Works with OpenClaw
Connect PraisonAI agents and OpenClaw over MCP, ACP and shared SKILL.md skills
OpenClaw is an MIT-licensed, self-hosted assistant gateway (source). It runs isolated agents, sub-agents and Swarm, and speaks MCP and ACP in both directions, so PraisonAI connects to it through standard protocols.
Quick Start
- Let a PraisonAI agent use OpenClaw
python from praisonaiagents import Agent from praisonaiagents.mcp import MCP agent = Agent( instructions="You are a helpful assistant.", tools=MCP("openclaw mcp serve"), ) agent.start("List my OpenClaw conversations") # Prints the conversations OpenClaw exposes - Let OpenClaw call PraisonAI
bash pip install praisonai-mcp openclaw mcp add praisonai --command praisonai-mcp --arg serve --arg --transport --arg stdio
Integration Paths
OpenClaw → PraisonAI (MCP)
Run PraisonAI as an MCP server. Transports are stdio and http-stream.
pip install praisonai-mcp
openclaw mcp add praisonai --command praisonai-mcp --arg serve --arg --transport --arg stdioSee PraisonAI MCP Package. OpenClaw reference: openclaw mcp.
PraisonAI → OpenClaw (MCP)
openclaw mcp serve exposes OpenClaw conversations to any MCP client.
from praisonaiagents import Agent
from praisonaiagents.mcp import MCP
agent = Agent(instructions="Summarise my chats", tools=MCP("openclaw mcp serve"))
agent.start("Summarise my latest conversation") # Prints a short summaryOpenClaw reference: openclaw mcp.
ACP
praisonai acp is an ACP server. OpenClaw runs external ACP harnesses through acpx and accepts custom agents with --agent <command>.
See ACP. OpenClaw references: openclaw acp, ACP agents setup.
Skills
Both sides use SKILL.md with YAML frontmatter. PraisonAI parses metadata.openclaw into SkillRequirements.openclaw_hints.
See Hermes & OpenClaw Skills Import. OpenClaw reference: skill format.
Capabilities at a Glance
| Capability | PraisonAI | OpenClaw |
|---|---|---|
| MCP server | praisonai-mcp serve | openclaw mcp serve |
| MCP client | MCP(...) tool | openclaw mcp add (stdio, streamable-http) |
| ACP | praisonai acp | openclaw acp and /acp spawn |
| Multi-agent | Multi-agent workflows | Isolated agents in one Gateway |
| Sub-agents | Agent handoffs | sessions_spawn, Swarm |
| Skills format | SKILL.md | SKILL.md |
| Channels | Bot Gateway | Gateway channels (docs) |
Best Practices
Bind MCP to localhost
Use stdio, or bind http-stream to 127.0.0.1, so only local clients reach your tools.
One virtualenv per side
Install PraisonAI and OpenClaw in separate environments to avoid dependency clashes.
Pin versions
Pin praisonai-mcp and the OpenClaw release so protocol changes do not break the link.