LLM cost
agentbreeder cost reports LLM spend per deployment and per model from the gateway's LiteLLM spend logs. An estimate from list prices, not a bill.
Preview
Run on macOS with Docker Desktop against Northwind on Gemini (2026-11-15): a release, one Try it
turn and 5 invoke turns made 260 model calls, and cost, by deployment and by model, equalled
a SQL sum of LiteLLM's own spend log (USD 0.264939). Nothing was compared with a provider
invoice.
agentbreeder cost reports LLM spend per deployment and per model from the gateway's LiteLLM spend logs. It is an estimate from list prices, not a bill.
agentbreeder cost --stack . --environment dev # per deployment, last 30 days
agentbreeder cost --stack . --by model --since 2026-11-01 --jsonEach deployment calls the gateway with its own key, so LiteLLM's spend log can be added up per
deployment and per model. cost only reads: it asks the gateway with the master key from the
operator config and never prints it.
- Amounts are LiteLLM's estimate from its model price list, in USD. Your provider's invoice is the record.
- Spend logs arrive about a minute late, so the newest calls can be missing.
- Removed deployments still count. Their keys are revoked; their log rows stay.
- Only AgentBreeder's keys are counted, including eval judge keys (
eval run <run>). - Ollama calls have no price and show
no price (local model), never $0.00. CACHED INshows prompt tokens the provider read from its cache (runtime/adk@1.4.0).
Cost per turn
runtime/adk@1.2.0 and later return each turn's cost as cost_usd, the sum of the gateway's
per-call cost header. In the Northwind run the turns matched the spend log's figure for that key's
chat calls within rounding; the turn figure leaves out the knowledge embedding calls. Older runtimes show
"Not measured".
In the console
Home has an LLM cost card per project, and an agent's Overview has one for its own
deployments. Both showed the same figures as cost and the spend log in the Northwind run. Try it
shows each turn's cost_usd.
Budgets
A key's budget is enforced by LiteLLM, on the same estimate. It checks before each call against the spend so far, so the turn that crosses the budget still completes. Ollama calls are recorded at 0.0, so a budget would not count them (not run). See Models and gateway.
Canonical reference: cost.md.
Console (agentbreeder up)
agentbreeder up starts a loopback, single-user web console (formerly the command center) that runs the same CLI commands to build, release and inspect agents.
Agent traces
agentbreeder traces lists each deployment's recent agent traces from the stack's Phoenix: models, tokens, tool calls and latency, never message content.