For AI agents

What Systemix does for your stack

The design system is a machine-readable contract: plain MDX and JSON in the repo, exposed as tools by the systemix-mcp server. Any MCP-capable agent can query components and tokens, open and measure experiments, and write evidence back — the same files humans edit, no bespoke integration.

Autonomy is a dial, not a switch: ghost / assisted / autonomous tiers decide what agents write directly vs propose. Self-modification (skills, guardrails) is always human-approved, at every tier. Agents propose; humans close.

Your 10-minute start

1. Get the server. npx @getsystemix/cli init registers systemix-mcp in the repo's .claude.json (and Claude Desktop, if present). It runs from packages/mcp-server with --project-root <your repo>.

2. Read before writing. Start with experiment_list, experiment_learnings (the memory ledger), contract_get_token, contract_list_drifted — the instance's current truth.

3. Drive the loop.

   Tool	What it does
   experiment_new	author an experiment file
   experiment_measure	write evidence into the frontmatter
   experiment_close	record result · decision · confidence, append the learning
   push_hitl_task / resolve_hitl_task	queue and resolve decision cards
   contract_write_evidence	attach evidence to a contract record
   emit_event	leave a trace in the event log

4. Respect the gates. Writes that matter route through the HITL queue (.systemix/queue.json). If a tool returns a queued card instead of a direct write, that's the covenant working — don't route around it.

Your weekly loop

• Cite what you use: when a learning from experiment_learnings informs an action, the Used by: backlink keeps the memory graph honest.
• Check contract_list_drifted before generating UI — drifted tokens are the ones a human hasn't ruled on yet.
• New capability worth keeping? Propose it as a skill via the HITL queue — skills are always human-approved.

Where next

• The contract file structure
• Why files & loops — why state lives in files, not sessions
• Your landing page: /for/agents