An agent control plane, not a replacement agent.

TokenSize helps agents coordinate the tools already on a developer's machine. The hosted API recommends and learns from verified outcomes; the local agent remains authoritative.

The boundary#

01
Your machine

The client discovers installed harnesses, models, permissions and normalized allowance. Credentials and raw account output stay local.

02
TokenSize API

The service receives task features, candidate metadata, policy and a request identifier. Prompt text is omitted unless you explicitly opt in.

03
Your chosen harness

The client validates the signed route receipt, enforces the permission ceiling and runs the selected local process with bounded depth.

Privacy model#

The default route is metadata-only. This is a product boundary, not a promise hidden in a marketing footnote.

Sent to the service

  • Task features: role, complexity, risk and context size
  • Candidate model metadata and capability flags
  • Local policy, objective and content-free client version
  • Opaque installation and request identifiers

Kept on the machine

  • Credentials, provider keys and browser callback payloads
  • Repository contents, source files and model output
  • Raw allowance screens and local account identity
  • Prompt text unless you explicitly enable sharing

Why allowance matters#

Quality comes first. Once the router establishes a quality-equivalent tier, normalized allowance helps distribute work across healthy subscriptions so one account does not become the bottleneck.

No silent quality trade-downA larger balance cannot make a weaker model eligible when it falls outside the safe quality set.
Refresh on realityExpired caches and unavailable targets trigger a refresh rather than a blind retry.
Local policy winsThe hosted response cannot raise permissions, bypass a receipt or select the active root host.

Router Brain: improve without collecting work#

Verified outcomes make the next decision more useful, while the data boundary stays the same. Learning is bounded, auditable, and reversible.

Bucket by task shapeRoutes use coarse non-content features — role, complexity, context size and capabilities — never a prompt fingerprint or transcript.
Count verified outcomesReceipt-bound agent feedback and provider-reported outcomes are idempotent success/failure observations.
Require confidence before influenceSignals move from baseline to observing to calibrated; only a conservative lower bound can affect a route.
Quarantine and roll backA failing candidate is removed from learning influence; static quality, allowance, readiness and local policy still decide.

The public demo uses static catalog priors. Router Brain signals are available to authenticated agent and model routes after migration0005_router_learning.sql is applied.

Recursion is blocked by contract.

Embedded Codex and OpenCode clients mark their root harness active. The router excludes that harness even at depth zero, and every child receives an incremented depth.