Saer and AI coding agents

In short

saer mcp, a local MCP server in the beta, gives Claude Code, Cursor and Codex your tokens, components, hooks, quality warnings, a recommendation for each task and a check of their own changes. system.json is also a machine-readable record of the system that exists in your code.

Today

  • Give an agent the tokens and values from system.json, resolved per theme and mode.
  • Scan after the agent's change. Exact drift is a mechanical fix; near-matches go to a person.
  • Use the component inventory to remove duplicate names before an agent has to choose between them.

Connect your agent

From your project folder, run the command for your coding agent once. Saer checks premium access, refreshes a missing or stale scan and saves the project connection. You do not need to copy configuration or start a server in another terminal.

Terminal
saer agent codex
Terminal
saer agent claude
Terminal
saer agent cursor

Open a new agent session and approve Saer if asked. Configuration is scoped to this project; your global agent settings and other servers are preserved. Saer uses absolute launch paths so a GUI does not depend on your shell PATH. No sign-in credentials are copied into agent configuration. A custom Saer entry is preserved rather than replaced. Newly created connection files are locally excluded from Git commits when Git is available; existing configuration is not hidden.

An empty project can connect too: Saer says that no interface source exists yet rather than inventing system evidence. You can also use /agent codex, /agent claude or /agent cursor in the interactive workspace. Setup does not enable cloud uploads. If you already approved cloud sync for this project, normal agent rescans keep its approved numerical history up to date.

Ask your agent to call saer_overview for coverage and next actions. When no existing reusable component matches, a review-only local foundation proposal can be offered without pretending it is an existing component. Approve chosen changes before adoption. Then inspect the relevant tokens and components, use saer_recommend before building, then saer_check_changes before finishing. After edits, call saer_rescan or run the full scan again. Your agent provider’s own data handling still applies to context you give it.

  • saer mcp is a local MCP server for Claude Code, Cursor and Codex, read-only, over the repository's scan.
  • saer_recommend turns a task ("add a confirm-delete dialog") into the existing components, hooks and variants to build it from, and names the screen the change probably belongs on. Deterministic: no model, every answer cites its evidence.
  • saer_component, saer_tokens and saer_value give props, variants, real usage examples and the token for any raw value.
  • saer_check_changes checks only the lines the agent wrote, against the same rules as a scan: hard-coded token values, off-scale values, invalid variants, duplicate components and static accessibility findings.

Return, check or disconnect

Keep using your agent normally. It starts the local server for the connected project. Refresh with saer_rescan after edits and check changed lines before finishing. You do not have to reconnect after each scan.

Terminal
saer agent status
saer agent check codex
saer agent disconnect codex

The connection check verifies active access and a fresh local scan. Disconnect removes only the entry Saer created and keeps other servers and reports. It also works after premium access ends. If you deliberately edit Saer's entry, automatic replacement or removal stops until you review it. Private backups of previous configuration are kept in your Saer user configuration folder.

For another MCP client, configure a local stdio server with your absolute Saer launcher path and arguments mcp plus your absolute project path. Run saer mcp PROJECT to inspect the protocol manually; a quiet terminal is expected. The connection is local, not a hosted MCP URL.

In development

In development

Not available yet. Described here so you know what is coming, not as something you can use.

  • A measured benchmark of coding-agent output with and without Saer. A first round has run; results are published with their full method once reviewed.

See it on your own codebase.

Request early access