIs are on PATH
   

   Pick the first installed non-Claude agent that fits the task (codex is the default; fall back through gemini, qwen, cursor, copilot, droid, opencode, ...). Do not assume an agent is installed — verify, then use it.

2. Dispatch the user's actual request to the chosen agent (prompt for multi-turn, exec for one-shot, compare for cross-agent).

3. Do not duplicate upstream skill content. Before invoking an agent for a task, check whether the upstream acpx skill/reference material already documents the capability needed (e.g. session lifecycle, flow authoring, output formats, permission policies in references/). If it does, defer to that knowledge and run the documented command — do not re-derive or re-explain it. Only synthesize new steps when the request is not already covered upstream.

The SubAgent returns a short summary of what ran, which agent it used, and the distilled result — not the raw ACP stream.

4. Reflect on every result with a fresh blank SubAgent before acting

CRITICAL: never accept an acpx SubAgent's result as final. Whatever the acpx SubAgent returns — a review, an evaluation, a diff, a recommendation, or a "done" — is a proposal from an outside agent, not a verdict. Before acting on it, spawn a SECOND, independent, blank SubAgent and have it reflect on that result; only after the reflection returns do you decide the next step.

• The reflection SubAgent MUST be a brand-new agent (a fresh Task/Agent invocation) with no context of the acpx run, the original prompt, or the implementation. It is NOT the SubAgent that ran acpx, and it is NOT the main thread. Reusing the acpx SubAgent or judging inline destroys the independence that makes the reflection worth anything — producer and judge must be different agents.
• Give the reflection SubAgent only the artifact under review (the acpx output plus the relevant files/diff), not the reasoning that produced it. It must reach its own conclusion.
• The reflection drives the branch: accept, rework (re-dispatch to acpx with corrections), or escalate to the user. Example: after acpx ... review, the blank SubAgent critiques the review itself — which findings are real, which are noise, what it missed — and that critique, not the raw review, decides what you do next.

This is the same GAN-evaluator / independent-audit pattern used elsewhere in this project: the agent that produces a result is never the agent that judges it.

Command model

prompt is the default verb.

acpx [global_options] [prompt_text...]
acpx [global_options] prompt [prompt_options] [prompt_text...]
acpx [global_options] exec [prompt_options] [prompt_text...]
acpx [global_options] compare <agent>... '<prompt_text>'
acpx [global_options] compare <agent>... --file <path>
acpx [global_options] cancel [-s <name>]
acpx [global_options] set-mode <mode> [-s <name>]
acpx [global_options] set <key> <value> [-s <name>]
acpx [global_options] status [-s <name>]
acpx [global_options] sessions [list | new [--name <name>] | ensure [--name <name>] | close [name] | show [name] | history [name] [--limit <count>] | export [name] --output <path> | import <archive> [--name <name>] [--cwd <dir>] | prune [--dry-run] [--before <date> | --older-than <days>] [--include-history]]
acpx [global_options] config [show | init]
acpx [global_options] flow run <file> [--input-json '<json>' | --input-file <path>] [--default-agent <name>]

If prompt text is omitted and stdin is piped, acpx reads prompt text from stdin.

Built-in agent registry

Friendly agent names resolve to commands:

• pi -> npx pi-acp@^0.0.31
• openclaw -> openclaw acp
• codex -> npx -y @agentclientprotocol/codex-acp@^1.1.5
• claude -> npx -y @agentclientprotocol/claude-agent-acp@^0.60.0
• gemini -> gemini --acp
• cursor -> cursor-agent acp
• copilot -> copilot --acp --stdio
• droid -> droid exec --output-format acp
• fast-agent -> uvx fast-agent-mcp acp
• grok-build -> grok agent stdio
• iflow -> iflow --experimental-acp
• kilocode -> npx -y @kilocode/cli acp
• kimi -> kimi acp
• kiro -> kiro-cli-chat acp
• mux -> npx -y mux@^0.28.0 acp
• opencode -> npx -y opencode-ai acp
• pool -> pool acp
• qoder -> qodercli --acp
• qwen -> qwen --acp
• trae -> traecli acp serve
• zeroclaw -> zeroclaw acp

Rules:

• Default agent is codex for top-level prompt, exec, compare, and sessions.
• factory-droid and factorydroid also resolve to the built-in droid adapter.
• grok-build authenticates through the installed grok CLI: its agent-managed cached login when the server advertises cached_token, or XAI_API_KEY when it advertises xai.api_key.
• Unknown positional agent tokens are treated as raw agent commands.
• --agent <command> explicitly sets a raw ACP adapter command.
• Do not combine a positional agent and --agent in the same command.
• CRITICAL: never invoke the claude adapter. This skill runs inside Claude Code, so acpx claude would spawn a nested Claude instance — redundant, slower, and it adds no model diversity. Prefer codex (default), gemini, qwen, or another non-Claude agent. Only call claude if the user explicitly requests a second Claude instance by name.

