What is Memdoor?
Memdoor is an open-source coding agent and workflow runner distributed as one Go binary for macOS, Linux and Windows. In a project folder it reads files, writes patches, runs the build or the tests and fixes what it broke, with the model of the user's choice. The source repository is published by the GitHub account guregodevo under the Apache License 2.0, and the project site at memdoor.ai hosts the installers and the documentation. The repository was created in October 2026, and releases use a version scheme in the 1.x range with frequent builds.
The project describes its headline feature as Team Mode: pressing Shift+Tab in the terminal window switches a turn from a single agent call to the built-in team-solve workflow, in which a planner, two ideators, a reviewer and a judge work from fresh contexts on files written to disk. According to the project, five agents on the open-weight Qwen3.8-27B model reached a mean of 9.8 of 10 on the first ten problems of a LiveCodeBench Hard list over five passes, against 8.0 for Fable 5 in a single call, with all 50 runs published in the repository. The project itself notes that ten problems are a sample and that contamination cannot be ruled out. The team-solve result page documents the method, which rebuilds the GVS5H multi-agent harness with its authors' prompts.
Every run is a conversation on a local gateway, a background process on port 18789 that holds conversations, schedules and workflow runs. Several entry points open onto the same gateway: the terminal window (memdoor or memdoor tui), headless runs with memdoor run, local cron, editors over the Agent Client Protocol, GitHub Actions, a phone through /remote, and remote boxes reached over SSH with memdoor --gateway ssh://you@box. A hosted decision model, Jev by TypeSafe, answers typed questions about each turn so that only relevant search hits, file sections and log lines reach the chat model.
Memdoor does not sell model access. Turns run on the user's own provider key, a company AI gateway, a ChatGPT Plus or Pro plan through Sign in with ChatGPT, or a local runtime such as janis, Ollama, llama.cpp, vLLM or LM Studio. The project states that code does not pass through memdoor.ai. The agent, workflows, schedules and remote control are free; an Enterprise offer adds setup and support from the maintainers on a company's own machines, by invoice, through the Enterprise page.
The latest GitHub release is Memdoor v1.828.0, published on November 13, 2026. It ships binaries for darwin-amd64, darwin-arm64, linux-amd64, linux-arm64 and windows-amd64 with a SHA256SUMS file, and the release notes state that the install script picks the matching binary and checks it against those sums. Building from source needs Go 1.25. Similar terminal coding agents listed on iThub include Aider, Codex CLI and goose; Archon also runs coding agents from YAML workflows.
Key capabilities
- Single binary installer
curl -fsSL https://memdoor.ai/install.sh | bashplaces one binary at~/.local/bin/memdoorwith no sudo, Docker or Python. The download is checked against the published SHA-256 sums, and whenminisignis installed the manifest signature is verified as well. Windows usesirm https://memdoor.ai/install.ps1 | iex, which checks the checksum only. A GitHub-release install path is documented for forks or fixed versions. - Terminal window The window shows each tool call (locate, read, edit, bash) as a frame, a diff for every edit and a closing receipt line such as checked, files changed and the test command with PASS. Esc interrupts, ctrl+o expands a frame, ctrl+f opens the files the conversation touched with previews and diffs, and ctrl+g draws the turn as a graph of reads, edits and runs. On macOS an image on the clipboard can be pasted in.
- Team Mode Shift+Tab toggles between auto and team. In team, three direct attempts run first, each writing a check; a pick step keeps the change that most checks pass, and only when the Check fails do two ideators propose fixes for a solver. The direct attempt is saved before the team touches it and is put back unless the team's work passes a check it did not, as described in the Team mode guide. The project puts the cost at about two to three times a single call.
- Workflows as a deterministic harness A workflow is one YAML file per task under
.memdoor/workflows/<name>/, run as a graph by the mario engine. Task types areagent,commandandllm; a task is done when its target exists or its command exits 0, not when the model says so. Independent tasks run in parallel, a task marked external is an approval gate, and a failed run resumes at the failed step. The coder can write the files from a plain description, as shown in the workflows guide. - Built-in workflows The release includes pr-triage, pr-review, security-review, write-tests, pr-description, explain-failure, issue-triage, fix-tests, release-notes, team-solve, team-backlog, qa and pr-create. team-backlog takes the next open issue, runs team-solve on it in a branch and opens a pull request when the Check passes. pr-create opens a draft pull request with
ghand never merges. - Decision model Jev by TypeSafe, a hosted model, answers typed questions about a turn with a calibrated probability in under half a second, according to the project. It drives judged reads, a per-turn toolbox that only sends the tool schemas a turn needs, and a stop for turns that stop making progress. It runs on an OpenRouter key or a dedicated decision key, or locally on janis; on an Anthropic, OpenAI or Gemini key alone it is off. In the project's 44 paired runs, a codebase question used 74 percent fewer input tokens while edits changed within 11 percent.
