Skip to content

Budgets, usage limits, and models

Outcome

You will set project guardrails for AI work, understand what stops at a limit, and change model settings without silently breaking retrieval.

Before you begin

  • You need Admin or Owner to change project settings.
  • Visibility of LLM & budgets can be raised by the owner under Settings → Organization → Access control.
  • Budget changes do not repair already failed runs.

Set the agent budget

  1. Open Settings → LLM & budgets.
  2. Set Agent monthly budget (USD).
  3. Select Save.
  4. Open Settings → Cost and record the current-month usage.

The agent budget is checked before a new autonomous agent run. At or above the limit, new agent runs are refused for the rest of the current UTC month unless the cap is raised. A value of $0 disables new agent runs.

Per-run Max tool calls and Max requests are separate safety ceilings in Operator settings. They bound one runaway task; they do not replace the monthly budget.

Set the search-ranking budget

  1. In LLM & budgets, set Search ranking monthly budget (USD).
  2. Select Save.
  3. Use the displayed meter and Settings → Cost to watch ranking spend.

When paid reranking is unavailable, retrieval falls back to its non-reranked ordering rather than making all project knowledge unavailable.

Current release limitation

The project field is persisted and displayed as a reporting cap, while runtime paid-reranking eligibility is also controlled by the organization's plan. Do not treat this field alone as a contractual hard limit. Confirm the effective plan and the Cost caps meter for the deployment.

Read the Cost page

Open Settings → Cost:

  • Cost caps shows the current UTC month's spend cap and create-issue volume cap;
  • the create-issue cap may come from the organization plan and can be unlimited;
  • Cost summary can show 7, 30, or 90 days;
  • spend is grouped by model, skill, user or non-user actor, and day;
  • System / automated covers scheduled work, webhooks, and background jobs;
  • MCP (API key) identifies MCP usage without a resolved human actor.

At a create-issue volume cap, affected skills pause new issue creation rather than silently exceeding the limit.

Change a model safely

  1. Expand Operator settings only when you own the model migration.
  2. Record the current Default LLM provider, Default LLM model, Default embedding model, and agent model.
  3. Change one setting at a time and select Save.
  4. Run a low-risk question or skill and inspect Activity and Cost.
  5. If the embedding model changed, coordinate and complete Re-index project before judging search quality.

Warning

Changing the embedding model without rebuilding the index can break search. Do not use the manual retrieval-mode override as a shortcut for a failed migration; involve the platform operator.

Recover from a risky budget or model change

Lowering a budget affects new runs; it does not cancel spend already incurred or repair a failed run. Raising it later allows future eligible runs but does not automatically resume work that was refused. Model changes affect subsequent requests, and an embedding-model change without a consistent re-index can make existing retrieval unreliable.

  1. Before the change, record the current settings, current-month spend, and one known-good low-risk query.
  2. Change one value, reload the page, and run the same low-risk check.
  3. Compare Activity, Cost, and the observable answer or skill result.

For a configuration rollback, restore the recorded provider, model, or budget and verify a new run. Retry refused or failed work only after confirming the guardrail and external side effects. For an embedding change, stop further model changes and have the platform operator complete a consistent re-index with the selected model; toggling the field alone is not a rollback. Historical spend cannot be rolled back.

Verify the result

  • Reloading LLM & budgets shows the saved values.
  • Cost shows current-month caps and attributed usage.
  • A safe test run uses the expected model and completes within its guardrails.
  • At a tested zero agent budget, a new autonomous agent run does not start.

Common problems

Symptom Likely cause What to do
A member cannot see LLM & budgets The tab's minimum role is higher Ask the owner to review Organization → Access control
Agent runs stop at the start of a month The monthly cap was reached or set to zero Check Cost caps and the saved agent budget
Search still works after ranking is capped Retrieval fell back without paid reranking This is expected; compare relevance and raise the effective cap only if justified
Cost appears under System / automated A schedule, webhook, or background job initiated it Correlate the time and skill with Activity
Search quality drops after an embedding change The knowledge index was not rebuilt consistently Stop further changes and coordinate a re-index
The displayed ranking cap differs from expected plan behavior Project reporting cap and plan entitlement are not aligned Confirm the effective plan with the operator before changing policy

Next step

Configure write-back policies and retain evidence with the audit and export controls.