Point the agent at the plane.

What you need

The two headers

Authorization: Bearer YOUR_REA_API_KEY
X-Upstream-Api-Key: YOUR_MODEL_KEY
Content-Type: application/json

For Claude, add X-REA-Upstream: anthropic, or use a claude-* model name. The Anthropic key still goes in X-Upstream-Api-Key.

A first call

curl -s https://api.reacontrolplane.com/v1/chat/completions \
  -H "Authorization: Bearer YOUR_REA_API_KEY" \
  -H "X-Upstream-Api-Key: YOUR_MODEL_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-4o-mini",
    "messages": [{"role":"user","content":"Reply with exactly: REA LIVE"}]
  }'

Python

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_REA_API_KEY",
    base_url="https://api.reacontrolplane.com/v1",
    default_headers={"X-Upstream-Api-Key": "YOUR_MODEL_KEY"},
)

resp = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": "Reply with exactly: REA LIVE"}],
)
print(resp.choices[0].message.content)

Health

curl -s https://api.reacontrolplane.com/health

You want "status": "INTEGRATED", "byok": true, version 1.4.8-prod.

Continue or Cline

1. Choose an OpenAI-compatible provider.
2. Set the base URL to https://api.reacontrolplane.com/v1.
3. Set the API key to your REA key.
4. Add a custom header X-Upstream-Api-Key with your model key.

Those values live in the tool config. There is no website login. A 402 means the monthly loop pool is empty — the agent is not logged out.

What each request does

1. Check the REA key. Apply the rate window and loop budget.
2. If the messages carry code, hash the structure.
3. Redis hit → return the stored completion. No model call.
4. Otherwise probe the snippet against the ~150 ms ceiling.
5. On a trap, the GPT path may attach logit bias for the next generate.
6. Forward the same messages with your upstream key.
7. Return the vendor text plus rea_meta.

LoopGuard fixtures

Fences are optional. json.dumps is required. Continue and Cline often send a fenced block. Hermes and raw SDKs often send raw text. Prose that only mentions while True is NO_CODE, not a trap and not a 422.

Raw (benches / SDK):

{
  "model": "gpt-4o-mini",
  "messages": [{
    "role": "user",
    "content": "import time\nwhile True: pass"
  }]
}

Fenced (Continue):

{
  "model": "gpt-4o-mini",
  "messages": [{
    "role": "user",
    "content": "```python\nwhile True:\n    pass\n```"
  }]
}

First call: HTTP 200, mode=block, BLOCKED_TIMEOUT, about 150–160 ms. Repeat: HTTP 200, mode=cache, cache_hit=true, same structural_hash.

The small receipt: rea_meta

Status codes