skill
ClawCross
Unify agents from multi-platform (OpenClaw, Codex, Claude, Gemini) into one Team with shared prompts & skills — then drop the whole agent workflow into any Python script.
Set up
Skills are installed by placing their folder where your agent loads skills.
The download comes from github.com/ClawCross/ClawCross itself, not from our server, so what you get is exactly what its author published.
Install instructions come from the README
README
synced Oct 21, 2026Clawcross
A local AI workspace where multiple expert agents collaborate, debate, and execute — with a visual workflow engine, living memory, and one-click public access.
What is Clawcross?
Clawcross turns a single chatbot into a programmable multi-expert system. You create a Team — a group of AI agents with distinct roles and personas — and let them collaborate on tasks through visual workflows. Each discussion builds a living knowledge graph that persists across sessions.
Key concepts at a glance:
| Term | Meaning |
|---|---|
| Agent | Any agent on your machine — WeBot, Codex, Claude Code, Gemini, OpenClaw, or an OpenAI-compatible service — behind one interface (/v1/agents) |
| Team | A named set of agents, each playing a role (one may lead), plus the team's personas, workflows and skills |
| OASIS | The visual workflow engine that orchestrates multi-expert discussions (sequential, parallel, branching, DAG) |
| OASIS Town | A pixel-town visualization in the Studio sidebar where you watch live discussions and inspect the swarm graph |
| WeBot | A Claude-Code-style delegated runtime with role-based subagents, plan/todo/verification, and approval-aware tool policies |
| GraphRAG | A living knowledge graph built from each discussion, stored locally in SQLite (optionally mirrored to Zep) |
| Team Presets | 15 ready-to-use expert teams — strategists, content creators, tech titans, and more — installable in one click |
| ACP (acpx) | Agent Client Protocol for communicating with external AI agents (OpenClaw, Codex, Claude, Gemini, Aider) |
| OpenClaw | An agent runtime whose agents join Teams, group chats and workflows like any other agent |
Product Video
Quick Start
Prerequisites
- Python 3.11+, prepared by the startup wrapper through
uvif absent - Node.js 18+ only for optional acpx and frontend builds
- Git
- macOS / Linux / Windows (WSL or PowerShell)
Install via AI Code CLI
Open any AI coding assistant such as Codex, Cursor, Claude Code, CodeBuddy, or Trae, and say:
Clone https://github.com/ClawCross/ClawCross.git, read AGENTS.md, and install Clawcross.
Manual Setup
Fastest start
bash run.sh start
powershell -ExecutionPolicy Bypass -File .\run.ps1 start
Open the UI at http://127.0.0.1:51209.
Stop / status
bash run.sh status
bash run.sh stop
powershell -ExecutionPolicy Bypass -File .\run.ps1 status
powershell -ExecutionPolicy Bypass -File .\run.ps1 stop
Notes
run.sh/run.ps1are root-level wrappers around the canonical scripts underlaunch/.startprepares Python 3.11, a virtual environment, core dependencies,.env, and services. Useinstall-componentonly for optional integrations; see SKILL.md.- If
LLM_MODELis missing, the web UI still starts; configure a model in the first-login wizard before sending LLM requests. SetCLAWCROSS_REQUIRE_LLM_MODEL=1to make startup strict.
Canonical launch commands
bash launch/run.sh start
bash launch/run.sh status
bash launch/run.sh stop
powershell -ExecutionPolicy Bypass -File .\launch\run.ps1 start
powershell -ExecutionPolicy Bypass -File .\launch\run.ps1 status
powershell -ExecutionPolicy Bypass -File .\launch\run.ps1 stop
For the full install guide (OpenClaw, Antigravity, MiniMax, WSL, manual CLI config, troubleshooting), see SKILL.md.
