memdoor

Five Qwen3.8-27B agents beat Fable 5 on LiveCodeBench Hard, as a Memdoor workflow. GVS5H's harness rebuilt with the authors' exact prompts and loop, on the first ten problems of their list, five passes each: a mean of 9.8 of 10, against 9.6 for the authors' own harness and 8.0 for Fable 5 in a single call. All 50 runs are published: every request, reply, program and grader output. The workflow is YAML and names no model: point it at yours, or at a local one with memdoor connect local (Ollama, llama.cpp, vLLM, LM Studio; how to run it on an open-weight model, no key). examples/workflows/team-solve · the result page · the HN thread

On this machine:

curl -fsSL https://memdoor.ai/install.sh | bash
export OPENROUTER_API_KEY=sk-or-...    # or ANTHROPIC_API_KEY, OPENAI_API_KEY, GEMINI_API_KEY, …
# or no key: memdoor connect local    # an open-weight model on this machine — https://memdoor.ai/docs/open-weight
cd your-project && memdoor             # the first run sets the machine up

Or on a box you can ssh to, with nothing installed there:

curl -fsSL https://memdoor.ai/install.sh | bash          # here, once
memdoor --gateway ssh://you@box                           # installs memdoor on the box, starts its gateway, opens the window on it
memdoor --gateway ssh://you@box connect openrouter        # your key, kept on the box; every turn there runs on it
memdoor --gateway ssh://you@box run "fix the failing test" # the same box from a script or CI

The runs live on the box: close the laptop, the work goes on, and your phone or the next terminal opens the same conversation there.

No account. Telemetry off unless you turn it on. macOS, Linux, Windows, one static binary, Apache 2.0.

In GitHub Actions

One step in a workflow, on your key, with nothing to add to your repo (Marketplace):

on: pull_request
jobs:
  review:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v5
      - uses: guregodevo/memdoor-action@v1
        with:
          workflow: pr-review
          openrouter-api-key: ${{ secrets.OPEN_ROUTER_API_KEY }}

Nine are built in:

• pr-triage sorts a pull request for a maintainer: ready for review, needs work, or likely slop, from checks run in the job (the tests, claimed tests that are not there, functions nothing calls).
• pr-review reviews the pull request, verifies each finding by running the code, and fails the job only on a blocker.
• security-review does the same for security only: secrets, injection, path traversal, missing authorization.
• write-tests writes tests for what the change touches, and fails the job when one shows the change is wrong.
• pr-description writes the description: title, What, Why, How to test, Risk.
• explain-failure reads a failed run's logs and the diff: the cause, file:line, and the fix.
• issue-triage suggests labels, the kind, a summary and duplicates for a new issue.
• fix-tests finds the test command, fixes the code until the tests pass, and leaves the fix in the working tree.
• release-notes turns the commits since the last tag into Added / Changed / Fixed.

No key at all: local-model: Qwen3-8B-4bit instead of the key runs any of them on a model inside the runner (janis, one file, the model cached), in one model call. It is a free first pass, not a reviewer: on a free runner it caught an obvious and a subtle bug in about a minute each, and can file a correct change's own fix as a blocker.

Or workflow: <name> for your own, .memdoor/workflows/<name>/ in the repo. The job ends with the run's outcome (0 done, 2 failed, 1 stopped, 3 stopped at an approval gate) and its summary shows the review, the fix or the notes. Trigger on pull_request, never pull_request_target: GitHub gives a fork's pull request no secrets, so a stranger's code never runs beside your key.

On a ChatGPT Plus or Pro plan? Skip the key: memdoor connect chatgpt (or /connect chatgpt in the window) opens Sign in with ChatGPT, and your plan's allowance answers Memdoor's turns. A Claude subscription cannot be used this way (Anthropic's terms); Claude needs an API key.

No key at all? The first memdoor signs you in with OpenRouter in the browser, a free account included (or with ChatGPT), from a gateway on this machine, over ssh or on a box. An account with no credit runs on a free model picked from OpenRouter's list, and the connect says so: 50 requests a day, and the free hosts train on what they are sent, so a trial, not private code. Or no key at all: memdoor connect local finds a model running on this machine (janis, Ollama, llama.cpp, LM Studio), and nothing leaves it. janis is ours: one Go server that runs the open models on a Mac's GPU through MLX or on an NVIDIA box through CUDA, serves chat and the decision model with no key, and runs Qwen3-Coder-30B-A3B on a $0.71-an-hour Google L4; memdoor --gateway ssh://you@box connect local makes that box yours. And that box answers your laptop: memdoor peers add box http://box:18789 --dir ~/app names it, and @box:coder … in any conversation here runs the turn there, on the box's own model, and brings the answer back. Recorded on memdoor.ai: a laptop with no GPU asks, Qwen3.8-27B on an L4 through janis reads, patches and runs a Go program on the box, the window on the laptop shows it land.

Every run is a conversation on a gateway on your machine, whatever door it came in by: a terminal window, memdoor run from a script or CI, a cron schedule, your editor, or the page on your phone. The run a CI job started is the one you open with memdoor resume <id>, watch live, and answer. One gateway, every door (the walk through them, with exit codes).

memdoor run "fix the failing test"        # one turn, no window; the exit code is the receipt of its checks
memdoor workflow run release --wait       # a workflow from a shell or CI, exit code the run's outcome
memdoor resume --last                     # the same run, in a window

memdoor reads your files, writes patches, runs the build and fixes what it broke, in the directory you launch it from. Every model call goes from your machine to the provider you connected, on your account, at their list price.

Two more things set it apart.

A decision model in front of the chat model. Jev (TypeSafe) answers a typed question such as is this hit relevant to the task? with a calibrated probability in under half a second, and the harness acts on it: search, file and log reads return only what counts; only the tools the turn needs are sent; a turn that stops making progress ends instead of looping to a cap. Measured over 44 paired runs, decisions off and on: 22 of 22 passed each way; a question about the codebase went from 54,661 to 14,367 input tokens (−74%); an edit with a test and ten one-file fixes read the same files either way (within ±11%) (docs/features/DECIDE.md).

Workflows. Say the steps and the coder writes them as task files, one YAML per step, run as a graph by mario. Independent steps run in parallel; a step is done when its check passes, never because the model says so; a gate waits for your approval with the diff in front of you; a failed run resumes at the failed step. The files live in your repo, so a workflow is kept, shared and run again, by hand or on your local cron (docs/features/WORKFLOWS.md). Why workflows rather than skills: the write-up. Worked examples, each run before it was added: [`examples/workflow