Spend (FinOps)

See spend in Admin and stop spend at admission — export and alerts included; not a full BI warehouse.

Configure USD (pricebook)

Tokens show up automatically from the model response. Dollars do not — you must map each model id string the runner reports in finops.pricebook (or use a gateway that inlines usage.cost / cost_usd).

{`{
  "finops": {
    "pricebook": {
      "gemini-2.0-flash": { "input_per_1k": 0.0001, "output_per_1k": 0.0004 }
    },
    "budgets": {
      "tenants": { "default": { "max_usd_per_day": 25, "soft": true } }
    }
  }
}`}

If Admin Spend shows tokens but $0 and an usage_unpriced alert, the model id in the alert attrs is missing from the pricebook — add that exact string, restart the control plane, and new runs will price. Full field list: docs/configuration.md.

What ships now

• Optional finops in langgraph.json: UTC-day budgets, alerts.soft_pct, reservation holds (optional hold_ttl (keep longer than longest run), per-agent agents overrides), and routing.aliases cheaper-agent rewrite near soft threshold. Optional on_hard_breach: cancel_inflight cancels pending/running runs when a hard day cap is already breached.
• Terminal Output.usage → SQL usage_events (LangGraph Python/TS; CrewAI/LlamaIndex/LangChain/AutoGen adapters when metrics exist). A HITL pause or a mid-run cancel still meters whatever tokens ran before the pause/cancel — it does not wait for a terminal success to record them. Webhook event budget_alert + Admin alerts strip (approach alerts deduped once per UTC day per process).
• Admin /admin/spend labels dollars Estimated USD, not "Cost": tokens are exact (from the model response), dollars are tokens × your finops.pricebook rate — unless the runner is talking to an LLM gateway that reports cost inline with usage (OpenRouter's usage.cost, and OpenAI-compatible gateways following the same convention), in which case that reported figure is used instead and the pricebook is skipped for that run. With a non-empty pricebook, a terminal run with real tokens, no reported cost, and a model id missing from the book raises a usage_unpriced alert instead of silently looking free (empty pricebook = tokens-only, no alert; a present $0-rate row is intentional free tier). If a run's reply carries no token data in any shape the runner recognizes at all (a brand-new provider, or an integration that has not adopted LangChain's standardized usage_metadata fields), that raises a usage_unmetered alert instead — the one failure mode worse than under-pricing a run is making it invisible. APIs: /admin-api/usage/{summary,alerts,holds,export}.
• Live editing from Admin, no redeploy. The finops block in langgraph.json is only the bootstrap baseline — open Admin → Spend and use the Live FinOps config panel to add pricebook rows or change a tenant/agent day budget directly, no file edit or control-plane restart required. The change is live on that replica the moment you save, and every other replica picks it up within about 15 seconds (same fingerprint-poll mechanism as durable policy grants). The panel shows file baseline and effective (live) side by side so an overridden value is never a surprise, and Clear overlay reverts everything back to whatever the file says. Writes are validated server-side (no negative rates or budgets, a same-magnitude-swapped input/output rate is flagged, and a genuinely-intentional $0 free-tier rate is confirmed on save rather than silently rejected). API: GET|PUT|DELETE /admin-api/finops.

What is not shipped yet

• Streaming mid-run token metering (kill of inflight is opt-in via on_hard_breach: cancel_inflight, not continuous spend tracking) — usage is still only recorded once a run reaches a terminal or interrupted status.
• Reconciling against a provider's account-level billing API (OpenAI/Anthropic Costs/Usage endpoints) after the fact; multi-currency / warehouse BI.
• Reading cost from gateways that report it out-of-band instead of inline on the response (Portkey response headers, Helicone's async dashboard) — routing calls through one works today (it's a plain client base_url change, nothing Runkite-specific), but its cost won't reach Runkite unless it inlines cost the way OpenRouter does; those runs price from tokens × pricebook like any direct-provider run.
• Automatic model-catalog shopping (routing only uses configured routing.aliases).
• Frameworks with no token metrics still report empty Output.usage (best-effort only).

Practical stance

Cost caps and metering sit next to kill, grants, and HITL. Wire finops in config, watch /admin/spend, and hard-deny when a day budget is gone.

Related: Security · Kill & break-glass · Limitations · docs/configuration.md