screenpipe

Screenpipe MCP Server

https://github.com/user-attachments/assets/7466a689-7703-4f0b-b3e1-b1cb9ed70cff

MCP server for screenpipe - search your screen recordings, audio transcriptions, and control your computer with AI.

Usage attribution

When Screenpipe analytics are enabled, successful nonempty retrievals produce qualified_value_event with surface (api or mcp) and agent_client. Supported app identifiers are claude, chatgpt, codex, cursor, gemini, openclaw, hermes, runner, windsurf, vscode, grok, and grokbot. Claude Desktop and Claude Code share the existing claude bucket. This identifies the caller, not the app appearing in a recording.

MCP uses SCREENPIPE_MCP_CLIENT when it names a supported app, falling back to the initialization handshake's client name. HTTP sessions resolve independently. Only the fixed identifier is forwarded; raw client names are never analytics properties. Unrecognized MCP clients remain unknown.

Installed REST skills send X-Screenpipe-Client: api and a fixed X-Screenpipe-Agent header for their target app. Other API clients can send the same headers. Missing or unrecognized app identifiers retain the legacy direct_api bucket on qualified-value events. search_performed also includes the fixed agent_client (unknown when unidentified), alongside request_source. These are self-reported identifiers, not verified app identities; copied skills and callers that omit the headers cannot be reliably attributed. Historical generic API events cannot be assigned an app retroactively.

The engine samples search_performed for every source, including app, MCP, named agents and callers without attribution headers. Each fixed source/client pair gets its first event per 60-second monotonic window per engine process, separately for empty/non-empty and parsed/ordinary results (at most four events per pair per window). The first event in each category after UTC midnight is also preserved. One client's polling cannot consume another client's budget.

API/MCP qualified_value_event emissions have an independent first-per-minute budget per surface, fixed client and value kind: search, meeting, artifact retrieved or artifact created. This covers cached API searches, activity summaries and MCP outcomes. Explicit frontend value actions are not sampled.

Both event families carry telemetry_sample_interval_seconds: 60 and telemetry_sampling: "first_per_interval_per_source_client_result_kind" for search diagnostics, or "first_per_interval_per_surface_client_value_kind" for qualified values. These are presence observations, not exact retrieval counts or unbiased samples of query properties. Counts and property breakdowns are not comparable to earlier releases. api_usage_5min.request_count still counts aggregate traffic across all routes. No search diagnostics are added for cache hits; actual searches, recording and local monitoring are unaffected.

Resource events additionally report telemetry_window_samples, window_cpu_percent_mean, window_cpu_percent_max, window_memory_gb_mean, window_memory_gb_max, and window_phys_footprint_gb_max across local samples since the previous reporting attempt. These supplement the existing point-in-time properties. Weight window means by their sample counts when aggregating them. Transport remains best effort; these summaries are not a durable sample archive.

Sampling preserves the first successful retrieval for daily presence and cohort-relative retention, not every occurrence. It reduces outgoing analytics volume after clients upgrade; actual billing savings depend on the account's pricing and usage tiers. Health/resource reporting is unchanged by retrieval sampling.

Attribution adds fields to existing events and uses the existing analytics opt-out. It does not scan running apps or send prompts, search text, or raw client names. HTTP MCP now reports successful searches through the same best-effort local telemetry endpoint as stdio MCP; empty/failed searches do not produce qualified-value events.

Call diagnostics and source references

New engine builds accept POST /internal/telemetry/mcp-call through the same local API authentication and analytics consent boundary as other private MCP telemetry. mcp_tool_call_completed records one best-effort observation per completed dispatcher call: schema version, random request ID, package version, fixed client/tool/transport names, ok/empty/error, normalized error category, duration, content bytes, optional search count/truncation and known report loss. No arguments, result text, error strings or raw client names are included. The existing qualified_value_event definition is unchanged.

ok means the dispatcher returned without a protocol/tool error. It does not mean the answer was correct or the user completed their task. Search alone sets result_count and truncated; absent fields on other tools remain unknown. response_bytes counts UTF-8 text and base64 content bytes, not total serialized JSON or model tokens. Duration includes tool execution, excluding report delivery.

The request ID is returned in _meta["screenpipe/request_id"] and forwarded to the local API as x-screenpipe-request-id. It is one call's correlation key, not a user, session or task ID. This change does not add engine log persistence for that header. Local helpers report to the engine, which applies its telemetry opt-out. SCREENPIPE_DISABLE_TELEMETRY and SCREENPIPE_TELEMETRY_DISABLED also suppress call reporting; CI suppresses it by default. Delivery is capped at eight pending requests per reporter and expires after one second. Old engines, opt-out, backend outages and process exit can lose reports. dropped_reports reports known local losses on a later delivery, not complete observation coverage.

