CoachAven

A coding agent your team — or your agency — runs in Slack, Microsoft Teams or a web chat. One warm sandbox per project, on hosts you own.

• Each project gets a long-running container with its repositories and a persistent Claude Code process; a chat channel (or web room) drives it.
• The agent reads the code, answers questions, runs builds and tests, and opens pull requests; people review and merge.
• Per project, opt-in: it reviews pull requests as they open, runs routines on a schedule, investigates YouTrack tickets, and calls the MCP servers the project declares (Grafana, Sentry, trackers, your own).
• The bot is a small TypeScript control plane with no model calls of its own: all the thinking happens in the sandboxes, on your model credit — a Claude subscription token to try it, an Anthropic API key for a team or clients.

Try it on one machine — Docker and a Claude subscription; no GitHub token, Slack, registry or privileged container: docs/quickstart.md. Clone, run quickstart/coach-aven up, paste a Claude token, and the agent answers in a web chat on localhost about our demo repository — or yours, with a GitHub token for its pull requests — in about three minutes of machine time. Slack beside it is two more tokens.

The question and the agent's status report (Slack)	A question answered in a reply chain (Microsoft Teams)

Built for an agency or a small team that runs many client systems: one installation, one sandbox and one set of tools per client project, each client's work billed to a credit you choose. It grew up on Laravel projects (Sail runs inside the sandbox), YouTrack and Grafana, and runs any stack whose tools fit in a container.

Why not Anthropic's own Claude Code in Slack? Their integration runs each thread in an Anthropic-hosted sandbox tied to Claude Code on the web. CoachAven keeps a warm sandbox per project on your own Docker host, with your toolchain and the MCP servers your project declares, and adds routines, PR review and ticket investigation. If all you need is thread-to-pull-request with zero operations, their path is the better choice.

Capability	Status
Slack driver, pull requests, PR review and autofix, routines, YouTrack	production (Slack since June 2026)
Project sandboxes on your own Docker host	production for every project since v0.43.0
Microsoft Teams driver	in releases since v0.42.0; canary on a staging installation
Agent messaging between installations, through a relay hub	in releases since v0.42.0; canary between three installations
Anthropic API key as the model credit	in v0.44.0; canary on the maintainer's installation
Web chat (OIDC login; a sign-in link on localhost)	canary; the quickstart runs it
Extensions — sources (a scrubbed store of a project's mail, chat, calendar and tickets)	opt-in; production on one installation
Extensions — attribution, the lab bench	opt-in, experimental, on the maintainer's own tenant

Every change reaches the maintainer's installation from main first and production with a tag; what is staged where is the flag ledger, docs/rollout.md.

Free and open: everything here is Apache-2.0; there is no hosted tier. Rather not run it yourself? We run and support installations on a client's own infrastructure — coachaven.com — and that is what funds the project. No telemetry: an installation never reports to this project. Besides the services you configure (chat platform, GitHub, Anthropic, Docker host, MCP servers, YouTrack, sources, an error tracker or relay hub if you set one) it reads the npm registry once a day to check the Claude Code pin, and pulls its images from ghcr.io. What goes where and what the sandbox can reach: docs/security.md.

Versions: vX.Y.Z tags, while the project is below 1.0: a minor release (0.Y) may change configuration or behaviour and says so in its notes, a patch (0.Y.Z) fixes. Only the latest release is supported (SECURITY.md); main is the canary. Every release is the very image the canary ran.

Status: a single-maintainer project with installations for real users; self-hosting is documented, support is best-effort (SUPPORT.md, CONTRIBUTING.md, SECURITY.md).

Start here:

• Why CoachAven: who it is for, what it does differently, what it is not.
• How it is built and run: invariants, specs first, canary then tag, who does what — built with Claude Code, every change specified first (OpenSpec) and canaried.
• All docs: running an installation, what the agent can do, under the hood.

以上为 README 的开头部分。其余内容请阅读原文。 查看完整 README ↗