MCP Server
The Apex MCP server (@apex-inc/mcp-server) is how your coding agent talks to the same Apex your dashboard uses. Ask it to list experiments, switch shops, write a letter, or check what is live on a file. You stay in the editor.
One server. Not a server per product. Splitting Experiments from Letters would break workspace switching.
What you can ask
The editor Tools list is the live catalog. Do not treat this page as the count. The connected server registers every tool. They group the way a shop works:
| What you are doing | Examples |
|---|---|
| Point the session at the right shop | list_workspaces, switch_workspace, list_orgs, switch_org |
| Finish Set up Apex | get_setup_state, list_data_sources, get_wiring_status, list_conversions |
| Run an experiment | plan_experiment, create_experiment, check_live_tests, activate_experiment, get_results |
| Write a letter or Adaptive Journey | list_communications, list_journeys, list_segments, recommend_communications |
| Adoption | list_adoption_milestones, get_adoption_metrics |
| Partners | list_partner_programs, list_pending_payouts |
| Measurement | list_events, get_schema, get_channel_economics |
Reads that take no arguments were smoke-tested against an empty shop. Writes (invite, publish, approve a payout, mint a key) are real tools. Confirm before the agent runs them on a live shop.
Resources, prompts, and skills
Resources are read-only context the agent can open without a tool call:
| Resource | What it is |
|---|---|
apex://beliefs | Recorded beliefs |
apex://experiments/active | Experiments that are running |
apex://ledger | The decision log |
apex://intelligence | Intelligence breakdown |
apex://intelligence-score | The score |
apex://calibration | How well predictions match outcomes |
apex://adoption | Path progress |
apex://communications | Letters |
apex://communications/events | Letter events |
apex://data-sources | Website, apps, and server sources |
Prompts are conversation starters: new-experiment, experiment-review, apex-graduate, apex-experiment, apex-link, apex-followup, apex-login, apex-keys.
Skills ship with the package. They teach the agent how Apex works (experiments, letters, Adaptive Journeys, Adoption, Partner Network, the Apex Spec) instead of only listing tool names.
What the agent will not do for you
- It will not invent a conversion. Set conversions in Apex, or ask it to
create_conversion. - It will not see what a SKU costs unless you turn cost access on for that key in Settings → AI Agents. Cost of goods stays off by default.
- If you have never saved scoring weights, asking for the model returns that the vertical preset is still in use. That is not a failure.
- Experiments it creates are code experiments. Apex returns a recipe. The agent implements it. The snippet still changes pages you do not have in a repo.