c69fbd8a4d
* Add OpenClaw + Plano intelligent routing demo Demonstrates preference-based routing for personal AI assistants: Kimi K2.5 handles conversation and agentic tasks, Claude handles code generation, testing, and complex reasoning — with zero application code changes and ~48% cost savings. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Remove redundant provider_interface from Kimi K2.5 config The openai/ prefix in the model name already sets the provider interface. Setting provider_interface explicitly conflicts with it and fails config validation. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Simplify config to v0.3.0 format, remove explicit Arch-Router entry Arch-Router is implicit when routing_preferences are defined. Aligns with the preference_based_routing demo pattern. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Clean up Ollama/Arch-Router references, make Jaeger optional Router is handled internally by Plano — no need for Ollama or explicit Arch-Router setup. Jaeger is kept as an optional step in the README for developers who want tracing visibility. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Remove run_demo.sh, use planoai CLI directly The planoai CLI already handles startup. README now uses planoai up/down directly instead of a wrapper script. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Remove docker-compose.yaml, use inline docker run for Jaeger No need for a compose file when Jaeger is the only optional service. A single docker run command in the README is simpler. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Clarify testing: OpenClaw channels vs direct Plano requests Primary testing is through messaging channels (Telegram, Slack, etc.) with log monitoring. The test_routing.sh script is now documented as an optional direct verification tool. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Add OpenClaw onboarding instructions to README Includes install, onboarding wizard, channel setup, doctor check, and how to point the gateway at Plano. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Use OpenClaw onboarding wizard for Plano provider setup Replace manual JSON config with instructions to use the openclaw onboard wizard to set up a custom OpenAI-compatible provider pointing at Plano. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fixed readme and removed unnecessary testing.sh file --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com> Co-authored-by: Salman Paracha <salmanparacha@MacBook-Pro-342.local>
5 KiB
OpenClaw + Plano: Smart Model Routing for Personal AI Assistants
OpenClaw is an open-source personal AI assistant that connects to WhatsApp, Telegram, Slack, and Discord. By pointing it at Plano instead of a single LLM provider, every message is automatically routed to the best model — conversational requests go to Kimi K2.5 (cost-effective), while code generation, testing, and complex reasoning go to Claude (most capable) — with zero application code changes.
Architecture
[WhatsApp / Telegram / Slack / Discord]
|
[OpenClaw Gateway]
ws://127.0.0.1:18789
|
[Plano :12000] ──────────────> Kimi K2.5 (conversation, agentic tasks)
| $0.60/M input tokens
|──────────────────────> Claude (code, tests, reasoning)
Plano uses a preference-aligned router to analyze each prompt and select the best backend based on configured routing preferences.
Prerequisites
- Docker running
- Plano CLI:
uv tool install planoaiorpip install planoai - OpenClaw:
npm install -g openclaw@latest - API keys:
MOONSHOT_API_KEY— from Moonshot AIANTHROPIC_API_KEY— from Anthropic
Quick Start
1. Set Environment Variables
export MOONSHOT_API_KEY="your-moonshot-key"
export ANTHROPIC_API_KEY="your-anthropic-key"
2. Start Plano
cd demos/llm_routing/openclaw_routing
planoai up --service plano --foreground
3. Set Up OpenClaw
Install OpenClaw (requires Node >= 22):
npm install -g openclaw@latest
Install the gateway daemon and connect your messaging channels:
openclaw onboard --install-daemon
This installs the gateway as a background service (launchd on macOS, systemd on Linux). To connect messaging channels like WhatsApp or Telegram, see the OpenClaw channel setup docs.
Run openclaw doctor to verify everything is working.
4. Point OpenClaw at Plano
During the OpenClaw onboarding wizard, when prompted to choose an LLM provider:
- Select Custom OpenAI-compatible as the provider
- Set the base URL to
http://127.0.0.1:12000/v1 - Enter any value for the API key (e.g.
none) — Plano handles auth to the actual providers - Set the context window to at least
128000tokens
This registers Plano as OpenClaw's LLM backend. All requests go through Plano on port 12000, which routes them to Kimi K2.5 or Claude based on the prompt content.
If you've already onboarded, re-run the wizard to update the provider:
openclaw onboard --install-daemon
5. Test Routing Through OpenClaw
Send messages through any connected channel (WhatsApp, Telegram, Slack, etc.) and watch routing decisions in a separate terminal:
planoai logs --service plano | grep MODEL_RESOLUTION
Try these messages to see routing in action:
| # | Message (via your messaging channel) | Expected Route | Why |
|---|---|---|---|
| 1 | "Hey, what's up? Tell me something interesting." | Kimi K2.5 | General conversation — cheap and fast |
| 2 | "Remind me tomorrow at 9am and ping Slack about the deploy" | Kimi K2.5 | Agentic multi-step task orchestration |
| 3 | "Write a Python rate limiter with the token bucket algorithm" | Claude | Code generation — needs precision |
| 4 | "Write unit tests for the auth middleware, cover edge cases" | Claude | Testing & evaluation — needs thoroughness |
| 5 | "Compare WebSockets vs SSE vs polling for 10K concurrent users" | Claude | Complex reasoning — needs deep analysis |
OpenClaw's code doesn't change at all. It points at http://127.0.0.1:12000/v1 instead of a direct provider URL. Plano's router analyzes each prompt and picks the right backend.
Tracing
For fast dev/test cycles, Plano provides built-in tracing to visualize routing decisions and LLM interactions. Start the trace listener in a separate terminal:
planoai trace
Then send requests through OpenClaw. You'll see detailed traces showing:
- Which model was selected and why
- Token usage and latency for each request
- Complete request/response payloads
Learn more about tracing features and configuration in the Plano tracing guide.
Cost Impact
For a personal assistant handling ~1000 requests/day with a 60/40 conversation-to-code split:
| Without Plano (all Claude) | With Plano (routed) |
|---|---|
| 1000 req x Claude pricing | 600 req x Kimi K2.5 + 400 req x Claude |
| ~$3.00/day input tokens | ~$0.36 + $1.20 = $1.56/day (~48% savings) |
Same quality where it matters (code, tests), lower cost where it doesn't (chat).
Stopping the Demo
planoai down