Cost controls
Keep model spend bounded: only models you connected and selected can be called, one throughput dial sets how many strategy-creation runs a day agents may start, chat has a round cap and a per-conversation cap, and spend is visible per agent, per run and per day.
Forven's AI work runs on your own model key. Every agent run and every assistant message spends against your provider account, not Forven's. This page covers the controls that keep that spend deliberate: which models may be called at all, how much work the agents may start, how far a chat may go, and where to watch what you spent.
The rest of the lab — data, backtests, the gauntlet, paper trading — needs no key and costs nothing to run.
Forven never resells model tokens and never sees your key. Spend is between you and your provider. Numbers on this page are the current build's defaults, not a quote.
Spend safety: only models you chose
The strongest control is structural. Forven only calls a (provider, model) pair that is connected (Agents → Setup → Providers) and selected — on your Shortlist, or picked for an agent, a fallback, a background task or the backup model (Agents → Setup → Agent models). A key sitting in an environment variable isn't enough: until you connect it in the app, its provider can't be called.
The check sits beneath every caller — agents, the Brain, the assistant, deep-dive and background tasks — so nothing bypasses it. Fallbacks never cross to a provider you didn't add yourself, so a throttled model can't quietly re-route your prompt, and your bill, to another vendor. See models & providers.
Throughput presets
How much autonomous work the agents may start is set by one dial: Throughput preset, in Settings → System → Throughput (an advanced section).
| Preset | Strategy-creation runs a day | Cadence and workers | Use it when |
|---|---|---|---|
| Trickle | 4 | Hours-scale; one worker each | Hard-capped free models |
| Conserve | 12 | Slowed; small batches | Free or rate-limited routes — the fix if you're seeing 429s |
| Balanced (default) | 40 | The shipped defaults | A normal paid key |
| Max | 120 | Minutes-scale; all workers up | A high-limit paid or local model |
The daily strategy-creation budget is the biggest driver of model calls: it's how many times a day the Strategy Developer may be sent to write an idea and build strategies from it. A preset also writes the loop intervals, how many agent runs may be claimed at once, the backtest subprocess budget and the gauntlet workers. Edit any of those by hand and the dial shows Custom.
The budget only applies in Auto mode. Semi and Manual create nothing on their own, and ideas you submit yourself (The Forge → + Submit idea) don't count against it. See autonomy modes.
If Forven notices you routing to free OpenRouter models, a banner suggests switching to Conserve.
Chat: tool rounds and the per-conversation cap
The in-app assistant, including deep-dive on a strategy, lets a model loop through tool calls within one reply. Each round re-sends the whole turn, so cost grows faster than the number of rounds. Two limits bound it:
- Chat tool rounds — Settings → System → Assistant (default 12, allowed 2–40). The cap lands softly: near the limit the model is told to wrap up, and at it the assistant gives a final summary of what it did instead of an error.
- Cost per conversation — each assistant conversation has a spend cap, $5 by default (
assistant.cost_cap_usd; read or set it withGET/PUT /api/assistant/cost-cap). Once a conversation reaches it, further messages are refused; start a fresh conversation or raise the cap.
The daily agent cap (advanced)
The engine also has an opt-in daily guard for background agent work: agent_daily_cost_cap_usd, default 0 (off). Set above zero, Forven totals today's agent-run costs before each new run starts and refuses to start more once the total reaches the cap; refused runs stay queued until the next UTC day or until you raise it. Runs already going finish and book their cost.
The Settings page doesn't expose this key. Treat model choice and the throughput preset as your everyday levers, and your provider's own budget alerts or hard limits as the backstop.
Where cost comes from
Forven doesn't read your provider's invoice. It estimates cost locally:
- A call returns its token usage.
- A per-provider, per-model price table turns that into dollars.
- The figure is stored on the run that made the call.
Estimates err on the safe side. Known-free routes (a local LM Studio server, OpenRouter :free models) cost nothing. A known provider with an unrecognized model is charged that provider's highest published rate, and an unknown provider a conservative default, so unknown usage never reads as free. If a provider changes its prices, the number can drift from your invoice until the catalog is refreshed. Your provider's dashboard is the source of truth.
Watching spend
- Agents → Overview — today's spend (or the last 7 days) and the 30-day total, overall and per agent.
- Agents → Runs and each run's Run Detail — the cost, tokens and model of every run, next to its transcript.
- Diagnostics — Cost (24h): dollars, runs and tokens over the last day.
Steps: set a sensible budget posture
Connect only what you'll pay for
On Agents → Setup → Providers, connect just the providers you're willing to spend on.
Put cheap models where the volume is
On Agent models, give busy roles and the background tasks cheaper models, keep your best model where reasoning matters, and leave fallbacks empty unless you want cross-provider failover. Use the Shortlist to keep the models you want at the top of every picker.
Pick a throughput preset
In Settings → System → Throughput, choose a preset that matches your provider's limits. Unsure? Start with Conserve.
Watch for a day
Check Cost (24h) on Diagnostics and the spend on Agents → Overview, then adjust.
You should see a steady daily cost, per-agent spend that matches the roles you put on expensive models, and no surprise vendors in the run records.
Good to know
- Cost figures are estimates from token counts and list prices, not a billing ledger.
- Presets change how much work runs, not model prices. A cheap preset on an expensive model can still cost more than a fast preset on a cheap one.
- Cost controls bound model spend; they say nothing about trading risk. For that, see risk controls.
Forven is research software. Nothing here is financial advice, and no number shown — cost or backtest — predicts future outcomes.
Related
- Models & providers — the connected-and-selected rule and fallbacks
- Agents — the runs that spend your budget
- The assistant & deep-dive — what the chat caps bound
Models & providers (BYOK)
Bring-your-own-key model setup in Forven: sixteen providers, the connected-and-selected spend rule, fail-closed fallbacks, the backup model, and background tasks — all on Agents → Setup.
MCP server
Connect Claude Code, Claude Desktop, Cursor, Codex or any MCP client to Forven so an AI can design, register, backtest, validate and promote strategies through the real gates — and give Forven's own agents external MCP tool servers.