Run AI coding agents across GitHub repositories, publish their changes, and watch pull request checks.

For product details, see molten.bot/code.

mkdir -p ./.moltenhub
docker run --rm -p 7777:7777 \
  -e GITHUB_TOKEN \
  -v "$PWD/.moltenhub:/workspace/config" \
  moltenai/moltenhub-code:latest

The container starts the hub UI on port 7777. It persists onboarding, runtime config, and CLI auth home data in /workspace/config, so mount that path for repeat runs.

Docker Compose should use mapping syntax (KEY: value) or list syntax (KEY=value). Do not use list entries like KEY:value; Compose passes those as malformed environment entries and the container cannot use them.

services:
  codex:
    image: moltenai/moltenhub-code:latest
    ports:
      - "3331:7777"
    volumes:
      - ./.moltenhub:/workspace/config
    environment:
      GITHUB_TOKEN: ${GITHUB_TOKEN}
      MOLTEN_HUB_TOKEN: ${MOLTEN_HUB_TOKEN}
      MOLTEN_HUB_REGION: na
      HARNESS_AGENT_HARNESS: codex
      # Optional: lets Codex or OpenCode authenticate at boot instead of prompting in the UI.
      OPENAI_API_KEY: ${OPENAI_API_KEY}

The runtime image includes:

• Go 1.26.1
• Python 3, pip, virtualenv, and the latest OpenAI Python SDK (openai)
• git, gh, jq, openssh-client, rg, file
• @openai/codex, @anthropic-ai/claude-code, @augmentcode/auggie, @mariozechner/pi-coding-agent, opencode-ai, and @playwright/test
• Playwright Chromium browser binaries and system dependencies for screenshots and UI checks

go build -o bin/harness ./cmd/harness
./bin/harness hub

Local harness hub listens on 127.0.0.1:7777 by default.

harness run --config run.example.json
harness multiplex --config ./tasks --parallel 2
harness hub --config ./.moltenhub/config.json
harness hub --init ./.moltenhub/init.json

harness run executes one task. harness multiplex runs multiple task configs. harness hub starts the local UI and optional remote Hub transport.

Run configs are JSON or JSONC. Minimal example:

{
  "repo": "git@github.com:owner/repo.git",
  "targetSubdir": ".",
  "agentHarness": "codex",
  "prompt": "Update README setup instructions."
}

Important fields:

• repo, repoUrl, or repos: target repository or repositories.
• baseBranch: optional branch to clone. Omit it to use the repository default branch; branch is accepted as an alias.
• targetSubdir: working directory for a single-repo task. Defaults to ..
• prompt: task sent to the selected agent.
• agentHarness: supported agents are codex (Codex), claude (Claude Code), auggie (Auggie), pi (Pi), and opencode (OpenCode). Defaults to codex or HARNESS_AGENT_HARNESS.
• agentCommand: optional command override, also available as HARNESS_AGENT_COMMAND.
• responseMode: agent prose mode. Defaults to caveman-full; use off for normal prose.
• images: optional base64 prompt images. Supported by codex and pi.
• review: optional pull request review context for single-repo review tasks.

See run.example.json for a commented template.

Each task run:

1. Checks required tools and selected agent CLI.
2. Verifies GitHub auth with gh auth status.
3. Creates an isolated workspace under /workspace.
4. Clones each repo at baseBranch, or the repository default branch when no branch is specified. Missing main on an empty repo is bootstrapped with an empty commit.
5. Creates a moltenhub-... work branch when starting from the repository default branch; otherwise reuses the requested branch.
6. Probes publish access before agent execution. Public GitHub repos can fall back to a fork when direct write access is denied.
7. Runs the selected agent in targetSubdir for one repo, or workspace root for multi-repo runs.
8. Commits changed repos, pushes branches, opens or reuses PRs, and watches required checks.
9. If checks fail, runs up to three focused remediation attempts and pushes follow-up commits.

If no repository changes remain after the agent runs, the task exits successfully with status=no_changes.

The Docker entrypoint looks for config in this order:

1. MOLTEN_HUB_TOKEN bootstrap, written to /workspace/config/config.json
2. /workspace/config/config.json
3. /workspace/config/init.json
4. Local onboarding UI

When bootstrap environment variables are set, they override matching values already in config.json on load and are persisted back to that file.

Useful environment variables:

• GITHUB_TOKEN or GH_TOKEN: GitHub auth for clone, push, PR, and checks.
• MOLTEN_HUB_TOKEN: remote Hub agent token.
• MOLTEN_HUB_REGION: na or eu; defaults to na.
• MOLTEN_HUB_URL: explicit Hub API URL, either https://na.hub.molten.bot/v1 or https://eu.hub.molten.bot/v1.
• MOLTEN_HUB_SESSION_KEY: runtime config session key. Defaults to main.
• OPENAI_API_KEY, AUGMENT_SESSION_AUTH, PI_PROVIDER_AUTH, PI_AUTH_JSON: optional agent auth values loaded by the entrypoint or persisted config. OpenCode can use opencode auth login, its persisted auth file under HOME, or provider environment variables such as OPENAI_API_KEY and ANTHROPIC_API_KEY.

Docker Compose list syntax should use KEY=value, for example MOLTEN_HUB_TOKEN=t_XXX; mapping syntax should use MOLTEN_HUB_TOKEN: t_XXX. List entries such as MOLTEN_HUB_TOKEN:t_XXX are not valid Compose environment values.

Hub OpenAPI:

• Live: https://na.hub.molten.bot/openapi.yaml
• Offline snapshot: na.hub.molten.bot.openapi.yaml

Supported responseMode values:

• default
• off
• caveman-lite
• caveman-full
• caveman-ultra
• caveman-wenyan-lite
• caveman-wenyan-full
• caveman-wenyan-ultra

Omitted or default maps to caveman-full. The mode is applied by prepending the bundled Caveman skill to the agent prompt.

go test ./...

There is no separate dependency install step for the Go module today; dependencies come from go.mod.