Agent Quickstart

The one path that works today, start to finish. Every step below was executed against the running system before it was written down; the table at the end says where each one is checked.

This is the procedure. For what the platform is, the quota table and the honest list of limits, read For AI Agents or its plain-text original at /llms.txt — that page is the reference and this one does not restate it.

Two of the six steps are not yours to do. They are marked, and no amount of retrying will move them to your side.

Step 1 — a person creates the accounthuman

Account creation happens in the browser at meta-council.com. There is one signup step with two mechanisms — an email address and password, or continuing with a Google account. Either produces the same kind of account; nothing later in this guide depends on which was used.

Note for anyone scripting this: /signup is an alias that redirects to the homepage, where the form lives. It is not a separate page to fetch and parse.

Step 2 — that person mints your keyhuman

Three things about the key that will otherwise surprise you later:

• It is shown once. The full key is returned at the moment of creation and never again — only its short prefix is visible afterwards. If it is lost, the fix is a new key, not a lookup.
• It expires. Keys are created with an expiry, 90 days by default. A credential that worked for months and then stops has probably not been revoked; it has aged out. Ask for a fresh one.
• It carries scopes. A key created with the defaults can run councils and read their results, which is everything this guide needs. Tools outside that set — settings, outreach and the rest — need a key minted with the matching scope, and refusing them is not a bug.

Step 3 — connectyou

Nothing to install:

claude mcp add --transport http meta-council https://meta-council.com/mcp --header "Authorization: Bearer mc_YOUR_KEY"

Speaking the protocol directly instead: POST JSON-RPC to https://meta-council.com/mcp with Accept: application/json, text/event-stream, beginning with an initialize call. Bearer keys are the only way in — MCP OAuth is not implemented.

Step 4 — confirm the endpoint before the key mattersyou

Call list_panels. It takes no arguments and needs no credential, which makes it the cheapest way to prove you are talking to the right endpoint and that your transport works. It returns every expert panel with its description, tags and agent count; the slugs it hands back are what panel accepts in the next step.

Do the same with tools/list to see what your own credential actually reaches, rather than trusting any list written on a web page. This page deliberately names no catalogue size for that reason.

Step 5 — run a councilyou

Call run_council. Only query is required:

{"name": "run_council", "arguments": {"query": "Should we migrate off the monolith this quarter?"}}

Optional arguments: panel (a slug from step 4, or "auto" to let the platform choose), model, and wait_seconds — an integer from 0 to 90.

Treat this as asynchronous. wait_seconds defaults to 0, which returns a session id immediately rather than holding the connection open. That default exists to stay clear of reverse-proxy timeouts, so raising it is not free: a synchronous wait is capped at 90 seconds and a panel can take longer than that. Take the session id and move to step 6.

This is the first step that requires your key. Without one you get a JSON-RPC error telling you to add the bearer header — that response means the endpoint is fine and the credential is missing, which is a different problem from a malformed call.

Step 6 — read the resultyou

Call get_session with the id from step 5:

{"name": "get_session", "arguments": {"session_id": "<the id from step 5>"}}

It returns the finished result: each agent's recommendation, confidence and full reasoning, plus the synthesis, consensus points, dissenting views and risk matrix. Poll it rather than assuming a fixed delay — a panel's duration depends on how many agents it has and which model served it.

What can stop you

Four refusals you should expect to meet, and what each one actually means. All four are the platform telling you something specific — none of them is a reason to retry blind.

What this page does not cover

Where each step was verified

Every row was executed or read before the step above was written. The guard file walks this same sequence against the real routes, so these are checks that keep running rather than a record of one afternoon.

Repository paths are given so the claims are auditable by anyone who can read the source. The repository is private, so they are provenance rather than a link you can follow.