Skip to main content
MCP and Buttons meet in two directions. One ships today; the other is planned.
  • Buttons → MCP (shipped): buttons mcp runs Buttons as an MCP server, so any agent can list and press your buttons as deterministic tools.
  • MCP → Buttons (planned): capture an agent’s exploration of an MCP server — list tools, find the right tool and parameter shape, confirm the call works — and freeze that known-good invocation as a button. This is the page’s namesake feature and is not built yet.

​
What ships today: expose buttons as MCP tools

buttons mcp runs an MCP server over stdio that exposes a small meta-tool surface to any MCP client (Claude Desktop, Cursor, etc.):
  • buttons_list — list buttons available as tools
  • buttons_press — press a button by name with args
  • buttons_inspect — read a button’s spec
  • buttons_create — only when started with --allow-create (off by default)
Point an MCP client at it:
Add "--allow-create" to args to also expose buttons_create. Only buttons that opt in are visible or pressable: set "mcp_enabled": true in the button’s button.json. The server is rate-limited (10 calls/min, 1 concurrent, 120s cap per call) so an agent can’t hammer it. This is the inverse of the planned “freeze an MCP tool” flow: instead of wrapping someone else’s MCP tool as a button, you take any button you already have — shell, HTTP, anything — mark it mcp_enabled, and it becomes a scoped, deterministic tool an agent can call without re-discovering anything.

​
What’s planned: MCP → Buttons

The MCP-tool-to-button import is not implemented yet. The intended entry point is buttons import mcp <server>, which today returns an IMPORT_ERROR (“not implemented yet”). Only import code, import skill, and import url create buttons right now. There is no buttons create --mcp flag.
When it lands, the idea is to save a frozen MCP invocation as a button so future presses skip the exploration entirely and go straight to the known-good call:
The same button would then work in two contexts:
  • An agent reviewing the button sees it’s a safe, scoped call — no re-discovery of the MCP server required
  • A deterministic workflow (drawer, cron, webhook) can press it without involving an LLM at all — the MCP call is already parameterized and approved

​
Why this matters

Without this, every agent session has to rediscover the same tools: list servers, list tools, read schemas, guess parameters, verify the call. Freezing that one-time exploration into a button turns it into a reusable asset — and, pressed as JSON, returns the tool’s structured response in the same envelope as every other button.