MCP & Composio setup

Hermite's enterprise API registers as a Composio toolkit, which means any MCP-compatible agent — Claude, Cursor, or something you build — can call Hermite's capabilities as ordinary tool calls: generate an easing curve, transcribe a file, clean up captions.

Preview. This guide describes the integration as designed; it becomes usable when Stage 4 ships. The tool names and flows below are the contract we're building against.

How it fits together

Model Context Protocol (MCP) is the open standard agents use to discover and call external tools. Rather than hand-building and hosting our own MCP server, Hermite registers its API as a toolkit on Composio, which handles auth, tool discovery, and execution for MCP clients. You authenticate once with your Hermite API key; your agent sees a set of HERMITE_* tools.

Tool	Wraps
`HERMITE_GENERATE_CURVE`	POST /v1/curves/generate
`HERMITE_APPLY_PRESET`	POST /v1/presets/apply
`HERMITE_TRANSCRIBE`	POST /v1/transcribe
`HERMITE_CLEANUP_CAPTIONS`	POST /v1/captions/cleanup

Claude (claude.ai or Claude Code)

Add Hermite through Composio's MCP endpoint. For Claude Code, that's one command:

claude mcp add hermite \
  --transport http \
  https://mcp.composio.dev/hermite?api_key=<your-composio-key>

On claude.ai, add the same URL as a custom connector in Settings → Connectors. Claude then discovers the four tools automatically.

Cursor

In .cursor/mcp.json at your project root (or globally in ~/.cursor/mcp.json):

{
  "mcpServers": {
    "hermite": {
      "url": "https://mcp.composio.dev/hermite?api_key=<your-composio-key>"
    }
  }
}

Custom agent (Composio SDK)

import { Composio } from "@composio/core";

const composio = new Composio({ apiKey: process.env.COMPOSIO_API_KEY });
const tools = await composio.tools.get("default", {
  toolkits: ["HERMITE"],
});

// Hand the tools array to your LLM call; execute results with:
const result = await composio.tools.execute("HERMITE_GENERATE_CURVE", {
  arguments: { preset: "overshoot", duration_frames: 24 },
});

Example agent prompts

"Transcribe interview.wav with speaker labels, clean up the punctuation, and give me an SRT."
"Generate an overshoot easing curve over 18 frames and apply a slide-in-left preset for a 1080p clip."
"Batch-clean the captions in this JSON and flag any segment you changed significantly."

Endpoint shapes live in the API reference.