AI features
Bring your own Anthropic key
Why ProductOS is BYO-key, how to plug yours in, and what it costs in practice.
ProductOS uses Anthropic's Claude models for every AI feature: chat copilot, ProductSense, spec generation, release notes, the works. We don't resell access. You pay Anthropic directly with your own API key. The one non-Anthropic piece is embeddings for semantic search, which run on OpenAI's text-embedding-3-small through ProductOS's own key.
This page covers why, how to plug yours in, what onboarding does without one, and rough cost expectations.
Why BYO
Three reasons:
- Honest pricing. We charge a flat subscription. If we resold Claude, we'd either over-charge heavy users or eat their costs on top of paying for the platform. Neither is sustainable.
- Your data, your contract. With BYO, your workspace's prompts and completions go through your Anthropic agreement, your data-retention settings, your privacy posture, your audit. Calls are made from ProductOS's servers, not your browser: the server assembles each prompt from your request and the workspace context it needs, then sends it to Anthropic with your key.
- No throttling. Heavy AI usage doesn't degrade for other workspaces. If you've negotiated a higher rate limit with Anthropic, it applies; we don't cap.
Adding your key
- Get an Anthropic API key at console.anthropic.com.
- In ProductOS, open Settings → AI (admin or owner).
- Paste the key and click Save. It's encrypted and stored; there's no validation call, so a bad key surfaces on the first AI request.
The key is encrypted at rest (AES-256-GCM). It's saved against the signed-in user, not the workspace: the wizard's API-key step saves it for the owner, and each admin who uses AI features saves their own under Settings → AI. The field loads the saved key back masked, with a show/hide toggle, so you can check or rotate it in place; pasting a new key overwrites the old one. Once it's saved, every AI call you make runs on your key, including where ProductOS also provides a hosted key.
Keys are per user, and Settings → AI is available to admins and owners. Members can't add their own key yet. Their AI calls run on the hosted key where ProductOS provides one, and otherwise fail with No API key configured; the workspace's admins can run AI features on their behalf in the meantime.
What runs without a key
Onboarding, and a daily allowance of coaching. The new-signup wizard at /onboarding uses a ProductOS-funded key so first-time users can experience the AI before plugging in their own. The PM Coach (the Copilot in Learning Mode) is subsidized the same way, capped at a fixed number of questions per user per day, after which it falls back to your key. Once onboarding completes (the wizard finishes its commit step), every other AI feature uses your key.
The hosted key, where provided. ProductOS can also provide a hosted key that covers AI calls from anyone who hasn't saved their own. It never takes precedence: the moment you save a key, your calls run on it.
Without a saved key after onboarding, and with no hosted key provided:
- The copilot chat answers with "No API key configured. Add your Anthropic API key in Settings → AI."
- ProductSense, spec generation, release notes and the other generate features return the same error.
- Non-AI features (Inbox, Themes, Features, Initiatives, Releases, Outcomes, API, integrations) work fully.
Cost expectations
Illustrative numbers for a ~40-user workspace:
| Feature | Calls / month | Avg cost / call | Monthly total |
|---|---|---|---|
| Copilot chat | ~1,200 | $0.02 | ~$24 |
| Auto-tag (per feedback) | ~3,000 | $0.001 | ~$3 |
| Theme clustering (weekly) | ~4 | $0.50 | ~$2 |
| ProductSense reports | ~6 | $1.20 | ~$7 |
| Release notes drafts | ~12 | $0.04 | ~$0.50 |
| Total | ~$37 / month |
Embeddings for the feedback index run on OpenAI through ProductOS's key, so they never show up on your Anthropic bill. Heavier usage (high-volume feedback ingest, larger workspaces, deep Claude Code sessions) scales linearly. Lighter workspaces are well under $10/month.
The Usage panel on Settings → AI shows spend for the last 7, 30 or 90 days, broken down by source (spec generation, Claude Code sessions, onboarding, ProductSense, embeddings, other) with a daily strip, so you can spot patterns before the bill arrives.
Which model is used
ProductOS picks the model per feature based on what's worth paying for:
| Feature | Model class | Why |
|---|---|---|
| Copilot chat | Claude Sonnet 4.6 | Multi-turn reasoning and tool calls over the workspace context block |
| Auto-tag, title suggestions, urgency scoring | Claude Haiku 4.5 | High throughput, low stakes, cheapest tier |
| ProductSense, spec generation, release notes, onboarding | Claude Sonnet 4.6 | Structured output over a full page of entities |
| PM Coach | Claude Haiku 4.5 | Short answers on a subsidized key |
| Embeddings | OpenAI text-embedding-3-small | Semantic search and clustering; billed to ProductOS, not your key |
ProductOS calls Anthropic's model aliases (claude-sonnet-4-6, claude-haiku-4-5), so point releases roll in automatically and major upgrades roll out centrally. There is no per-workspace model picker.
What the Settings → AI page lets you change
- Anthropic API key: paste, rotate.
- Workspace context: a button that takes you to Knowledge, where vision, competitors, personas, and OKRs feed every AI prompt.
- AI features list: a read-only catalogue of what's unlocked once a key is configured (ProductSense, Copilot Chat, AI Insights, Release Notes).
- AI actions: a workspace-wide switch, Allow AI to apply changes. Off, the Copilot and the MCP server can still answer and propose, but Apply and the MCP write tools are blocked for everyone regardless of role.
- Usage panel: AI spend by source over the last 7, 30 or 90 days.
There is no switch that turns AI reads off entirely. Where ProductOS provides no hosted key, not saving a key switches AI off: every AI surface returns No API key configured. Add your Anthropic API key in Settings → AI. (HTTP 400) and nothing is sent to Anthropic. Where a hosted key is provided, calls from people without a saved key run on it, so the AI actions toggle is the control: reads stay on, writes stop.
Next
- Connecting Claude Code: how to use the workspace from inside your editor.
- ProductSense bulk cleanup: the heaviest AI feature; budget for it.