REA 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)
import OpenAI from "openai";
const client = new OpenAI({
apiKey: "YOUR_REA_API_KEY",
baseURL: "https://api.reacontrolplane.com/v1",
defaultHeaders: { "X-Upstream-Api-Key": "YOUR_MODEL_KEY" },
});
const resp = await client.chat.completions.create({
model: "gpt-4o-mini",
messages: [{ role: "user", content: "Reply with exactly: REA LIVE" }],
});
console.log(resp.choices[0].message.content);
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"}]}'
Same shape as an OpenAI chat completion, including streaming. Your REA key and your vendor key both go to REA. Use a header, or put them in one API key with a dot between them. The vendor key is used for that call and is not stored. REA keeps no backup vendor key. You pay that vendor directly. Claude Code, Codex, and Gemini CLI use their own paths. Set up each one.
Your agent still chooses the work. These three decide what is allowed to leave, and each one has its own page.
LoopGuard
A retry of broken code can be another vendor bill. LoopGuard hashes the structure of the code, not the wording of the prompt. A new structure is probed in a Firecracker microVM. If the probe finds the loop, the model is not called. If the pool cannot run the probe, the check is skipped and the call can still be forwarded. A later repeat of a stopped structure is the stored stop. Unique chat still pays the vendor.
How LoopGuard decidesPreFlight
PreFlight checks the deny rules you declared. A secret marker on a declared path is refused the same way. The prompt is not read for that check. A match costs $0, and the vendor is never called. A clean call continues on the same hop.
How PreFlight decidesTraceGuard
Lock the names this key may use, or keep up to 25 lists and apply one. A repeated tool is Contain. The first attempt to leave the sandbox is Contain. Kill is a second escape, or a secret-shaped call after the session was already contained. Contain still calls the model. Kill returns HTTP 403 traceguard_killed and ends the hop. Until you lock, every tool is recorded and the call is not stopped. A name off a locked list is Direct, and Direct does not stop the call. No message and no tool arguments are stored.
How TraceGuard decides
REA does not host the model, store the vendor key, or keep a backup key. The vendor bills you for the calls that are forwarded.
LoopGuard, PreFlight, and TraceGuard are on every plan. They are not an add-on, and a refused or stopped call is not a forwarded call.
Each forwarded call returns the outcome, the model, and the token counts. Message text and tool arguments are not stored. The field names are on Architecture.
Names identify products a user can point at REA. They are not partners, and the marks are not an endorsement. Bedrock and OpenRouter have their own routes. Agents and CLIs each have a page.
Your agent
Hermes, Cursor, Continue, Cline, OpenCode, Aider, or an SDK.
POST /v1/chat/completions
Authorization is Bearer rea_, or the REA key alone with X-Upstream-Api-Key.
REA ControlPlane
https://api.reacontrolplane.com
Your vendor
OpenAI, Anthropic, xAI, Gemini, Mistral, DeepSeek, Groq, Together, Fireworks, Amazon Bedrock, or OpenRouter. Your key. Your bill.
A non-stream answer comes back with a small rea_meta receipt. Normal streams do not carry it.
You do not change models. You point the base URL at REA. An SDK uses that same address. REA does not ship one. Claude Code, Codex, and Gemini CLI use their own paths. SDKs are on Set up.
REA sits between your agent and your model vendor. You set a tool list for a key. A repeated tool is Contain. The first attempt to leave the sandbox is Contain. Kill is a second escape, or a secret-shaped call after the session was already contained. Contain still calls the model. Kill returns HTTP 403 traceguard_killed and ends the hop. Until the list is locked, the call is not stopped. PreFlight checks policy before the call. LoopGuard is the cost control: a repeated code structure is not sent again.
No. REA ControlPlane is a product of REA Research Labs, Inc. It is not the product at controlplane.com, and it is not a Kubernetes control plane. The short description is on Facts.
A forwarded call is one request REA sends to your vendor. A PreFlight refusal, a LoopGuard stop, and a TraceGuard Kill are not counted. Contain still calls the vendor, so it counts. You pay your vendor for tokens on the calls that go through. When the forwarded-call allowance runs out, REA returns 402. Going over the plan's requests per minute returns 429.
REA replays the stored stop from an earlier LoopGuard decision. It never replays a vendor answer, and a cache hit does not create a new vendor call.
Signup asks for your name, work email, company, and a password. Phone and role are optional. Stripe collects the billing address. The trial needs a card and lasts 7 days. Calls stop at 2,000 forwarded calls, and the card is not charged early. Cancel before the 7 days end and the card is not charged. You can start Developer Pro or Growth Team today instead. The key is shown once, so copy it immediately. The monthly letter is a separate list. An account does not join it unless the signup box is checked. The box is not checked.
Hermes, Claude Code, Codex CLI, Cursor, Gemini CLI, Cline, Continue, OpenCode, Kilo Code, Roo Code, Aider, Goose, Qwen Code, Zed, and Pi. Claude Code, Codex, and Gemini CLI use their own paths and a combined key. Cursor chat uses the chat path. Tab completion in Cursor does not. The others use the chat path. Subscription logins are refused. The setup page has the settings.
Yes. The probe on api.reacontrolplane.com runs in a Firecracker microVM. If the pool cannot run it, the code check is skipped and the call can still be forwarded. The status name is on Architecture.
There is no universal savings percentage. A repeated failure can be answered from the stored stop. A new failure the probe does not stop still calls the model. Results depend on your traffic, models, and repeat patterns.