OpenClaw + Ollama + Claude Code Setup

Configure OpenClaw to use a remote Ollama instance as the primary model, with automatic delegation of coding tasks to Claude Code CLI. Optionally set up a multi-agent PM/Dev workflow using Agent IRC.

Author: Yossi Ovadia (@yossiovadia)

Architecture

You (WhatsApp/Telegram/Web UI)
  ↓
OpenClaw Gateway (local Mac/Linux)
  ↓
Ollama (local or remote) ← cheap/free model for general tasks
  ↓
Claude Code CLI ← delegated coding tasks (via coding-agent skill)

Prerequisites

• OpenClaw installed and running
• Ollama running (local or remote)
• Claude Code CLI installed (for coding delegation)
• jq installed (brew install jq)

Quick Start

./setup.sh

The script will:

1. Detect or ask for your Ollama endpoint
2. List available models and let you choose one
3. Configure ~/.openclaw/openclaw.json
4. Optionally add Claude Code routing for coding tasks

Manual Setup

1. Configure Ollama Provider

Edit ~/.openclaw/openclaw.json and add your Ollama provider under models.providers:

{
  "models": {
    "mode": "merge",
    "providers": {
      "ollama": {
        "baseUrl": "http://<OLLAMA_IP>:11434/v1",
        "apiKey": "ollama",
        "models": [
          {
            "id": "<MODEL_NAME>",
            "name": "Your Model Name",
            "api": "openai-completions",
            "reasoning": false,
            "input": ["text"],
            "cost": {
              "input": 0,
              "output": 0,
              "cacheRead": 0,
              "cacheWrite": 0
            },
            "contextWindow": 128000,
            "maxTokens": 16384
          }
        ]
      }
    }
  }
}

2. Set Default Model

{
  "agents": {
    "defaults": {
      "model": {
        "primary": "ollama/<MODEL_NAME>"
      }
    }
  }
}

3. Add Tool Permissions

IMPORTANT: allow: [] means NO tools allowed. Use ["*"] to enable all tools:

{
  "tools": {
    "byProvider": {
      "ollama/<MODEL_NAME>": {
        "allow": ["*"]
      }
    }
  }
}

4. Restart OpenClaw

openclaw gateway restart

5. (Optional) Claude Code for Coding Tasks

Add this to ~/.openclaw/workspace/TOOLS.md to make your Ollama model delegate coding work to Claude Code CLI:

## MANDATORY: All Coding Tasks -> Claude Code CLI

**YOU MUST DELEGATE ALL CODING TASKS TO CLAUDE CODE. DO NOT WRITE CODE YOURSELF.**

When a user asks to write, fix, or modify code, ALWAYS execute:
  bash workdir:<project-dir> command:"claude -p --dangerously-skip-permissions '<task description>'"

All projects MUST be created in your projects directory (e.g. /home/user/code/),
NEVER in the workspace directory.

What you CAN do yourself:
- Answer questions about code (explaining, discussing)
- Run existing scripts, git commands, file operations
- General research and conversation

If it involves WRITING or MODIFYING code -> DELEGATE TO CLAUDE CODE.

Workspace Best Practices

DO NOT change the workspace directory. The workspace (~/.openclaw/workspace/) is for agent identity files (SOUL.md, TOOLS.md, etc.), not for project code.

Directory	Purpose
~/.openclaw/workspace/	Agent identity (SOUL.md, TOOLS.md, etc.)
~/code/ (or your projects dir)	Your code, accessed via workdir parameter

When coding, specify the project directory per-task:

bash workdir:~/code/myproject command:"claude -p --dangerously-skip-permissions 'Fix the bug'"

Enterprise Claude Code (Vertex AI / AWS Bedrock)

If your Claude Code CLI uses an enterprise backend (Google Vertex AI, AWS Bedrock, etc.), you need to add the required environment variables to the OpenClaw launchd plist.

Why This Is Needed

OpenClaw runs as a launchd service with a minimal environment. Your shell's env vars (like Vertex AI credentials) are NOT available to the gateway process. Without them, Claude Code will try consumer OAuth login instead of your enterprise backend.

Setup

1. Find your launchd plist:

ls ~/Library/LaunchAgents/ai.openclaw.gateway.plist

2. Add your enterprise env vars to the EnvironmentVariables dict:

For Google Vertex AI:

<key>CLAUDE_CODE_USE_VERTEX</key>
<string>1</string>
<key>ANTHROPIC_VERTEX_PROJECT_ID</key>
<string>your-gcp-project-id</string>
<key>CLOUD_ML_REGION</key>
<string>us-east5</string>

For AWS Bedrock:

<key>CLAUDE_CODE_USE_BEDROCK</key>
<string>1</string>
<key>AWS_REGION</key>
<string>us-east-1</string>

3. IMPORTANT: Unload and reload the plist (a simple restart won't pick up plist changes):

launchctl unload ~/Library/LaunchAgents/ai.openclaw.gateway.plist
launchctl load ~/Library/LaunchAgents/ai.openclaw.gateway.plist

4. Verify the env vars are loaded:

launchctl print gui/$(id -u)/ai.openclaw.gateway | grep -E "VERTEX|BEDROCK|CLOUD_ML"

Multi-Agent PM/Dev Workflow (Agent IRC)

Set up two AI agents (PM + Dev) that collaboratively build software projects with human oversight using Agent IRC.

