AI assistants

An assistant can run the whole flow for you: send the data, answer the review, start the training, follow it to the result and wire the model into your code. Every training says what it needs next (next_action, see Trainings), so the assistant always knows the next step.

Remote MCP serverCopy Anchor Link

https://api.typestrict.ai/mcp

Nothing to install: add the URL to your assistant and sign in to typestrict when it asks (OAuth). Assistants that take custom headers can use an API key instead (Authorization: Bearer ts_..., from Settings). It has a tool for every step; starting a training always takes a quote and your confirmation, so no credit is spent without asking.

The remote server cannot read your disk: the assistant sends text files (CSV, JSON, text) by their contents. For large files, Excel, PDF or zip, upload them in the dashboard and let the assistant take the training from there, or use the local server.

Local MCP serverCopy Anchor Link

The typestrict CLI includes the same server over stdio, typestrict mcp. It runs on your machine with your API key, so the assistant can send any local file and save the model into your project.

curl -fsSL https://typestrict.ai/install.sh | sh
typestrict login                 # an API key from https://typestrict.ai/settings#api-keys

SkillsCopy Anchor Link

typestrict-lab/skills puts your assistant to work on what makes a model good: it checks and prepares the data, reviews the agent's proposal and fixes it before training, and integrates the result. They use the MCP server (remote or local) and the CLI, and carry the tips for the best model.

git clone https://github.com/typestrict-lab/skills.git
mkdir -p .agents/skills && cp -r skills/typestrict-* .agents/skills/   # Codex, Cursor, Copilot, Antigravity, opencode
mkdir -p .claude/skills && cp -r skills/typestrict-* .claude/skills/   # Claude Code

Each assistant's page says where it reads skills.

Without MCPCopy Anchor Link

Any assistant with a terminal can also run the CLI directly (typestrict status --json, typestrict run...). Add this to its instruction file (AGENTS.md, CLAUDE.md...):

## typestrict

- typestrict trains small decision models on our data. Use the typestrict CLI (install: curl -fsSL https://typestrict.ai/install.sh | sh).
- Docs for assistants: https://docs.typestrict.ai/llms-full.txt. Read it before the first command.
- Drive each training by its `next_action` (`typestrict status <id> --json`).
- Starting a training costs a credit: show the cost and ask before `typestrict run`.
- Never print the API key or commit it.