Memdoor
An open-source CLI coding agent and workflow runner in a single Go binary, with a Team Mode that coordinates multiple model agents.
At a Glance
Apache-2.0 licensed Memdoor, run on your laptop or company machines for free. No account. BYOK model costs are paid to providers.
Engagement
Available On
Listed Nov 2026
About Memdoor
Memdoor is an open-source (Apache-2.0) CLI coding agent and workflow runner distributed as one Go binary. It reads files, edits code and runs checks, and works with a local model or a provider on the user's own key. The project's homepage and repository describe a terminal window, scripted runs, editor integration and remote access to the same gateway.
What It Is
Memdoor is a coding agent plus a deterministic workflow harness. In the terminal window, you give it a task and it reads what it needs, patches code, builds and runs tests before answering. Shift+Tab switches between auto mode and team mode. For repeatable work, a workflow describes checked steps that run as a graph, with independent steps in parallel and approval gates.
How Workflows Work
You describe the steps and the coder writes one YAML file per step. A step counts as done when its target exists, meaning a file is present or a command exits 0. A gate pauses the run for approval with the diff in view, and a comment sends the step back. A failed run resumes at the failed step. Workflows can be started by hand, from CI, or on local cron. Bundled examples include issue-to-pr, release-check, a digest, a deep-research fan-out and team-solve.
Team Mode and the Decision Model
Team-solve is a workflow in which a planner, two ideators, a solver and a manager each work from a fresh context and communicate through files. The project reports that a team of five Qwen3.8-27B agents scored a mean of 9.8 of 10 on the first ten LiveCodeBench Hard problems from the GVS5H list, against 9.6 for the authors' harness and 8.0 for a single Fable 5 call, and publishes all 50 runs. It notes that these are ten problems, not the authors' full 100, and that contamination cannot be ruled out. Separately, a hosted decision model (Jev, by TypeSafe) judges which search hits and file sections the coding model reads. The project reports 74% fewer input tokens on a codebase question in paired runs.
Ways to Reach It
The same gateway on your machine can be reached from the TUI, from a shell or CI with memdoor run (exit codes report whether checks passed), from a remote box over ssh, from VS Code, Zed and JetBrains over the Agent Client Protocol, from a schedule, and from a phone via an end-to-end encrypted /remote link. A GitHub Action runs built-in workflows such as pr-review and security-review. Model access is via OpenRouter, OpenAI, Gemini, DeepSeek, Baseten, Groq, xAI, Anthropic, a ChatGPT plan, or local runtimes such as Ollama, llama.cpp, vLLM, LM Studio and janis.
Community Discussions
Be the first to start a conversation about Memdoor
Share your experience with Memdoor, ask questions, or help others learn from your insights.
Pricing
Open source
Apache-2.0 licensed Memdoor, run on your laptop or company machines for free. No account. BYOK model costs are paid to providers.
- The terminal agent: reads, edits, runs the tests, fixes what it broke
- Workflows with a proof per step and a gate
- One gateway, every door: window, memdoor run, schedule, editor, ssh, phone
- BYOK: OpenRouter, OpenAI, Gemini, DeepSeek, Baseten, Groq, xAI, Anthropic or a company gateway, paid to them at list price; or a ChatGPT Plus/Pro plan
- Decision model on your key
Enterprise
Self-hosted on your company's machines, supported by us. By invoice.
- Gateway on a box in your network
- Approval mode and tool guards
- Every host the binary can name is listed; no usage ping
- Full support from a named person
Capabilities
Key Features
- Single Go binary CLI coding agent with terminal UI
- Team Mode via Shift+Tab with multi-agent team-solve workflow
- YAML workflows run as a DAG with parallel steps, gates, resume and cron scheduling
- Steps complete when a target file exists or a command exits 0
- Decision model that judges reads to reduce input tokens
- Headless runs with exit codes for scripts and CI
- Editor integration over Agent Client Protocol (VS Code, Zed, JetBrains)
- Remote control from a phone with end-to-end encrypted /remote link
- Remote gateway over ssh and peer machines
- MCP server support
- Subagents and scheduled checks
- GitHub Action with built-in review, triage and test workflows
- Local model support with no key
- Git worktree mode and approval before changes
