Latchkey

• CollectionCode Execution
• Classificationagent-native
• MCPAvailable
• Agent SkillUnavailable
• InterfacesMCP · REST · CLI · openapi · http · API

• Website ↗
• Docs ↗
• Official repo ↗
• Full dossier on GitHub ↗

What it does

Latchkey gives coding agents a place to actually execute and verify work. run_job (or latchkey run from the CLI) runs one command on a fresh ephemeral Ubuntu 24.04 x86_64 runner with no workspace and no GitHub in the path, then returns the command’s own exit code and logs, so an agent can run a test suite on a clean machine before it opens a pull request.

The same runners execute GitHub Actions jobs and diagnose failing steps while the job is still running: exit codes, then a pattern library, then a time-boxed sandboxed AI stage. Transient failures (registry timeouts, OOM kills, full disks, missing tools) are fixed on the ephemeral runner and retried; real defects are never retried or masked. What the runner cannot fix reaches the agent as a failure bundle: root cause, exact failing file and secret-scrubbed logs.

Agent-facing interfaces

• Catalog protocols: MCP, REST, CLI, openapi, http, API
• MCP: ✅ Available
• Agent Skill: ⚠️ Not yet published
• Onboarding type: mcp

How agents onboard

# Agent instructions (when to use it, how to call it, MCP / REST / CLI):
Read https://latchkey.dev/agent.txt and follow the instructions.

# MCP (Streamable HTTP):
claude mcp add --transport http latchkey https://latchkey.dev/mcp --header "Authorization: Bearer $LATCHKEY_TOKEN"

# CLI:
npm install -g @latchkeydev/cli  # then: latchkey run 'npm test'

See the full onboarding section in the dossier.

Protocol surface