Key commands

Prompt (default, persistent session)

acpx codex 'fix flaky tests'
acpx codex prompt 'fix flaky tests'
acpx prompt 'fix flaky tests'   # defaults to codex

• Uses a saved session for the session scope key, auto-resumes prior session
• If no session exists for the scope, exits with NO_SESSION and prompts for sessions new
• Is queue-aware when another prompt is already running for the same session
• On interrupt during an active turn, sends ACP session/cancel before force-kill fallback

Prompt options: -s, --session <name>, --no-wait, -f, --file <path>

Exec (one-shot)

acpx exec 'summarize this repo'
acpx codex exec 'summarize this repo'

Runs a single prompt in a temporary ACP session. Does not reuse or save persistent session state.

Compare (multi-agent, one-shot)

acpx compare codex gemini qwen 'summarize this repo in 3 lines'
acpx compare codex gemini --file ./prompt.md
acpx compare codex gemini -- '--looks-like-a-flag'

Runs the same prompt across multiple agents, each in a temporary exec-style session. Honors the same global execution controls as exec (--cwd, --timeout, permission flags, --policy, auth, terminal advertising, retries, model/system options, --format).

• Use -- after the agent list when prompt words might be parsed as flags (see the third example).

• --format text prints one summary-table row per agent (timing, token usage, stop reason, permissions, final output)

• --format json or command-local --json prints a CompareRow[] summary payload

• --format quiet prints <agent>\t<status> per row

• CompareRow.status is ok, cancelled, permission_denied, or error

• Agents run serially in the requested workspace; no saved sessions or separate transcript directories are created

Cancel / Mode / Config / Model

acpx codex cancel
acpx codex set-mode auto
acpx codex set model gpt-5.2[high]
acpx codex set model gpt-5.4

• cancel: sends cooperative session/cancel through queue-owner IPC
• set-mode: calls ACP session/set_mode
• set: calls ACP session/set_config_option
• set model <id>: calls session/set_model for mid-session model switching

Sessions

acpx sessions list                          # list all sessions
acpx sessions new --name backend            # create fresh session
acpx sessions ensure --name backend         # idempotent: get or create
acpx sessions close backend                 # close a session
acpx sessions show backend                  # show metadata
acpx sessions history backend --limit 20    # show turn history
acpx sessions export backend --output backend-session.json
acpx sessions import backend-session.json --name backend-restored
acpx sessions prune --dry-run --older-than 7
acpx sessions prune --older-than 30 --include-history
acpx status                                 # check local agent process

Prefix any command with an agent name: acpx codex sessions ensure --name backend

Global options

• --agent <command>: raw ACP agent command (escape hatch)
• --cwd <dir>: working directory for session scope (default: current directory)
• --mcp-config <path>: load mcpServers from an external JSON file for the invocation, replacing project/global MCP config. Relative paths resolve from --cwd. A live persistent session rejects MCP config changes until it is closed.
• --approve-all: auto-approve all permission requests
• --approve-reads: auto-approve reads/searches, prompt for writes (default mode)
• --deny-all: deny all permission requests
• --non-interactive-permissions <policy>: when prompting is unavailable, choose deny or fail
• --permission-policy <json-or-file> / --policy: per-tool ACP permission rules
• --format <fmt>: output format (text, json, quiet)
• --json-strict: strict JSON mode; requires --format json and suppresses non-JSON stderr output
• --suppress-reads: suppress raw read-file contents while preserving the selected format
• --timeout <seconds>: max wait time (positive number)
• --ttl <seconds>: queue owner idle TTL before shutdown (default 300, 0 disables TTL)
• --model <id>: request an agent model during session creation
• --system-prompt <text>: replace the agent system prompt (persisted in session)
• --append-system-prompt <text>: append text to the agent system prompt
• --allowed-tools <list>: comma-separated tool whitelist (use "" for no tools)
• --max-turns <count>: cap session turn count
• --prompt-retries <count>: retry failed prompt turns on transient errors (default 0)
• --no-fs: advertise ACP filesystem read/write capabilities as disabled so compatible agents use their native filesystem implementation
• --no-terminal: do not advertise the ACP terminal capability

Permission flags are mutually exclusive.

System prompt override (Claude-specific mechanism)

Note: per the agent-registry rule above, the claude adapter is blocked from normal use. The override mechanism itself is only honored by the Claude adapter, so the examples below show the Claude form for reference. Only use them when the user explicitly asks for a nested Claude instance.

