Memdoor

Memdoor is an Apache-2.0 coding and automation agent delivered as a single CLI binary. It can run on the local machine or through SSH to a gateway on another machine, then continue the same run from a terminal, editor, script, CI job, schedule, or phone. Its useful distinction is durable work rather than a one-off chat: a job can contain check-gated steps, approval points, and a resume path after an interruption.

What it does

Memdoor documents a headless-first workflow. memdoor run can start an agent turn from a shell or CI, while memdoor resume reopens that recorded run. Project workflows live as YAML files in the repository. Independent steps may run in parallel; a step is only complete when its declared check passes, and an approval gate pauses before the next step. Built-in workflows cover pull-request triage and review, security review, test writing, failure explanation, issue triage, test repair, and release notes.

The runtime can use a configured provider, compatible endpoint, or a local model connection. The project says that files, sessions, and commands remain on the operator's machine, while model-visible material is sent to the selected provider. Its decision model is intended to limit irrelevant context before a chat-model call, but the project's benchmark claims are maintainer measurements and should be validated on the repository and workload that matter to you.

A practical workflow

Start in a disposable repository or branch. Install from the official instructions, connect a provider or a local runtime, then create a narrow workflow such as pr-review: require the test command to pass, place the final patch behind an approval gate, and retain the generated run record. In GitHub Actions, use the documented action only on pull_request, not pull_request_target, so untrusted fork code does not receive secrets. Resume a failed run locally and inspect the diff and checks before accepting an agent-written change.

Costs, limits, and alternatives

Memdoor itself is open source under Apache-2.0. Model inference, remote infrastructure, and any provider account remain the operator's responsibility. The documentation warns that free hosted models can train on submitted inputs; do not use that route for private code. A coding agent can read files, execute commands, and alter a working tree, so use least-privilege credentials, isolated test environments, and explicit review gates.

• Codex CLI is OpenAI's terminal coding agent for supported accounts and API workflows.
• Claude Code is Anthropic's terminal agent for repository work.
• CowAgent is a broader self-hosted personal-agent runtime with channels and memory.

Conclusion

Choose Memdoor when you want an agent workflow to be durable, inspectable, and runnable from CI or a scheduler, not only interactive in a terminal. Begin with one reversible task and a real check, then decide whether its gateway and provider model fit your security boundary. Read the official headless workflow guide before connecting a production repository.