LLMKit / request controlOpen source / MIT

Cost control for agents that actually run.

Give every request an identity, a budget decision, and a receipt. Start locally, then move enforcement to the gateway when the system needs it.

Start locallyRead the source
Hosted accounts are temporarily unavailable. Local tools remain available.
request signalactive
  1. 01

    identify

  2. 02

    reserve

  3. 03

    dispatch

  4. 04

    settle

decisionALLOWED / BUDGET RESERVED
Run the first signal locally

Add local cost tools to your coding agent

$npx @f3d1/llmkit-mcp-server

Runs locally. No LLMKit account or provider proxy required.

Full setup ->

The request lifecycle

A small boundary with hard edges.

In gateway mode, pricing describes the request. Enforcement decides whether it is allowed to become spend.

  1. 01

    Identify

    Bind provider, model, key, session, and price to one request.

  2. 02

    Reserve

    Hold the estimated cost against the budget before dispatch.

  3. 03

    Decide

    Allow or reject at the boundary, before provider spend occurs.

  4. 04

    Settle

    Reconcile actual usage and preserve the decision trail.

Four ways in

Meet the code where it already runs.

Integration docs ->
MCPInspect coding-agent sessionsRead supported Claude Code sessions and Cline task data without an LLMKit account.CLIWrap an existing processAdd a cost summary to an agent command without changing its application code.SDKInstrument the request in processUse typed Python and TypeScript clients when attribution belongs in your code.APIMove enforcement to the gatewayApply the same identity and budget policy across provider-compatible traffic.