๐ 1st Place Winner โ Z.ai Bounty @ UK AI Agent Hackathon ร OpenClaw ๐ฆ.
๐ Finalist Winner โ Anyway Bounty @ UK AI Agent Hackathon ร OpenClaw ๐ฆ.
Give 0xClaw a hackathon URL โ it researches, codes, tests, and submits, all on its own.
conda create -n 0xclaw python=3.11 -y
conda activate 0xclaw
pip install -e .
cp .env.example .env
./scripts/verify_setup.shThen fill in ZAI_API_KEY or FLOCK_API_KEY in .env.
For a quick sanity check after installation:
python launcher/__main__.py --help
0xclaw --helppip install -e . installs both the runtime dependencies and the local CLI in editable mode.
If you also want development tooling such as Ruff, use pip install -e .[dev].
requirements.txt is kept as a compatibility mirror of the same runtime dependency set.
conda activate 0xclaw
0xclaw # interactive CLI
0xclaw --logs # with debug output
0xclaw gateway # start Telegram/other chat channels from repo config
0xclaw whatsapp login # start WhatsApp bridge and scan QR0xclaw is the canonical runtime entrypoint.
./scripts/start.sh is an optional convenience wrapper that activates conda, loads .env, then runs 0xclaw.
./scripts/verify_setup.sh is a preflight checker.
Legacy scripts:
scripts/run_phase.py: single-phase engineering/debug runner.scripts/hackathon_runner.py: deprecated compatibility path; avoid for normal workflow.
Integration docs:
- Claude Code coding backend:
docs/integrations/claude-code-subagent.md. - Docs index:
docs/README.md.
The agent runs a 7-phase pipeline to produce a complete hackathon submission. Trigger any phase with natural language โ routing is automatic.
| # | Phase | Output |
|---|---|---|
| 1 | research | hackathon/context.json |
| 2 | idea | hackathon/ideas.json |
| 3 | selection | hackathon/selected_idea.json |
| 4 | planning | hackathon/plan.md ยท hackathon/tasks.json |
| 5 | coding | hackathon/project/ |
| 6 | testing | hackathon/test_results.json |
| 7 | doc | hackathon/submission/ |
Some prompts for quick understand the processe:
research the hackathon requirements
generate project ideas
plan the architecture
implement the project
run tests
write the documentation
/status | Pipeline progress + session token usage |
/resume | Resume from last checkpoint |
/redo <phase> | Reset phase and all downstream, then re-run |
/new | Clear all outputs, fresh start |
/stop | Cancel running task |
/help | Show all commands |
? | Alias for /help |
!<cmd> | Run a shell command without leaving the agent (e.g. !git status) |
Long-running phases hand off to the background after the first reply โ you can keep chatting while work continues.
Ctrl+C interrupts the current task but never exits โ type /exit to quit.
Inference powered by FLock.io and Z.AI
| Phases | Model | Context |
|---|---|---|
| research ยท idea ยท selection ยท doc | minimax-m2.1 | 196k |
| planning ยท coding ยท testing | minimax-m2.5 | 205k |
Per-phase timeout and fallback behavior are defined in 0xclaw/config/model_profiles.json (source of truth).
The runtime has built-in support for 10 messaging platforms (2/10 now passed solid test). Enable any via config.json + one env var:
| Discord | Slack | Feishu | ||
| DingTalk | Matrix | WeChat Work |
Telegram uses long polling โ no public IP or webhook required.
Configure it in 0xclaw/config/config.json under channels.telegram, then run 0xclaw gateway.
WhatsApp uses a local bridge process. Configure channels.whatsapp, run 0xclaw whatsapp login to scan the QR code, then start 0xclaw gateway. Bridge assets and auth state are stored under ~/.0xclaw/.
