REA Your agent retries
Coding tools send broken code, get an answer, try again. Each try can be another paid model call — even when the mistake is the same infinite loop with different comments.
Stop paying for the same failed loop twice.
REA sits between your coding agent and the model. It stops runaway code in about 150 milliseconds, fingerprints the shape of the failure, and answers repeats from cache so the model is not called again. You bring your own key. We never bill OpenAI or Anthropic tokens.
API live · 1.4.8-prod
Bring your own key
Structural memory
~150 ms stop
Firecracker when the host allows it
In plain language
REA watches the shape
It looks at the structure of the code, not the wording of the prompt. The same trap hashes to the same fingerprint.
Repeats skip the model
The first time, the model still answers. The next time that structure appears, REA returns the stored answer. That is the savings. Unique work still pays the vendor.
What is on today
A public API
HTTPS at api.reacontrolplane.com. Same shape as OpenAI chat completions. Health reports INTEGRATED.
Two keys, always
A REA key opens the plane. Your model key rides in X-Upstream-Api-Key. REA does not keep a backup vendor key for customer traffic.
A hard stop
Runaway snippets are killed at about 150 ms. An empty monthly loop pool returns 402. A burst over your tier returns 429.
A receipt on every call
Responses include rea_meta: cache hit or miss, the fingerprint, and whether the sandbox blocked, crashed, or saw no code.
Where it sits
Your agent (Continue, Cline, or an SDK)
|
| POST /v1/chat/completions
| Authorization: Bearer <REA key>
| X-Upstream-Api-Key: <your model key>
v
https://api.reacontrolplane.com
|
v
REA ControlPlane
checks your REA key and monthly loop budget
fingerprints code structure
returns cache on a known trap
otherwise probes the snippet (~150 ms)
|
v
OpenAI or Anthropic — your key, your bill
|
v
Vendor answer, plus a small rea_meta receipt
You do not change models. You change the base URL and add one header.
What it is not
Not a model
No training. No rewritten answers. The first answer you see is the vendor’s.
Not a single savings number
Cache helps when the same failure repeats. New failures still call the model. We do not publish a universal percentage.
Not Cursor-first
The supported path is Continue, Cline, or any client that can set a custom base URL and header.