# Replace the system prompt for a named session, persisted across reuse
acpx --system-prompt "You are a code reviewer who challenges every implicit assumption." claude -s review

# Append a guideline on top of the default system prompt
acpx --append-system-prompt "Always explain trade-offs before recommending a fix." claude -s impl

The override is forwarded via ACP _meta.systemPrompt on session/new and stored in session_options.system_prompt. Subsequent prompt/ensure calls in the same scope keep the override unless you explicitly create a new session. Non-Claude adapters ignore the field.

Config files

Config files are merged in this order (later wins):

• global: ~/.acpx/config.json
• project: <cwd>/.acpxrc.json

Supported keys: defaultAgent, defaultPermissions, nonInteractivePermissions, authPolicy, ttl, timeout, format, agents map, auth map.

Use acpx config show to inspect the resolved config and acpx config init to create the global template.

Custom agents on Windows: define structured agents.<name>.argv (command plus arguments array). Unambiguous legacy command + args entries migrate automatically; raw or ambiguous commands and .sh wrappers must move to argv (with explicit migration guidance), and saved custom-agent sessions without argv must be recreated.

For ACP authenticate handshakes, use either config auth entries or explicit ACPX_AUTH_<METHOD_ID> environment variables such as ACPX_AUTH_OPENAI_API_KEY. Ambient provider env vars like OPENAI_API_KEY pass through to child agents but do not trigger ACP auth-method selection on their own.

Environment variables

• ACPX_CLAUDE_INCLUDE_USER_SETTINGS=1: opt in to loading Claude Code user settings for built-in claude sessions. By default only project/local settings load, so globally enabled channel or daemon plugins cannot interfere with spawned ACP sessions.
• ACPX_AUTH_<METHOD_ID>: explicit credential for an ACP authenticate method.
• Session storage path is derived from the OS home directory (~/.acpx/sessions); child processes inherit the current environment by default.

Session behavior

Persistent prompt sessions are scoped by: agentCommand, absolute cwd, optional session name.

• Session records are stored in ~/.acpx/sessions/*.json
• -s/--session creates parallel named conversations in the same repo
• Changing --cwd changes scope and therefore session lookup
• Closed sessions are retained on disk with closed: true and closedAt until pruned
• Auto-resume by scope skips closed sessions
• Prompt mode attempts to reconnect to saved session; if adapter-side session is invalid/not found, acpx creates a fresh session and updates the saved record

Prompt queueing and --no-wait

Queueing is per persistent session. The active acpx process for a running prompt becomes the queue owner. Other invocations submit prompts over local IPC.

• Default: enqueue and wait for queued prompt completion, streaming updates back
• --no-wait: enqueue and return after queue acknowledgement
• Ctrl+C during an active turn sends ACP session/cancel, waits briefly, then force-kills only if cancellation does not finish in time
• cancel sends the same cooperative cancellation without requiring terminal signals
• After the queue drains, owner shutdown is governed by TTL (default 300s, configurable with --ttl)

Output formats

Use --format <fmt>:

• text (default): human-readable stream with updates/tool status and done line
• json: NDJSON event stream (good for automation)
• quiet: final assistant text only
• --suppress-reads: replace raw read-file contents with [read output suppressed]
• --json-strict: pair with --format json to suppress non-JSON stderr noise

Example automation:

acpx --format json codex exec 'review changed files' \
  | jq -r 'select(.type=="tool_call") | [.status, .title] | @tsv'

Flows (multi-agent workflows)

Flows let you declare a multi-agent workflow as a graph of typed nodes connected by edges, executed by the acpx runtime. The runtime owns persistence, retries, timeouts, and routing — the flow file declares the shape, not the engine.

acpx flow run ./my-flow.flow.ts --input-file ./flow-input.json
acpx flow run ./my-flow.flow.ts --input-json '{"task":"FIX: add a regression test"}'
acpx --approve-all flow run examples/flows/pr-triage/pr-triage.flow.ts \
  --input-json '{"repo":"openclaw/acpx","prNumber":150}'
acpx flow run ./my-flow.flow.ts --default-agent codex

Run artifacts persist under ~/.acpx/flows/runs/<runId>/. Default per-step timeout is 15 minutes when --timeout is unset.

The authoring surface lives in acpx/flows. Node types: acp (model-driven step), decision (constrained-choice LLM step), action (runtime-supervised deterministic operation), compute (pure local data transform), checkpoint (pause point for human or external trigger).

See references/advanced.md for the full authoring example, edge shapes, and detailed node type reference, plus the full Practical workflows example set (persistent assistant, named streams, specialized reviewer, idempotent bootstrap, --no-wait follow-up, one-shot exec, cross-agent compare, --mcp-config, JSON orchestration, raw adapter, periodic cleanup, triage flow, repo-scoped review).