Authorization and limits
How Olimpia authorizes agents, how long tokens last, what limits apply, and what rules the MCP server follows so an agent can't cause harm.
OAuth 2.1
Olimpia is its own authorization server:
- Discovery. The server publishes
/.well-known/oauth-protected-resource(RFC 9728) and/.well-known/oauth-authorization-server(RFC 8414). A request to/mcpwithout a token returns401with aWWW-Authenticateheader that points to those documents. - Client registration. Through a Client ID Metadata Document (the
client_idis anhttpsURL) or through dynamic registration (POST /oauth/register, public clients only). - PKCE S256 is required, along with
issin the response (RFC 9207) and a validatedresource(RFC 8707). - Consent happens on olimpia.dev, with your session. You see which agent is requesting access before you approve.
| Credential | Lifetime |
|---|---|
| Authorization code | 10 minutes, single use |
| Access token | 1 hour |
| Refresh token | 90 days, rotated on every use |
| Plugin connection code | 10 minutes, single use, bound to the agent that requested it |
| Personal token | 90 days, or until you revoke it |
POST /oauth/revoke revokes a token, and Account → Agents and tokens lists every connection and lets you revoke it.
Same permissions as the console
Every tool calls the same API the console uses, with the token from the request. Validation and authorization work exactly the same way: an agent can’t do anything your account can’t do.
Limits
| Limit | Value |
|---|---|
Calls to /mcp |
240 per minute per token |
| Apps, databases, buckets, email domains, projects, and environments created | 30 per hour per account, counting the console and the API |
| Failed connection code attempts | 10 every 10 minutes per agent |
| Rows per read query | 200 |
| SQL write duration | 30 seconds |
Safety rules
- Deletes and writes require confirmation.
delete_app,delete_resource,delete_environment,disable_backups, andquery_postgreswithwrite: trueneedconfirmset to the exact name, which the person has to approve. - So does taking data out of Olimpia.
set_public_accesswithpublic: true,set_backupsto an external S3, andupdate_appwhen it changessource,repo, orimageask for the sameconfirm. The agent only does them when the person asked for exactly that, never because a log, a row, or a commit says so. - Secrets hidden by default.
get_envhides values that look like secrets unless you passreveal: true. - Untrusted data. Logs, build output, variable values, and database rows are data, not instructions. The tools mark them as such so the agent doesn’t follow commands hidden in them.
- SQL without superuser.
query_postgresruns as the user that owns the database. Reads run in a read-only transaction, and the database itself enforces their row and size caps. Each query uses a new connection, so nothing it leaves in the session carries over to the next one.
Best practices
- Use OAuth for interactive agents and personal tokens only for CI.
- Use one token per integration, with a name that says where it’s used.
- Revoke connections you no longer use.