Skip to content

Pay for local runs with Agentry credits

The credits on your SecureAgentry account can pay for the model behind runs on your own computer. You need no API key (a long secret code that lets a program use your account) and no account at a model provider. Agentry chooses the model, and each run spends credits from your balance as the model works.

Every tier can do this, Standard included, from the same balance that pays for submissions and runs In Agentry.

This is for runs on your own computer: the desktop app, the agentry command and persistent agents. A run In Agentry (hosted) already pays with credits on Agentry's own model access (Running on Agentry).

What it costs

  • The same rate as runs In Agentry. Credits are counted from the model's own figures for what it read and wrote, whatever the model (or from an estimate, when it gives none: below).
  • Charged as it goes. About every 15 minutes while a run lasts, and at its end, the credits used so far come off your balance. A run is rounded up to a whole credit once, at its end. Nothing is held at the start.
  • Shown where you look. The run's receipt, the desktop's run card and your dashboard's credit history name what it cost. On the dashboard the line reads Local runs on Agentry's model.
  • Who pays. You do, or, if you belong to an organization, its pool, as for a run In Agentry (Subscription tiers and credits).
  • Publishers earn nothing from a run paid this way, even of another account's public QuickAgent (Earnings).

The model

Agentry chooses the model, and may change it. The desktop app names the model in use when you choose Agentry credits, and each run's receipt names the model it ran on. You cannot pick another one: to choose your own model, use your own provider and key instead.

Runs paid this way use Agentry's default harness, the program in the run's container that talks to the model. A run set to the Claude Code, Codex or GitHub Copilot harness refuses to start with Agentry credits; those harnesses are for a plan you already have.

Set it up in the desktop app

  1. Sign in to your SecureAgentry account in the app.
  2. Open Settings → Model provider. The first-run setup has the same choice.
  3. Choose the Agentry credits tile. Until you sign in, it is greyed out with Please sign in across it.
  4. The app shows your balance and the model Agentry uses now. It asks for no key and no model.
  5. Click Save provider settings, or Prepare images now to save and get the run images ready in one go.

From then on, every run on this computer pays with your credits: QuickAgent runs and chats, the Builder's Drafter, reviews, the Knowledge Creator, the Voice Writer and persistent agents.

Set it up on the command line

agentry login
agentry setup --provider agentry

agentry setup records the registry you are signed in to, and asks for no key. With --provider agentry it refuses a key or a base URL, and a --model you name is not used. Signed out, it stops and asks you to run agentry login first. agentry doctor says whether the sign-in is in place. Then run QuickAgents as usual with agentry run.

An AI agent of your own can make the same choice through the MCP server's agentry_setup tool, with provider set to agentry. Signing in stays in a terminal (From your own agent).

Persistent agents

A persistent agent opens one credits session, which pays for its model, when it starts. It keeps the session for as long as it runs, renewed about every 15 minutes. If your computer sleeps for more than an hour, the session ends. When the computer wakes, the agent opens a new session and restarts the work it was doing on it, and says so. The session ends when the agent stops.

When the credits run out

When your balance is used up, the model's next call is refused before it reaches the model.

  • A run ends with out of credits, and says that your Agentry credits are used up. Its receipt says why.
  • A persistent agent fails that turn, says so in its chat, and stays up.

You are never charged past your balance. Your monthly credits come back at the reset. On Pro you can buy a credit pack to carry on sooner. Or switch to your own provider in Settings → Model provider (agentry setup on the command line).

How your sign-in stays out of the run

A QuickAgent runs in a container on your computer. Paying with credits does not hand that container your SecureAgentry sign-in:

  • A token for this run only. When a run starts, the desktop app or the agentry command uses your sign-in to open a credits session, and hands the container that session's token. The token is good only for the model's calls. It cannot read your account, publish, install or start another run.
  • Kept alive from outside. The app or the command renews the session about every 15 minutes while the run lasts. Each renewal checks your sign-in again. The token never changes, so nothing in the run restarts.
  • Ended with the run. When the run ends, so does its session, and the token stops working. If a run stops without ending its session (the computer turns off, say), the token lapses within an hour.
  • Cut off with your sign-in. If you revoke the token your computer signed in with, under Settings → Authentication, the session cannot be renewed, and its token stops working when the session runs out, up to an hour later. Revoking does not stop it at once. agentry logout only removes the sign-in from this computer and does not revoke it (Personal access tokens).
  • One address. The run's model client is pointed at the model address of the registry you are signed in to, and the token is good there alone.

There is no spending cap on a session other than your balance. Anyone who copied a session's token could spend your credits until it lapses. The token is handed to the run's container alone, and is masked in the run's receipt and logs.

What happens to what you send

Your prompts and the model's answers pass through Agentry on their way to the model's provider and back. In each request Agentry changes only the key and the model's name and, when the answer is streamed, asks the provider to report the tokens it used. It passes the answer back as it comes, a streamed answer piece by piece. It keeps neither the prompts nor the answers. For each call it records the session, the result, how long it took and the model's token counts, which are what your credits are counted from. If the provider reports no counts, or a stream is cut short, the tokens are estimated from the size of the request and the answer. The model's provider handles the request under its own terms.

Limits

  • A request can be up to 8 MiB.
  • A run makes at most 4 model calls at a time.
  • One payer can have 8 credits sessions open at once, each a run or a persistent agent. The members of an organization share its 8.