// Council for business Early access

One answer your product can trust.

Your product already switches between AI models. Council tells you which answer to trust: several models from different labs work on the same task, a neutral chair compares them, and you get one answer with a confidence level. Clear cases go through. Unclear ones go to a person.

// The problem

Many models. No referee.

01

Models change every month

Most AI products keep a list of models they can switch between. But nobody knows which model gives the best answer for each request.

02

Testing them is slow

Teams test hundreds of tricky cases on every model, one by one, and have to start again whenever a new model comes out.

03

Some mistakes cost real money

In payments, a misread instruction can send money to the wrong place. One confident model is not enough to act on.

// How it works

Send a task. Get one answer.

01

Send the task

The user’s words, plus the format you need back, for example a payment intent in JSON.

02

Models work blind

Three models from different labs read it on their own, without seeing each other.

03

The chair compares

A neutral fourth model checks where they agree, and where they don’t.

04

One answer, with confidence

High confidence: go ahead. Low confidence: you get the disagreement, so you can ask the user or a person.

// How it connects

Checks the words. Then checks the transaction.

Most safety tools look only at the transaction, and most AI tools look only at the words. Council checks both, and only lets you sign when they match.

01You say it

“Send 0.5 ETH to Rahul tomorrow at 9.”

02Council reads it

Three labs agree on one exact intent: action, amount, asset, network, recipient, time.

03Your wallet builds it

Your app or wallet turns that intent into a real transaction. Council never does.

04Council checks the match

It simulates the transaction and compares the result with the agreed intent, line by line.

05You sign

Only if both match. Any mismatch or disagreement stops it and asks you first.

What you said

Send 0.5 ETH to rahul.eth, tomorrow 09:00

≠

What the transaction does

Approve unlimited USDC to 0x9f3c…a1

Stopped before signing. The transaction does not match what you asked for. This is how Council catches a wrong address, a hidden approval or a swapped token, even when the words were clear.

// Ways to plug it in

Live in preview

API & SDK

Your app sends the user’s words and the draft transaction. Council returns the agreed intent, a match result and a confidence level.

Planned

Wallet plugin

A wallet extension that shows Council’s reading right on the signing screen, next to the transaction it is about to send.

Planned

Guard for AI agents

For agents that spend money: a smart-account guard with spending limits. Small, clear payments go through. Anything unusual needs Council’s consensus and your signature.

Live today

MCP server

AI tools like Claude Code and Cursor can already call Council to check a plan or an instruction before acting on it.

  • ✓Never holds your keys
  • ✓Never sends money
  • ✓Disagreement is a hard stop
  • ✓Every check leaves a receipt

// Developer quickstart Live in preview

Keep your models. Add Council on top.

Council does not replace the AI you use. It puts several of them in one room and gives your product the answer they agree on. Two ways in, both work today with your free Council key.

Drop-in for any AI feature

Already on the OpenAI SDK? Change the base URL and the key.

from openai import OpenAI

client = OpenAI(
    base_url="https://ai-council-ashen.vercel.app/api/v1",
    api_key="ck_your_council_key",
)
r = client.chat.completions.create(
    model="council",            # or "council-quick"
    messages=[{"role": "user", "content": "Should we list this token?"}],
)
print(r.choices[0].message.content)

Read what the user meant

POST /api/v1/interpret: three labs fill the same fields

curl https://ai-council-ashen.vercel.app/api/v1/interpret \
  -H "Authorization: Bearer ck_your_council_key" \
  -H "Content-Type: application/json" \
  -d '{ "input": "send 0.5 ETH to rahul tmrw morning",
        "context": "Wallet app on Ethereum" }'

What comes back

Agreement per field, computed, not generated

{
  "intent": {
    "action": "transfer", "amount": 0.5,
    "asset": "ETH", "recipient": null,
    "when": "tomorrow morning"
  },
  "confidence": "low",
  "next_step": "ask_user",
  "question": "Please confirm the recipient.",
  "fields": {
    "amount":    { "agreement": "3/3", "status": "agreed" },
    "recipient": { "agreement": "3/3", "status": "missing" }
  },
  "notes": ["'rahul' is a name, not an address"],
  "readings": [ … one per model, side by side … ],
  "receipt": { "id": "rcpt_4f1c09…", "latency_ms": 6120 }
}

Three possible next steps

proceed

Every model agreed on every required field. Your app can build the transaction.

confirm

A required field had only a majority. Show the reading and let the person confirm it.

ask_user

Models disagreed, or something is missing. Ask the question Council wrote for you.

Test every model at once

Run your edge cases through /interpret once. Each response shows every model’s own reading next to the agreed one, so you see where models disagree or break in one pass, instead of testing them one by one.

Get a key on the sign-up page. Also available as the council_interpret tool on the MCP server.

// Where it helps

Anywhere a wrong answer is expensive.

Crypto and fintech

Turn what people type into exact payment and trading instructions, and stop before acting on anything unclear.

Customer support

Pick the right reply or route for a ticket, and send the uncertain ones to a person.

Legal and compliance

A second reading of a clause or a policy, with the disagreement shown instead of hidden.

Reading documents

Pull fields from invoices, forms and contracts, and flag the ones the models read differently.

Engineering

Review a plan or a change before you merge it. Works today inside Claude Code and Cursor through MCP.

Choosing a model

Run your hard test cases through Council once, and see where each model agrees or breaks.

// Private by design

Your data, your rules.

Today

Hosted, with no training on your data

Requests go only to the model providers that answer them, under their API terms. Council does not train on what you send.

Next

Choose which labs see your data

Pick the providers you trust for each task, and keep sensitive fields out of the request.

Later

Council on your own servers

Our own model, running offline on your hardware, so nothing ever leaves your building. The same kind of private AI that law firms already pay for.

Why Council, not one more model

Council is not tied to any AI lab, so the referee never marks its own homework. When a new model comes out, it joins the panel. Your product does not change.

You get a confidence level you can build rules on, instead of a confident tone you have to trust.

“Not a smarter model. A fair way to decide between them.”

// Questions

Before you get in touch.

Is the business API live?

Yes, in preview. /api/v1/interpret and the OpenAI-compatible /api/v1/chat/completions both work today with your free Council key, under the same fair-use limits as the website. Higher limits, a dedicated model panel and support come with early access: tell us your use case.

How fast is it?

About 10 seconds for a quick check and about 30 for a full council. For a payment flow, a short wait before money moves is usually worth it.

Which models are on the panel?

GPT, Claude and Gemini work on the task and Qwen chairs, with free backups from other labs when a model is busy. The panel can be tuned for your task.

Can Council move money on its own?

No, and that is on purpose. Council never holds keys and never sends a transaction. It reads, checks and explains. Your wallet or app builds the transaction, and a person signs it, or, for AI agents, a guard with spending limits that you set.

What does it cost?

Early-access pricing depends on volume and the models you need. Each call costs a few cents in model time, and we share that openly.

What happens to my data?

It goes only to the model providers that answer the request, and Council does not train on it. Please read the privacy policy, and talk to us before sending anything sensitive.

Building AI into something that matters?

Tell us what you are building. We will show you how Council handles your hardest cases.