https://github.com/jcartu/rasputin-omnitool-skill
Rasputin Omnitool: OpenClaw skill bundle with planner/executor/reviewer agent loop and 12 tools (research, browser, sandbox, multimedia). Manus-equivalent agent built from OSS.
https://github.com/jcartu/rasputin-omnitool-skill
agent-skill ai-agent anthropic claude comfyui crawl4ai docling llm-agent manus multimodal-agent openclaw playwright qwen rasputin
Last synced: about 1 month ago
JSON representation
Rasputin Omnitool: OpenClaw skill bundle with planner/executor/reviewer agent loop and 12 tools (research, browser, sandbox, multimedia). Manus-equivalent agent built from OSS.
- Host: GitHub
- URL: https://github.com/jcartu/rasputin-omnitool-skill
- Owner: jcartu
- License: mit
- Created: 2026-05-09T07:40:37.000Z (3 months ago)
- Default Branch: master
- Last Pushed: 2026-05-09T08:20:02.000Z (3 months ago)
- Last Synced: 2026-05-09T10:05:26.487Z (3 months ago)
- Topics: agent-skill, ai-agent, anthropic, claude, comfyui, crawl4ai, docling, llm-agent, manus, multimodal-agent, openclaw, playwright, qwen, rasputin
- Language: Python
- Homepage: https://github.com/jcartu/rasputin-omnitool
- Size: 610 KB
- Stars: 0
- Watchers: 0
- Forks: 0
- Open Issues: 3
-
Metadata Files:
- Readme: README.md
- Contributing: CONTRIBUTING.md
- License: LICENSE
- Code of conduct: CODE_OF_CONDUCT.md
Awesome Lists containing this project
README
[](LICENSE)
[](pyproject.toml)
[](https://github.com/jcartu/rasputin-omnitool-skill/actions/workflows/tests.yml)
OpenClaw skill bundle providing agent workflows: research, browse, parse, sandbox-execute, generate multimedia (image / video / audio / music), and deliver multi-format reports.
Built on top of [rasputin-omnitool](https://github.com/jcartu/rasputin-omnitool) for catalog, library-smoke, and deliverable primitives.
> Formerly `become-manus-skill`. Renamed to `rasputin-omnitool-skill` at v0.2.0. Sprint history (audits, evidence, governance docs) under `runlog/sprint-2026-05/` is preserved as-is.
## What this skill does
Given a user goal, the skill:
1. **Plans** — Planner model emits a typed task list using only tools in the manifest.
2. **Executes** — Dispatcher runs each tool and feeds results back.
3. **Reviews** — Opus 4.7 inspects the trace and final artifacts; returns APPROVE / REVISE / ABORT.
A REVISE verdict triggers exactly one re-plan cycle. ABORT halts cleanly with a summary.
## Tools (16)
| Tool | Status | Backend |
|---|---|---|
| catalog | available | rasputin_omnitool.catalog |
| docling | available | Docling library, sandboxed paths |
| crawl4ai | available | Crawl4AI with URL safety filters |
| sandbox | available | agent-infra/sandbox HTTP API |
| browser | available | Playwright sync (5 actions) |
| deliverables | available | Parameterized rasputin_omnitool.deliverables |
| tts | available | Voxtral TTS, Kokoro fallback |
| stt | available | Whisper-large-v3-turbo |
| image_gen | available | ComfyUI + FLUX.2 [dev] |
| video_gen | deferred | Wan 2.1 (requires 96GB VRAM GPU) |
| music_gen | deferred | MusicGen-Melody (requires audiocraft venv) |
| memory | available | RASPUTIN MCP @ 8808 |
| web_search | available | SearXNG |
| slides | available | Marp CLI (PDF/HTML/PPTX) |
| mail | available | Himalaya CLI (IMAP/SMTP) |
| coding_agent | available | aider subprocess |
## Models
- **Planner**: Configurable via `RASPUTIN_OMNITOOL_PLANNER_MODEL` (default: `gpt-oss-120b`).
- **Executor**: Same as planner. Configurable via `RASPUTIN_OMNITOOL_EXECUTOR_MODEL`.
- **Reviewer**: Claude Opus 4.7 (Anthropic API). Configurable via `RASPUTIN_OMNITOOL_REVIEWER_MODEL`.
## Install
```bash
# 1. Install the kernel
cd ../rasputin-omnitool
pip install -e .
# 2. Install the skill
cd ../rasputin-omnitool-skill
pip install -e .
```
## Run a goal
```bash
./examples/run-demo.sh
# Or programmatically:
python -c "
from agent import run_goal
result = run_goal('Crawl http://example.com and produce a 1-paragraph markdown summary saved to outputs/.')
print(result['review'].verdict)
"
```
## Tests
```bash
pytest # 121 tests
```
## Observability
Real Langfuse-backed tracing. All planner / executor / reviewer / tool calls produce spans with token counts and dollar costs. Cost ceiling enforcement halts goals cleanly when `RASPUTIN_OMNITOOL_MAX_COST_USD` is exceeded.
See [docs/OBSERVABILITY.md](docs/OBSERVABILITY.md).
## Evals
Promptfoo eval harness with 10 golden tasks. CI runs on PRs with `[run-evals]` label.
See [docs/EVALS.md](docs/EVALS.md).
## User Surface
Open WebUI plugin for chat-driven goal invocation. Streams executor status to chat, returns verdict + artifacts.
See [docs/OPEN_WEBUI_SETUP.md](docs/OPEN_WEBUI_SETUP.md).
## Compose Stack
Docker Compose with 3 profiles (cpu, gpu-single, gpu-multi). Bundles sandbox, SearXNG, and ComfyUI.
See [docs/COMPOSE.md](docs/COMPOSE.md).
## Skill manifest
OpenClaw consumes `manifest.json` to know what tools the skill provides. See it for full input/output/error schemas per tool.
## License
MIT.