OpenRouter management keys gain read-only and workspace scoping
- Source
- x.com
- Date
Management keys on OpenRouter now have tighter access controls. Make a key read-only, limit it to one workspace, or both, so every agent, script, and tool gets only the access it needs, and if a key leaks, the impact is controlled. It’s available today on all plans.

Context
OpenRouter, a service that gives developers one API for models from many providers, now lets account owners limit what a management key can do. A management key is an admin credential rather than one for calling models. OpenRouter's documentation says it cannot be used on the completion endpoints and exists to create and rotate API keys and to manage members and workspaces. Workspaces are separate environments inside one account, each with its own API keys, routing defaults, and spending limits, according to the workspaces guide, which still describes management keys as operating at the account level across all workspaces. That is now only one option.
The announcement adds two limits that can be combined. A workspace-scoped key acts only on its own workspace, and requests outside that, including listing, creating or deleting workspaces, setting workspace budgets and checking the account's credits, are refused with a 403 error. A read-only key can make read requests and nothing else: any request that would create, change or delete something returns 403 Forbidden. OpenRouter suggests read-only keys for monitoring, reporting and auditing, and its screenshot shows that use: a key named Usage report, scoped to a Production workspace, read-only, and set to expire after 90 days.
This applies least privilege, giving each program only the access its job needs, to admin credentials. A cost dashboard or an that only reads usage no longer has to hold a key that could mint new API keys for every workspace or delete existing ones, which is the point of the post's claim that a leak's impact is controlled. A leaked read-only key still reveals what it can read, so the change limits the damage rather than removing it. One documented exception matters for automation: reading a guardrail policy assignment needs a read/write key, because reading an assignment creates it if it does not exist yet. The controls are available on all plans.
- guardrails — The checks around a model that block bad inputs and outputs — filters, validators, and permission rules the model itself can't override.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
Checking sign-in…
Loading comments…





