# Facet

Facet builds a custom AI model for a task your product repeats, such as reading every email or sorting every support ticket.

You describe the task and say what good looks like. Facet builds a small model for it, scores every version on hidden examples, serves it behind an OpenAI-compatible API, and keeps it right as you flag misses. Every change shows what got better and what got worse before it goes live.

## Four ways in

- **Your agent:** connect Claude Code, Cursor, ChatGPT, Codex or VS Code to `https://facet.halfspacelabs.com/mcp` and ask it to make a Model or make one better. See [Use Facet from an agent](/docs/agents), or start with the [Quickstart](/docs/quickstart).
- **Slack:** a workspace Owner adds Facet to Slack from **Settings**. Then mention `@Facet` in a channel, or send it a message, with the task and a few examples. It replies with the draft, its evals and a quote, and you approve there. See [Use Facet from Slack](/docs/agents#slack).
- **The web app:** sign in at [facet.halfspacelabs.com](https://facet.halfspacelabs.com) to watch, review, approve and manage your Models.
- **Your code:** call a Model with any OpenAI SDK, by its name. See [Call a Model](/docs/call).

> Facet is invite-only for now. [Join the waitlist](https://halfspacelabs.com/waitlist) and we'll email you an invite.

## When Facet fits

Facet fits one narrow task that runs often, where a frontier model is slow, costly or not quite right:

- Sorting and routing: support tickets, leads, documents
- Pulling fields out of text: invoices, contracts, emails
- Rewriting in a house style, or redacting personal data
- Grading or checking text against your rules

Keep a frontier model for open-ended work that changes every time.

## How it works

1. **Describe the task.** A sentence or the prompt you use today, and a few sample inputs if you have them.
2. **Review the draft.** Facet writes the Model's Instructions and its evals: plain statements of what a good answer does, each with examples.
3. **Create.** Facet builds v1, checks it against every eval, and makes it live.
4. **Call it.** Use the Model's name as the model in any OpenAI SDK.
5. **Ask for better.** When an answer is wrong, say what should change. Facet quotes the fix before it spends anything.
6. **Keep it current.** Turn on Retraining and Facet checks the live version on new rows, and drafts a quoted retrain when a score drops.

## Next steps

- [Quickstart](/docs/quickstart): your first Model, start to finish
- [Concepts](/docs/concepts): the four parts every Model is made of
- [Pricing](/docs/pricing): what a Model costs to build and to call
- [Questions about Facet](https://halfspacelabs.com/faq): short answers on quality, pricing and your data