Cost & budgets

The Cost tab is your spend ledger. It aggregates every token your workers burned — broken out by worker, provider, and model — for the period you pick, so you always know what the AI is costing you. From here you also export a CSV for your accountant, see how much the shared cache has saved, and set budgets that can warn you or hard-stop a worker before it overspends.

Token counts come straight from each provider’s reported usage, so what you see here should match what shows up on their bill. Local Ollama runs are free.

Period & total

• Period toggle — switch between Daily, Weekly, and Monthly views. Default: Monthly.
• Total — the big number is your total spend for the selected period.
• Refresh — re-pull the latest numbers on demand.

The spend table

Below the total, a table lists each combination of:

• Persona — which worker spent it.
• Provider — which AI vendor (OpenAI, Anthropic, …).
• Model — the exact model used.
• Tokens — total tokens consumed.
• Cost — the dollar cost for that row.
• Runs — how many runs contributed.

If nothing’s recorded, it means agents haven’t run, or your providers don’t return usage (local Ollama, for instance, is free and reports none).

Export CSV (for tax time)

Click 💾 Export CSV, pick a month and year, and Download .csv. The file lists every cost event with worker names resolved — hand it to your accountant or keep it for tax filings. The filename includes the period (e.g. asktheodor-cost-2026-06.csv).

Shared tool-call cache (🔁)

When two workers in the same Company call the same URL or run the same web search within the cache window, the second call resolves from this shared cache instead of making a fresh network round-trip — saving time and tokens. Cached results never cross Company boundaries.

This section shows:

• Cache entries, Total hits, Bytes cached, and ≈ Tokens saved at a glance.
• A per-tool table with entries, hits, and estimated tokens saved, plus a Clear button per tool.
• Refresh and Clear all controls for the whole cache.

Budgets

Budgets cap spend and (optionally) stop a worker before it goes over. Create one with:

• Scope — per worker (cap one worker) or 🌐 global (total), a single cap on all spend — the master safety net for unattended and Fleet runs.
• Worker — which worker the cap applies to (only for the per-worker scope).
• $ cap — the spending limit for the period.
• Period — daily, weekly, or monthly.
• Hard-stop — when on, the agent loop halts at its next step once the period total crosses the cap, with an inline warning. When off, the budget is warn-only (it alerts at ~80% but doesn’t stop the work).

Existing budgets show in a table where you can toggle Hard-stop or delete a budget. Use + Add to create one.

Set a global hard-stop before running unattended

Before you let Plans, Routines, or a Fleet machine run on their own, add one 🌐 global budget with hard-stop on. It’s the single backstop that guarantees nothing can quietly run up a large bill while you’re away.

Where pricing comes from

Per-token prices come from a built-in pricing table. Token counts are taken verbatim from each provider’s usage field, so the ledger reflects real billed usage — update the table if a provider changes its prices.

In practice

A budget layout that works for most people: one 🌐 global monthly budget with hard-stop on as the backstop, plus a per-worker budget on anything running unattended. The global cap protects you from the scenario you didn’t predict; per-worker caps tell you which worker misbehaved.

Pick the global number by what you’d tolerate losing, not by what you expect to spend. The cap exists for the run that goes wrong at 3am, so set it where an unnoticed overrun would annoy you rather than hurt. You can raise it after a month of real data.

Warn-only has one legitimate use: learning what a new setup actually costs without interrupting work. Once you know, switch it to hard-stop. Leaving everything warn-only means the alert arrives while the spending continues.

Reading the table diagnostically: one worker dominating the spend usually means it’s on a frontier model for work that doesn’t need one, or it’s caught in a retry loop. High runs with low output is the classic runaway signature — check its allowlist and per-run caps.

Cheap wins on spend, roughly in order: move routine, high-volume workers to local models (free, and they report no token cost); keep workers that repeat the same lookups inside one Company so the shared tool-call cache can serve them; and trim allowlists so workers stop invoking tools they don’t need.

Export monthly, not at year end

The CSV resolves worker names at export time. Pulling it each month gives your accountant a clean series and avoids reconstructing a year of activity in one sitting.

🎓 Learn it hands-on: Approvals: humans approve, agents prepare

Terms in this page

• Cost ledger — the running record of what your workers have spent, totaled per period.
• Token — the unit AI models read and write in (roughly ¾ of a word). Providers bill per token.
• Persona / worker — a configured AI character; the table attributes spend to whichever one ran.
• Provider — the AI vendor behind a model (OpenAI, Anthropic, Gemini, and so on).
• Model — the specific AI model used for a run; different models cost different amounts per token.
• Run — one execution of a worker (a chat turn, a task step) that consumed tokens.
• CSV — a plain spreadsheet file you can open in Excel/Numbers or hand to an accountant; here it lists every cost event for a chosen month.
• Tool-call cache (🔁) — a shared store of recent tool results (fetched URLs, searches) that workers in the same Company reuse to avoid repeating work and re-spending tokens.
• Cache hit — a request served from the cache instead of making a fresh network call.
• Company boundary — the rule that cached results are shared only among workers in the same Company, never across Companies.
• Budget — a spending cap for a worker or for the whole workspace, over a period.
• Scope (per-worker / global) — whether a budget limits one worker or all spend combined.
• Hard-stop — a budget that actually halts the agent loop when the cap is reached, versus a warn-only budget that just alerts.
• Ollama — a local model runner; its runs are free and report no token cost.