- Provider choice and model ladder Supported connections include OpenRouter, Anthropic, OpenAI, Gemini, DeepSeek, Baseten, Groq, xAI, EUrouter, LLM Gateway, Vercel AI Gateway, AI/ML API, any OpenAI-compatible endpoint and a company AI gateway. Each agent has a ladder of models, cheapest rung first, and
/modelpins any tool-capable model at the provider's list price. A Claude subscription cannot be used through sign-in; Claude needs an API key. Details are on the providers page. - Local and open-weight models
memdoor connect localinstalls or finds janis, the project's own Go model server for MLX on Apple GPUs and CUDA on NVIDIA, or connects a running Ollama, llama.cpp, vLLM or LM Studio. No key is needed and nothing leaves the machine. janis picks a model by RAM, Qwen3.5-9B at 16 GB and Qwen3.8-27B from 32 GB, and also serves the decision model locally. The open-weight guide covers the setup. - Headless runs and exit codes
memdoor runtakes a prompt, writes the answer to stdout and returns 0 when checked, 2 when unverified or a check fails, 3 when a question went unanswered, 1 on error and 130 when interrupted.--stdintakes piped input such as a diff, and--jsonemits events. Any headless run can be reopened live withmemdoor resume, as described in Headless, then attached. - GitHub Actions The
guregodevo/memdoor-action@v1action runs a built-in or project workflow in one step with a provider key as the secret and adds nothing to the repository. Exit codes are 0 done, 2 failed, 1 stopped and 3 stopped at a gate, and the job summary holds the review or fix. Alocal-modelinput runs a small model inside the runner through janis; the project describes this as a first pass that can flag a correct change as a blocker. - Editor integration over ACP
memdoor acpserves the coder over the Agent Client Protocol. Zed and JetBrains host it natively, and VS Code uses an ACP client extension with one setting. Text, tool calls and approvals appear in the editor's own interface, per the editor guide. - Remote control from a phone
/remoteprints a link and a QR code that open the same terminal in a phone browser, end-to-end encrypted, with the key kept after the # in the link. The gateway serves the page on its own address over Wi-Fi, VPN or Tailscale with no account; a gateway kept to localhost relays sealed frames through memdoor.ai./remote viewcreates a watch-only link for up to 200 viewers. The remote control page notes that the full link grants control of the agent. - Gateway on a remote box
memdoor --gateway ssh://you@boxinstalls Memdoor on a machine reachable over SSH, registers a systemd user or launchd service and opens the window there, so runs continue with the laptop closed. Provider keys can be stored on the box. Peers let one gateway run a turn on another named gateway and return the answer. - Scheduled checks and cron The agent can schedule a check, such as polling CI every 20 seconds, and the answer wakes the conversation. Limits are documented: at most 60 runs, 12 hours and no faster than every 10 seconds.
memdoor cron addruns a prompt or a whole workflow on local cron. - MCP servers
/mcpadds Model Context Protocol servers from a URL, a command line or a.mcp.jsonsnippet, or from a search of the official MCP registry. Servers can be scoped to the project or to every project, and secrets are referenced as variables rather than written to the file. - Boundaries, guards and approval mode File writes stay inside the launch folder, while shell commands can reach the rest of the machine; the project states it is not a sandbox. Secrets printed in tool output are masked before the model sees them.
memdoor guards addrefuses tool calls matching a regular expression, andMEMDOOR_APPROVE=changesasks before every command, file write and MCP call, as listed under Boundaries. - Usage and savings reports
/usageshows this month's turns, tokens and cost on the user's key per model, andmemdoor savingsreports what the decision model kept out of the bill.memdoor meterlists every turn with the model served and its tokens.
Where Memdoor fits
Fixing a bug with a test in a local project
The user runs memdoor in the project folder and types the task, for example rejecting an invalid discount with an error and adding a test. The agent locates the relevant code, reads the files, writes the patch, runs the test command and reports a receipt line with the files changed and the test result. The footer names the model that answered and its rung on the ladder.
Escalating a hard task to Team Mode
When a single call is not enough, Shift+Tab switches the window to team and the next prompt runs as the team-solve workflow on the same model. The window draws each step with its time, tokens and cost, and the run reports the verdict, what was kept, the scope and the diff. From a shell the same run is memdoor workflow run team-solve --model <id> "<task>".
Reviewing pull requests in CI
A GitHub Actions job with guregodevo/memdoor-action@v1 and workflow: pr-review runs the review with findings checked by running the code, and the job fails only on a blocker. The project documents triggering on pull_request rather than pull_request_target, since a fork's pull request gets no secrets. Other built-ins cover security review, test writing and failure explanation, as listed in the CI guide.
Working an issue backlog overnight
memdoor cron add --workflow team-backlog "0 2 * * *" schedules one open issue per night. The workflow picks an issue nobody is working on, runs team-solve in a branch and opens a pull request when the Check passes, or comments on the issue when it does not.
Running on open-weight models without a key
memdoor connect local installs janis and a model sized to the machine's RAM, tests it and pins it, so turns and the decision model run on local hardware. The project's guide continues to running the team-solve workflow on a local GPU, and lists Ollama, llama.cpp, vLLM and LM Studio as alternative runtimes for the chat model.
Following a run from a phone or a remote box
A gateway installed on a VPS or home server over SSH keeps runs going after the laptop closes. /remote or memdoor remote --last produces a link to the same terminal on a phone, where the user can watch tool frames, answer a question, approve a gate or stop a turn. /remote off revokes the link.
Install
curl -fsSL https://memdoor.ai/install.sh | bashReviews of Memdoor
Be the first to share your experience with Memdoor.
Write a review
Loading review form…
Found Memdoor useful? Upvote it todayYour vote helps other IT teams find it in the daily ranking.





