Agent

Access and pricing

On this page

Who can provision one
What a turn costs
Choose a brain
Read what it spent
Keep a session or let it go

Running an agent costs two different things: a plan that lets you provision one, and credits for the work it does. They are separate, and a refusal will tell you which one it is.

Who can provision one #

Provisioning requires an active Studio, Team or Enterprise subscription. Starter and Pro accounts can use every model API and the rest of the platform, but POST /agent refuses them. One agent per account.

Method	Path	What it does
GET	`/agent`	Get the current user's NOLGIA Agent, if provisioned.
POST	`/agent`	Provision the current user's NOLGIA Agent (one per account; requires an active subscription).
DELETE	`/agent`	Deprovision the current user's NOLGIA Agent and revoke its access token.
GET	`/agent/usage`	Credits spent and jobs created by the current user's agent.
GET	`/agent/models`	Selectable agent brains (reasoning models) with their flat per-turn credit rates.
PUT	`/agent/model`	Choose which brain the user's agent runs; turns bill that brain's flat per-turn rate.
DELETE	`/agent/model`	Clear the agent's model pin to use the platform default brain and rate.

DELETE /agent deprovisions it and revokes its credential. Sessions and transcripts are yours and are not deleted with it; provisioning again gives you a working agent against the same history.

What a turn costs #

Every turn bills against the same credit balance as a generation, from the pool the credential draws on. Three rules cover it:

A flat minimum per turn, set by the brain the agent runs. Its rate is what GET /agent/models publishes for that brain, and the same catalogue marks which brain you get when you have chosen none.
Metering above the minimum. A turn that does an unusual amount of work bills what it actually cost rather than the minimum. An ordinary turn bills the flat rate.
Generations are billed separately, at their normal per-model price, exactly as if you had submitted them yourself.

Credits are held when the turn starts running and settled when it finishes. A turn that fails or is interrupted has its remaining hold returned; queue time is not billed, because the clock starts when the turn does. A wallet that cannot pay refuses the turn with insufficient_credits rather than running a turn you cannot afford.

Stopping a turn does not cancel or refund generations it already submitted: the work was done and the model was paid. Pricing and credits covers holds, settlement and refunds across the platform.

Choose a brain #

The brain is the reasoning model your agent thinks with. It decides how the agent plans, and it sets the turn's flat rate, so it is the one pricing choice you make.

Read the catalogue rather than hard-coding a rate. Rates are revised, and a brain that costs 10 credits a turn today is not promised to tomorrow.

shell

$ curl -sS https://api.nolgia.ai/v1/agent/models \
  -H "Authorization: Bearer $NOLGIA_TOKEN"
$ curl -sS -X PUT https://api.nolgia.ai/v1/agent/model \
  -H "Authorization: Bearer $NOLGIA_TOKEN" \
  -H 'Content-Type: application/json' \
  -d '{"model":"<model from the catalogue>"}'

Field	Type	Required	Description
`model`	string	Yes	Model identifier, as accepted by PUT /agent/model.
`turn_credits`	integer	Yes	Flat credits billed per agent chat turn while the agent runs this brain.…
`default`	boolean	Yes	True for the platform default brain (what an agent runs with no explicit selection).

PUT /agent/model takes effect on the next turn that starts; there is no wait and no restart, and a turn already running keeps the brain and the rate it began with. DELETE /agent/model returns you to the platform default. Anything outside the catalogue is refused with 400.

Read what it spent #

GET /agent/usage reports the credits the agent's own credential consumed and the jobs it created, newest first.

Field	Type	Required	Description
`credits_spent`	integer	Yes	Total credits consumed by generations the agent's token created.
`jobs_total`	integer	Yes	Number of jobs created with the agent's token.
`recent_jobs`	array of `Job`	Yes

In an organization, the agent bills the organization's shared pool while that organization is your active context, and its spend appears against you in GET /organizations/{id}/usage like any other member's. The agent cannot change that context itself; see Teams and organizations.

Keep a session or let it go #

Sessions accumulate. PATCH /agent/sessions/{id} archives one to keep it out of the list without losing the transcript, and unarchives it again; DELETE /agent/sessions/{id} deletes the session and its transcript. Media a session produced stays in your Library, because assets belong to the Library, not to the conversation that made them.