orkestra

name	orkestra
description	Route LLM tasks to the cheapest model that can handle them. Use for any standalone LLM task (summarization, Q&A, classification, code generation, translation, drafting) where cost matters. Simple prompts go to budget models, complex ones to premium — automatically. Typical savings 70-80%.
user-invocable	true
metadata	{"openclaw":{"emoji":"🎼","requires":{"bins":["python3","uvicorn"]},"homepage":"https://github.com/imperativelabs/orkestra"}}

Orkestra Skill

Orkestra is a local routing proxy that sits between you and the LLM providers you already have keys for (Google, Anthropic, OpenAI). It classifies every prompt by complexity using a KNN router and automatically sends it to the cheapest model that can handle it.

Typical savings: 70–80% on mixed workloads.

---

When to Use This Skill

✅ Use Orkestra for:

• Summarization, Q&A, classification, drafting, translation, code generation
• Any standalone LLM task where cost matters
• Tasks where you want to delegate model selection automatically

❌ Do NOT use Orkestra for:

• Multi-turn conversations — the proxy only sees the last message; earlier context is lost
• Tasks that need OpenClaw tools (exec, browser, canvas) — call those tools directly
• Streaming output back to the user in real time — the proxy returns full responses only

---

Checking the Proxy is Running

curl -s http://127.0.0.1:8765/health

Expected response:

{"status": "ok", "provider": "anthropic", "multi": false}

If the proxy is not running, start it:

python3 {baseDir}/proxy.py &
sleep 2
# Then verify:
curl -s http://127.0.0.1:8765/health

---

Calling Orkestra

Send a POST to /v1/chat/completions using the standard messages format:

RESULT=$(curl -s -X POST http://127.0.0.1:8765/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [{"role": "user", "content": "Summarize the following text in 3 bullet points: ..."}],
    "max_tokens": 2048
  }')

Extract the response text:

echo "$RESULT" | python3 -c "import sys,json; r=json.load(sys.stdin); print(r['choices'][0]['message']['content'])"

Optionally log cost transparency:

echo "$RESULT" | python3 -c "
import sys, json
r = json.load(sys.stdin)
o = r['_orkestra']
print(f\"Model: {o['model']} | Cost: \${o['cost']:.6f} | Saved: {o['savings_percent']:.1f}%\")
"

---

Strategy Override

By default the strategy is whatever ORKESTRA_STRATEGY is set to in your config (usually cheapest). You can override per-request:

# Force premium model for a hard problem
curl -s -X POST http://127.0.0.1:8765/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"messages": [{"role": "user", "content": "Prove the Riemann hypothesis"}], "strategy": "smartest"}'

# Force budget model
curl -d '{"messages": [{"role": "user", "content": "What is 2+2?"}], "strategy": "cheapest"}' \
  -H "Content-Type: application/json" \
  http://127.0.0.1:8765/v1/chat/completions

# Let Orkestra balance cost and capability
curl -d '{"messages": [{"role": "user", "content": "Write a Python quicksort"}], "strategy": "balanced"}' \
  -H "Content-Type: application/json" \
  http://127.0.0.1:8765/v1/chat/completions

Strategy	Behaviour
cheapest	Cheapest model that fits the prompt complexity
balanced	Mid-tier preference, ties broken by cost
smartest	Most capable available model

---

Provider Configuration

Orkestra has no API key of its own. It routes requests through the provider credentials you already have. The proxy reads them at startup from env vars you set in your OpenClaw skills config:

// ~/.openclaw/openclaw.json — single provider (Anthropic example)
{
  "skills": {
    "entries": {
      "orkestra": {
        "enabled": true,
        "env": {
          "ORKESTRA_PROVIDER": "anthropic",
          "ANTHROPIC_API_KEY": "sk-ant-..."
        }
      }
    }
  }
}

See integrations/openclaw/README.md for Google, OpenAI, and multi-provider config examples.