{
  "markdown": "# AI Workbench\n\n<!-- mcp-name: io.github.hrishikesh-thakre/ai-workbench-mcp -->\n\n[![License](https://img.shields.io/badge/License-Apache--2.0-blue.svg)](LICENSE)\n\nAI Workbench supervises AI coding agents, captures evidence, validates work,\napplies acceptance policy, and produces auditable PR-ready reports.\n\nThe PyPI package remains `ai-workbench-mcp` for this public alpha because the\n`ai-workbench` package name is already occupied. The product and CLI are\n**AI Workbench**:\n\n```bash\npip install ai-workbench-mcp\nai-workbench --help\n```\n\nCurrent source metadata targets unpublished `ai-workbench-mcp==0.8.0a0`.\nThis public alpha consolidates local supervision, evidence capture, validation,\nacceptance policy, and PR reporting into one product surface.\n\n## Public Alpha Warning\n\nThe supervisor is the preferred automated evidence path, but daemon, Codex hook,\nand OpenCode adapter coverage are alpha mechanisms. AI Workbench checks evidence\nquality and acceptance readiness; it does not prove the work is absolutely\ncorrect. High-risk work still requires human review.\n\n## Architecture\n\n- AI Workbench supervisor captures local evidence.\n- AI Workbench validation writes `validation_report.json`.\n- AI Workbench quality gate writes `revision_decision.json`.\n- AI Workbench PR/report surfaces render `accept`, `needs_review`, or `block`.\n\nAgent output is a proposal. Workbench accepts evidence.\n\nMCP is the connection protocol. AI Workbench MCP is the tool server.\nAcceptance is decided by the selected validation profile and quality gate.\nThe agent performs. Workbench accepts. MCP connects them.\n\n## Quick Start\n\nRegister a project once and start the local supervisor:\n\n```bash\npip install ai-workbench-mcp\nai-workbench supervisor setup --project-dir . --task-type code_change\nai-workbench supervisor start\n```\n\nRun Codex, OpenCode, Goose, or another supported local workflow in the project.\nThen inspect the latest report:\n\n```bash\nai-workbench supervisor status\nai-workbench reports show latest --project-dir .\n```\n\nRender PR-ready artifacts from a finalized run:\n\n```bash\nai-workbench pr-gate --run-dir runs/<run_id>\n```\n\nThe canonical local run ledger is:\n\n```text\nruns/<run_id>/\n  task_metadata.json\n  final_prompt.md\n  model_selection.json\n  model_output.md\n  validation_report.json\n  revision_decision.json\n  run_log.jsonl\n  metadata.json\n  transcript.jsonl\n  commands.jsonl\n  workspace/\n  validation/\n  artifacts/\n```\n\n`validation_report.json` and `revision_decision.json` are the final acceptance\nauthority. Supporting supervisor reports are local evidence, not a substitute\nfor those Workbench artifacts.\n\n## Codex Hooks\n\nInstall project-local Codex hooks:\n\n```bash\nai-workbench setup codex --project-dir . --task-type code_change\n```\n\nRestart Codex or start a new session, open `/hooks`, review the project hook,\nand trust it once. Until a hook event is observed, supervisor status reports\nCodex coverage as configured but unverified.\n\n## Goose MCP\n\nAI Workbench still exposes the same MCP tool lifecycle. Register the server with\nGoose or another MCP host using:\n\n```bash\nai-workbench mcp serve\n```\n\nThe seven MCP tools remain:\n\n```text\nworkbench_open_run\nworkbench_select_policy_pack\nworkbench_select_model\nworkbench_record_execution\nworkbench_validate_run\nworkbench_quality_gate\nworkbench_analyze_runs\n```\n\n## PR Gate\n\nWorkbench PR acceptance consumes real Workbench run evidence:\n\n```bash\nai-workbench pr-gate \\\n  --run-dir runs/<run_id> \\\n  --out runs/pr_gate/pr_comment.md \\\n  --json-out runs/pr_gate/pr_decision.json\n```\n\nOutcomes are exactly:\n\n- `accept`\n- `needs_review`\n- `block`\n\nMissing, unreadable, or scaffold-only evidence blocks. A green CI run, uploaded\nartifact, sticky PR comment, or model self-claim is not acceptance evidence.