Why Clawcross
Multi-Expert Collaboration, Not Just Chat
- Team-based orchestration — combine WeBot, Codex, Claude Code, OpenClaw and API agents into a single Team with one-click import/export
- 16 built-in Team Presets — LLM Council, Nuwa All-Stars, Content Empire, Strategists, Tech Titans, and more — install and run immediately
- Teams through conversation — Ask your main Agent to manage members, personas, workflows and schedules with optional management tools
- Visual orchestration — design workflows in OASIS with sequential, parallel, branching, or DAG-style expert coordination
Claude-Code-Style Delegation with WeBot
- Role-based subagents — profiles for general, research, planner, coder, reviewer, and verifier modes
- Persisted state — run/task lifecycle, plan/todo/verification primitives, context compression, and artifact logging
- Approval-aware tool policies — configure which tools require human approval, with event logging and hooks
Living Memory and Observability
- GraphRAG — every discussion builds a living knowledge graph, persisted locally with optional Zep mirroring
- OASIS Town — watch discussions unfold in a pixel-town visualization, inspect the swarm graph, and inject nudges in real time
- ReportAgent — ask why the current prediction leans a certain way and get graph-backed evidence
Real Operator Features
- OpenAI-compatible API — local
/v1/chat/completionsendpoint that works with standard clients and MCP tools - Bots — Telegram and QQ integrations with whitelist access control
- Multimodal I/O — images, files, voice input, TTS, provider-aware audio defaults
- Automation — scheduled tasks and long-running workflow execution
- TinyFish — internet search agent powered by TinyFish Web Agent API
- Remote access — Hub device subdomains with login-token / password flows; optional Cloudflare
- Flow distribution — browse and share workflows on ClawcrossHub
What You Can Do Today
| Capability | What It Gives You |
|---|---|
| OpenAI-compatible API | Local chat completions endpoint for apps, tools, and clients |
| Web UI | Chat, settings, OASIS panel, group chat, tunnel control, WeBot runtime panel |
| Team Presets | 15 ready-to-use expert teams — install and start collaborating immediately |
| Conversational Team management | Let the main Agent create Teams and configure members, workflows and schedules |
| OASIS workflows | Sequential, parallel, branching, and DAG-style expert orchestration |
| OASIS Town | Pixel-town visualization with live residents, nudges, and swarm graph |
| WeBot runtime | Claude-Code-style delegation with profiles, modes, tool policies, and per-Agent runtime state |
| GraphRAG memory | Persist each topic as a living graph in local SQLite, with optional Zep mirroring |
| ReportAgent | Graph-backed evidence for predictions and decisions |
| MCP tools | Built-in tools for commands, files, sessions, search, scheduler, OASIS, WeBot, and LLM API |
| Team system | Public/private agents, personas, workflows, and Team snapshots with import/export |
| ACP communication | Broadcast to OpenClaw, Codex, Claude, Gemini, Aider via Agent Client Protocol |
| Multimodal I/O | Images, files, voice input, TTS |
| Bots | Telegram and QQ integrations |
| Automation | Scheduled tasks and long-running workflow execution |
| TinyFish | Internet search agent for web crawling and structured data extraction |
| Flow distribution | Browse and share flows on ClawcrossHub |
| Remote access | Hub device subdomains plus login-token / password flows; optional Cloudflare |
Flow Distribution Platform
ClawcrossHub is the companion flow distribution platform:
- Browse published Clawcross flows
- Distribute reusable workflows to other users
- Share flow links as a lightweight workflow catalog entry
Typical Use Cases
- Local AI workspace — run a private AI assistant with a browser UI and OpenAI-compatible API
- Team debate and execution — let multiple experts challenge, refine, and conclude on the same task
- Ready-made expert councils — install a preset team (LLM Council, strategists, content creators) and run it immediately
- Live debate observability — watch an OASIS discussion from the Town sidebar, inspect the swarm graph, and inject nudges
- Delegated agent work — use WeBot to delegate research, coding, or review tasks to role-based subagents
- Prediction / GraphRAG cockpit — use OASIS topics as living world models with evidence-backed report answers
- AI integration hub — connect bots, external agent runtimes, and other OpenAI-compatible clients
- Competitor monitoring — schedule daily pricing crawls, compare stored snapshots, and detect changes
- Operational cockpit — manage settings, ports, audio, workflows, public access, and users from one place
Acknowledgements
Clawcross benefited from several open-source projects:
msitarzewski/agency-agents— inspiration for expanding the preset expert poolAGI-Villa/agent-town— reference for OASIS Town's interaction and presentation designtanweai/pua— inspiration for upgrading the critical expert into a stronger PUA-style reviewer
Documentation
| File | Audience | Purpose |
|---|---|---|
AGENTS.md | AI Agents | Behavior rules, task router, progressive disclosure |
SKILL.md | Agents + Humans | Complete install, config, debug, and troubleshooting guide |
docs/index.md | Both | Task-based documentation map to all deep-dive docs |
Community
- Issues & feature requests: GitHub Issues
- Flows & presets: ClawcrossHub
License
Apache License 2.0 — see LICENSE.