Human-approved sign-in
Runfiredrill login. The CLI starts a device/browser authorization flow with
WorkOS. A person approves the identity and organization; the client then exchanges
that approved session for Firedrill authority.
The CLI and native SDKs use a protected owner-only login store outside your
repository. They preserve pending exchange and refresh state for exact recovery.
Do not delete that state to retry a network failure.
WorkOS identity tokens belong only at the sign-in boundaries. They are not general
Firedrill API credentials. If your signed-in identity or organization changes,
stop the client, verify the intended context, and reconnect deliberately. Recover
existing work under its original authority.
Coding-agent MCP sign-in
Connect your MCP host tohttps://mcp.firedrill.run/mcp, then use its OAuth
sign-in action. A person approves access through WorkOS in the browser. The host
stores the resulting OAuth credential; it does not belong in a prompt,
repository MCP file, or recipe. See native host setup.
The MCP host and CLI have separate sign-in sessions. Authorize the CLI or SDK
when it needs to run the actual application locally, using the same intended
organization. Do not copy a host’s credential into another client. Authentication
never starts the application or approves its model-provider credentials.
Scoped credentials for CI
Use a scoped machine credential for unattended execution. An authorized human sign-in withcredentials:manage can issue one through POST /v1/credentials or
the CLI’s firedrill credentials commands.
Choose one project/environment pair, an explicit capability subset, and an expiry.
A machine credential cannot issue or manage other credentials, and cannot receive
more authority than its issuer.
Grant only the capabilities your workflow needs. Tool-bound simulations need the
applicable Tool capabilities as well as run authority. Declaring a capability in a
recipe or config file grants nothing.
Store the issued token directly in your CI secret store as
FIREDRILL_TOKEN.
Credential lists return metadata, not secrets. Exact issuance replay has a bounded
protected secret-recovery window; if it has ended, inspect and revoke that
credential before intentionally issuing a replacement.
Native client configuration
Both clients connect tohttps://api.firedrill.run. Select the exact IDs returned
by the project list or saved by firedrill init.
FIREDRILL_TOKEN for automation, or use protected sign-in when no explicit token
is supplied. An explicitly supplied invalid token is not silently replaced by a
different sign-in. Keep credentials out of firedrill.config.json.
For a direct API request:
Tool grants are not API credentials
Your target receives only the Tool connections selected for its run: exact instance, epoch, audience, allowed operations, and lifetime. It must not receive the coordinating API credential, controller-only operations, or evaluation rubric. HTTP, MCP, and browser audiences are checked separately. A Tool token cannot read another run’s report, manage projects, or issue credentials. Current parent authority, actor ownership, instance state, and operation grants are checked at issuance and use. Reusable Tool connections return their token or browser launch URL once. Replaying connection creation returns retained metadata without the secret. If the first response was lost, inspect and explicitly revoke that connection before issuing another with its own operation key. Resetting an instance advances its epoch and invalidates prior connections; extending compute lifetime does not renew them.Expiry and revocation
On401, inspect whether your sign-in or machine credential expired or was
revoked. Use the supported login flow or explicitly authorized replacement; do
not repeatedly submit the same invalid authority.
Signing out ends the interactive session, not separately issued machine
credentials. Revoke machine credentials by their exact retained ID when they are
no longer needed. Revocation does not erase run history or evidence.