ComputeLlamaJSON API

Stage 3 / private decision plane

Recommendations an agent can explain before anyone routes.

ComputeLlama ranks eligible routes against your real commercial state, then produces immutable gateway configurations. It does not receive provider credentials, apply configuration, or claim counterfactual savings as realized.

Local CLI + MCPAuthenticated hosted API + MCPNo active routing

Developers

One CLI, local or hosted.

pnpm optimizer recommend \
+  --decision-key coding-2026-08-03 \
+  --workload coding-agent \
+  --policy research-default

Agents

Nine typed MCP tools.

Recommend, explain, simulate, generate, inspect, diff, approve, roll back, and poll route-change alerts through the same service contract.

Control boundary

Approval is data, not a mutable flag.

Every decision, configuration, approval, rejection, rollback, and alert is tenant-isolated, append-only, canonicalized, and hashed.

The gated path

Inspect every transition.

Stage 4 applies later
  1. 01

    Recommend

    Join exact public evidence, independent weather, private terms, balances, budgets, and policy into one immutable decision.

  2. 02

    Simulate

    Replay the exact snapshot or test explicit counterfactuals without rereading changing data or sending inference traffic.

  3. 03

    Generate

    Emit credential-free LiteLLM, OpenRouter, or Portkey documents with source digests and a bounded fallback sequence.

  4. 04

    Approve

    Bind approval or rejection to one exact manifest digest. Rollback creates a new unapproved revision instead of rewriting history.