Screen/audio/input search results preserve backend source identifiers and deep links. Both package transports cap text at 1,000 characters per hit by default; max_content_length=0 requests full text for a scoped read. The cap excludes metadata, tags and images, so callers should also limit result count. Final pages do not suggest another offset. In stdio, pass a returned frame_id to frame-context or get-frame-elements to inspect evidence beyond the snippet. HTTP clients can open the returned Screenpipe link or make a narrower search. A missing source ID stays missing.

This describes this package, not the separate hosted personal or enterprise MCP services. A hung backend read no longer keeps a stdio helper alive after stdin EOF: shutdown is bounded to 500 ms and affects only that client's process. This does not diagnose desktop recording health or reap another client's process.

Installable plugins

Build the local Claude Desktop extension and the cloud plugin from source:

cd packages/screenpipe-mcp

bun install --frozen-lockfile bun run build:installers

• installers/screenpipe-local.mcpb: open in Claude Desktop's extension installer. Bundles the server and its dependencies; Claude supplies Node. Keep Screenpipe running. Includes an optional sensitive API key setting and API URL setting.
• installers/screenpipe-cloud.zip: portable remote plugin with OpenAI/Codex and Claude compatibility manifests. Requires Data Sync and OAuth sign-in; exposes only synced device listing and search. See the cloud setup guide.
• installers/SUBMISSION.md: listing copy, endpoint, permissions guidance, and acceptance cases for directory submission. Building does not publish a listing.

The archive builder uses an explicit file allowlist, so local environment files, credentials, development dependencies, and source tests are not packaged.

Installation

Option 1: The screenpipe desktop app (Recommended)

The most reliable setup is to install the screenpipe desktop app and connect Claude Desktop from Settings → Connections (or during onboarding). This writes a config that:

• uses the bundled `bun` shipped with the app (an absolute path — no Node/npx or PATH dependency, and ~3× faster cold start), and
• injects your `SCREENPIPE_LOCAL_API_KEY` into the server's env, so the MCP authenticates instantly instead of running slow key discovery at startup.

Both matter: a config without the key forces the server to discover it via subprocess fallbacks, which on a cold package cache can stall Claude Desktop's MCP startup and produce Could not attach to MCP server screenpipe.

Option 2: Manual NPX (no desktop app)

If you're not using the desktop app, edit your Claude Desktop config:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %AppData%\Claude\claude_desktop_config.json

{

"mcpServers": { "screenpipe": { "command": "npx", "args": ["-y", "screenpipe-mcp@latest"], "env": { "SCREENPIPE_LOCAL_API_KEY": "sp-…" } } } }

Requires Node/npx on PATH. Pin @latest so the first install doesn't cache a stale version forever. Get your key with screenpipe auth token. If you omit the key, the server will try to discover it through the bundled Screenpipe CLI and then npx. It never opens the live SQLite database directly. Discovery can be slower on first run, so setting the environment variable is still recommended.

Enterprise admins: add SCREENPIPE_ENTERPRISE_TOKEN (and, for orgs running their own query gateway, SCREENPIPE_TEAM_API_URL) to the same env block to get the team-* tools — see Enterprise team tools.

Option 3: HTTP Server (Remote / Network Access)

The MCP server can run over HTTP using the Streamable HTTP transport, allowing remote MCP clients to connect over the network instead of stdio. This is ideal when your AI assistant (e.g., OpenClaw) runs on a different machine than screenpipe.

# loopback only (default)

npx -y screenpipe-mcp --http --port 3031

expose to your LAN with bearer auth

npx -y screenpipe-mcp --http --listen-on-lan --api-key $(openssl rand -hex 16)

or from source — must build first so dist/ exists

bun install && bun run build bun run start:http -- --port 3031

Tip: npx screenpipe-mcp-http (without --http) does not work —

npm resolves by package name, and there is no screenpipe-mcp-http

package. The HTTP server ships as a transport inside the

screenpipe-mcp package; use --http as shown above, or invoke the

bin directly with npx -p screenpipe-mcp screenpipe-mcp-http.

The server exposes:

• MCP endpoint: http://localhost:3031/mcp — Streamable HTTP transport (POST for requests, GET for SSE stream)
• Health check: http://localhost:3031/health — always unauthenticated, for monitors

Options:

| --api-key | Bearer token required for non-loopback requests (`A