stuntman / agents

Built for agents that test their own work.

Claude Code, Codex, Gemini CLI, OpenCode, Cursor and GitHub Copilot connect over MCP. Anything else uses plain HTTP. Both reach the same 356 tools, and the timeline records who acted.

Get startedMCP/mcpHTTP/api

Works with

Claude CodeCodexGemini CLIOpenCodeCursorGitHub Copilot

The loop it runs on its own.

  1. Read the codefinds the providers your app calls
  2. Set up Stuntmaninstalls plugins, points base URLs at Stuntman
  3. Change the codethe fix or the feature you asked for
  4. Run the real flowyour app, end to end
  5. Provoke failuresthe cases staging never shows
  6. Read what the app sentcalls, retries, callbacks
  7. Check the resultagainst the timeline, not a guess

Connect it in one line.

MCP runs over streamable HTTP at /mcp; each environment also has /e/<env>/mcp. Client formats change: if one does not load, check its docs for “remote” or “streamable HTTP” servers.

Claude Code

claude mcp add --transport http stuntman \
  http://localhost:8025/mcp

Codex

codex mcp add stuntman --url http://localhost:8025/mcp

Gemini CLI · settings.json

{ "mcpServers": { "stuntman":
  { "httpUrl": "http://localhost:8025/mcp" } } }

OpenCode · opencode.json

{ "mcp": { "stuntman": { "type": "remote",
  "url": "http://localhost:8025/mcp" } } }

Cursor · .cursor/mcp.json

{ "mcpServers": { "stuntman":
  { "url": "http://localhost:8025/mcp" } } }

GitHub Copilot · .vscode/mcp.json

{ "servers": { "stuntman": { "type": "http",
  "url": "http://localhost:8025/mcp" } } }
There is no API key yet.Anyone who reaches the port can use every tool. Per-agent keys are planned; until then, keep Stuntman private.

Or plain HTTP.

Every tool is an HTTP route under /api, with the same arguments. Pass by to name the actor on the timeline; MCP clients are named automatically.

curl -s localhost:8025/api/tools | jq ".[].name"

curl -s -X POST localhost:8025/api/whatsapp/force \
  -H "content-type: application/json" \
  -d '{"answer": "500", "count": 1, "by": "codex"}'

It sets Stuntman up, too.

Ask it to read the repository, find every provider, install the plugins, point the staging URLs at Stuntman and prove nothing leaks.

Source files become provider lines through the agent into Stuntman

What an agent can do.

A sample of the real tool names, grouped by intent. The full list is in the tool reference.

IntentTools
Watchlist_sms · list_whatsapp_messages · list_api_calls · list_ai_calls · list_product_events · list_traces · wait_for_object · wait_for_api_call · wait_for_file
Provokecreate_rule · set_rule_enabled · force_answer · force_sms_failure · force_whatsapp_failure · force_ai_failure · force_events_failure · expire_tokens · send_bad_webhook · provoke_sign_in
Play the peoplereply_as_user · reply_sms_as_user · simulate_recipient · put_mail · put_file · produce · send_bot_activity
Set uplist_plugins · install_plugin · import_spec · import_mocks · start_recording · make_plugin · set_provider · create_environment
Move timeclock · advance_clock
Checkreal_provider_calls · validate_plugin · list_forced · would_fail · check_events · verify_xml_signature

Tips for reliable runs.

  • Wait, do not sleep. wait_for_object, wait_for_api_call and wait_for_delivery return the moment the thing happens, or fail with a clear timeout.
  • Fail a set number of times, not forever. count, first and for turn a failure off by themselves, so a crashed run leaves no trap behind.
  • One environment per pull request. See Environments.
  • Ask for evidence. A good report points at the timeline, where every call carries the actor's name.