Integration
MCP Server — AI agents read your transcripts directly
Claude Desktop, Cursor, Zed and other MCP clients talk to your DeepScript library directly. No copy-paste, no third-party cloud.
Rather than building our own summarize, translate or search endpoints, DeepScript ships a full MCP server. AI clients like Claude Desktop, Claude Code or Cursor connect via Bearer auth to `https://mcp.deepscript.com` and can then search your transcripts, fetch individual ones or summarise them — using the LLM of their choice. Three tools are exposed: `search_transcripts` (full-text search across your library), `get_transcript` (complete transcript content including word timestamps) and `list_recent` (the last N transcriptions). The MCP server is part of the Pro plan (€22/month) because it removes the default 30-day retention and unlocks unlimited agent traffic. Auth tokens are the same `ds_live_` API keys used by the REST API.
What you can build
- Ask Claude in the desktop app: "What was the supplier's name in last week's meeting?" — Claude searches your transcripts.
- Summarise podcast episodes inside your IDE agent without copy-pasting them.
- Extract interview quotes via the agent and write them into Markdown notes (e.g. in Cursor or Zed).
- Compliance audits: have the agent check whether specific terms appeared in archived meetings.
- Cross-recording research: "Find all meetings that discussed Project X."
Code samples
{
"mcpServers": {
"deepscript": {
"type": "http",
"url": "https://mcp.deepscript.com",
"headers": {
"Authorization": "Bearer ds_live_xxx"
}
}
}
}{
"servers": {
"deepscript": {
"transport": "http",
"url": "https://mcp.deepscript.com",
"auth": {
"type": "bearer",
"token": "ds_live_xxx"
}
}
}
}{
"tools": [
{
"name": "search_transcripts",
"description": "Full-text search across the user's transcript library. Returns the top N hits with snippet, transcriptId and createdAt.",
"inputSchema": {
"type": "object",
"properties": {
"query": { "type": "string" },
"limit": { "type": "integer", "default": 10, "maximum": 50 }
},
"required": ["query"]
}
},
{
"name": "get_transcript",
"description": "Return the full text and word-level metadata for one transcription.",
"inputSchema": {
"type": "object",
"properties": {
"transcriptId": { "type": "string", "format": "uuid" },
"includeWords": { "type": "boolean", "default": false }
},
"required": ["transcriptId"]
}
},
{
"name": "list_recent",
"description": "List the N most recent transcriptions with filename, duration and createdAt.",
"inputSchema": {
"type": "object",
"properties": {
"limit": { "type": "integer", "default": 20, "maximum": 100 }
}
}
}
]
}Setup in a few steps
- 1
Activate the Pro plan
The MCP server is Pro-only (€22/month). Upgrade in the billing dashboard. Pro simultaneously removes the 30-day auto-deletion, letting your agent search older transcripts too.
- 2
Use an API key as the bearer token
Generate a regular `ds_live_` key in the dashboard. The same key authenticates both MCP and REST calls. Recommendation: create a dedicated MCP key labelled accordingly so you can revoke it in isolation.
- 3
Configure your MCP client
Add the server entry shown above to Claude Desktop, Cursor or Zed. Restart the client. The three tools `search_transcripts`, `get_transcript` and `list_recent` appear automatically in the tool picker.
- 4
Use it in chat
Talk to the agent in natural language. It picks the right tool on its own. Example: "Search my transcripts for any mention of 'Q3 forecast'." — the agent calls `search_transcripts({ query: 'Q3 forecast' })`.
Frequently asked questions
Why is MCP Pro-only?
Agents typically issue many queries per conversation and need access to older transcripts. Pro removes the 30-day default retention and exempts MCP calls from the regular rate limit. For occasional usage, the REST endpoint on pay-per-use is enough.
Which transports does the server support?
Currently HTTP transport (Server-Sent Events for streaming responses). stdio is not supported because the server runs remotely. Most clients (Claude Desktop since 0.7, Cursor, Zed) support HTTP MCP out of the box.
Do AI providers see my transcripts?
Yes — once your local agent loads a transcript via `get_transcript`, it sends it as context to the LLM (e.g. Anthropic, OpenAI). To avoid that, run a local LLM client (e.g. Ollama via the Claude Code adapter) combined with DeepScript MCP — then data only ever lives on your machine plus our German server.
What's the MCP server availability?
The MCP server runs in the same Hetzner cluster as the REST API with our usual 99.9% SLA. Status and incidents are at deepscript.com/status. On outage the agent fails gracefully — tool calls return a clear error message.
Can I call MCP tools without an AI client?
Technically yes — MCP is an open JSON-RPC protocol, you can hit the tools via HTTP/curl. For pure server-to-server integrations the regular REST API is more practical though, since it's OpenAPI-typed and not agent-oriented.
Ready to ship this to production?
Create an account, generate an API key, ship. Three transcriptions free to try. Full OpenAPI 3.1 docs at api.deepscript.com/docs.