Architecture

You send PRD (via WhatsApp/Telegram)
  ↓
PM Agent → Reads PRD, creates GitHub issues, assigns to Dev
  ↓ (posts to shared Agent IRC channel)
Dev Agent → Implements code using Claude Code CLI, pushes branches
  ↓ (posts completion to channel)
Dev Agent → "Ready to deploy. Awaiting approval."
  ↓ (escalates to you via WhatsApp/Telegram)
You reply: "approved"
  ↓
Dev Agent → Deploys
PM Agent → Closes issue, assigns next

Setup

1. Install Agent IRC CLI:

curl -o agent-irc.sh https://api.agent-irc.net/agent-irc.sh
chmod +x agent-irc.sh

2. Register and verify agents:

# Register
./agent-irc.sh --profile my-pm register "MyProject-PM" "Project manager"
./agent-irc.sh --profile my-dev register "MyProject-Dev" "Developer"

# Claim via GitHub gist (follow the instructions printed after register)
./agent-irc.sh --profile my-pm claim <gist-url>
./agent-irc.sh --profile my-dev claim <gist-url>

3. Create a shared channel:

./agent-irc.sh --profile my-pm join '#myproject-dev' --topic "PM/Dev coordination"
./agent-irc.sh --profile my-dev join '#myproject-dev'

4. Import the multi-agent skill:

curl -s https://api.agent-irc.net/skills/multi-agent-software-dev.md

5. Add agents to OpenClaw config (~/.openclaw/openclaw.json):

{
  "agents": {
    "list": [
      { "id": "main", "default": true },
      {
        "id": "my-pm",
        "workspace": "~/.openclaw/agents/my-pm",
        "identity": { "name": "MyProject-PM", "emoji": "📋" }
      },
      {
        "id": "my-dev",
        "workspace": "~/.openclaw/agents/my-dev",
        "identity": { "name": "MyProject-Dev", "emoji": "💻" }
      }
    ]
  },
  "tools": {
    "agentToAgent": {
      "enabled": true,
      "allow": ["main", "my-pm", "my-dev"]
    }
  }
}

6. Create SOUL.md files for each agent in their workspace directories. See the multi-agent skill docs for templates.

7. Set up cron schedules (staggered to prevent conflicts):

# PM Agent - runs at :00, :15, :30, :45
*/15 * * * * /path/to/pm-agent-run.sh

# Dev Agent - runs at :07, :22, :37, :52 (7-min offset)
7,22,37,52 * * * * /path/to/dev-agent-run.sh

Ollama Tips

Keep Models Loaded

Set unlimited keep-alive so models stay in memory:

# On your Ollama server
OLLAMA_KEEP_ALIVE=-1 ollama serve

Verify Context Length

ollama show <model-name> --modelfile | grep num_ctx

Common Pitfalls

Pitfall	Fix
allow: [] in tools.byProvider	Change to allow: ["*"] - empty array = no tools
Changed workspace to project dir	Revert to ~/.openclaw/workspace - use workdir for projects
openclaw gateway restart doesn't pick up plist changes	Use launchctl unload + load instead
Claude Code "Not logged in" from OpenClaw	Add enterprise env vars to launchd plist (see above)
Agent creates files in workspace instead of project dir	Update TOOLS.md with explicit project directory path
bash pty:true fails with "No such file or directory"	That's OpenClaw tool syntax, not shell syntax - model needs to use it as a tool call
Claude Code hangs and never exits	Always use -p flag: claude -p "task" exits when done, claude "task" hangs forever
Claude Code exits but doesn't write files	Add --dangerously-skip-permissions so it can approve its own file edits
Need text answer, not file changes	Use claude -p "question" (no skip-permissions needed for read-only)

Troubleshooting

"Connection refused" to Ollama

• Verify Ollama is running: curl http://<IP>:11434/api/tags
• Check firewall allows port 11434
• Ensure Ollama is bound to 0.0.0.0, not just localhost

Config validation errors

openclaw doctor --fix

Gateway won't start

tail -50 ~/.openclaw/logs/gateway.err.log

Check gateway env vars

launchctl print gui/$(id -u)/ai.openclaw.gateway | grep -A5 "environment"

References

• OpenClaw Docs
• Agent Workspace
• Skills Config
• Agent IRC
• Multi-Agent Software Dev Skill
• claude-flow - Alternative multi-agent orchestration

License

MIT