\n\n## Bootstrap Assets\n\nTo add starter configs, prompts, recipes, docs, and the GitHub PR-gate workflow\nto a repository:\n\n```bash\nai-workbench bootstrap --target .\n```\n\nThe bootstrap keeps `runs/` ignored.\n\n## Package Demo\n\nFor a package-only synthetic demo:\n\n```bash\nai-workbench demo --target ./workbench-first-run\n```\n\nThis shows `accept`, `needs_review`, and `block` PR-gate outcomes with fixture\nevidence. It is not a real target-repository acceptance run.\n\n## Development\n\n```bash\npython -m pip install -e \".[dev,publish]\"\npython -m pytest -q -p no:cacheprovider\npython -m ruff check . --no-cache\npython -m mypy --no-sqlite-cache --no-incremental\nai-workbench demo --target runs/package_demo_smoke\nai-workbench validate --project ai_workbench_mcp --profile scaffold --run-dir runs/scaffold-smoke\n```\n\nDo not commit `runs/`. Committed sample evidence must be sanitized and live\nunder `examples/`.\n\n## Docs\n\n- [Supervisor docs](docs/supervisor/automated-evidence-supervisor.md)\n- [Evidence folder contract](docs/supervisor/evidence-folder-contract.md)\n- [Transcript schema](docs/supervisor/transcript-schema.md)\n- [Workspace hygiene](docs/supervisor/workspace-hygiene.md)\n- [Confidence rules](docs/supervisor/confidence-rules.md)\n- [How acceptance works](docs/concepts/how-acceptance-works.md)\n- [Contract baseline](docs/contracts/v0.2-contract-baseline.md)\n- [Package demo walkthrough](docs/walkthroughs/package-demo.md)\n- [Codex setup](docs/codex/setup.md)\n- [Codex live-test handoff](docs/codex/live-test-handoff.md)\n- [Codex acceptance walkthrough](docs/walkthroughs/codex-acceptance-demo.md)\n- [Acceptance analytics](docs/analytics/acceptance-analytics.md)\n- [Evidence dashboard](docs/analytics/evidence-dashboard.md)\n- [Event ledger](docs/analytics/event-ledger.md)\n- [Golden-case harness](docs/evals/golden-case-harness.md)\n- [Model registry](docs/configuration/model-registry.md)\n- [Dogfooding guide](docs/dogfooding/phase5-dogfooding.md)\n- [Goose demo walkthrough](docs/walkthroughs/goose-acceptance-demo.md) - recording-ready 3-5 minute public demo runbook\n- [Policy packs](docs/policy-packs/)\n- [PR gate](docs/github/pr-gate.md)\n- [Launch issues](docs/github/launch-issues.md)\n- [Repository topics](docs/github/repository-topics.md)\n- [Create launch issues](docs/github/create-launch-issues.md)\n- [Publishing guide](docs/publishing/pypi.md)\n- [Gemini fixture proof](docs/proof/gemini-fixture-accepted-run.md)\n- [Codex fixture proof](docs/proof/codex-fixture-accepted-run.md)\n\nRecipes:\n\n- [Engineering acceptance](recipes/workbench-engineering-acceptance.yaml)\n- [MCP tool smoke](recipes/workbench-mcp-tool-smoke.yaml)\n- [Docs-only acceptance](recipes/workbench-docs-only-acceptance.yaml)\n- [Python package maintenance](recipes/workbench-python-package-maintenance.yaml)\n- [Test-fix acceptance](recipes/workbench-test-fix-acceptance.yaml)\n\nSample evidence:\n\n- [Accepted tiny Python fix](examples/sample-runs/accepted-tiny-python-fix)\n- [Accepted Codex tiny Python fix](examples/sample-runs/accepted-codex-tiny-python-fix)\n- [Accepted docs-only smoke](examples/sample-runs/accepted-docs-only-smoke)\n- [Needs-review test fix](examples/sample-runs/needs-review-test-fix)\n\n## License\n\nApache-2.0. See [LICENSE](LICENSE). MIT-origin attribution for the consolidated\nProve It code is retained in [NOTICE](NOTICE).\n",
  "bytes": 7083,
  "sha": "82e6a5f0b65e5e318418e8ec213322774063e104e3739c02478254bb613d1a1d",
  "repo_slug": "hrishikesh-thakre/ai-workbench-mcp",
  "fonte": "repo",
  "truncated": false,
  "api": "https://agentalog.com/api/listings/mcp_io_github_hrishikesh_thakre_ai_workbench_1393a23c/readme"
}