SeaMeet MCP

@seameet/mcp

The SeaMeet MCP server — let AI agents record, transcribe, and mine your meetings. One install works whether or not you have the desktop app, in Claude Code, Claude Desktop, Codex, Antigravity, Cursor, OpenCode, GitHub Copilot CLI, Windsurf, and any MCP client.

It picks the richest available backend automatically and exposes the union of both:

• Desktop mode — when the SeaMeet desktop app is running, agents get the full recorder toolset (start/stop recordings, screenshots, live transcript, AI artifacts, search). No auth; nothing leaves your machine.
• Cloud mode — when the app isn't there, agents read your synced library (recordings, transcripts, summaries), upload a local audio/video file to be transcribed, and manage webhooks over the network. Authorized by logging into the web app — no key copy/paste, works on a headless terminal.

tools/list returns the superset of whatever's available; call seameet_status any time to see the current mode(s).

Using Claude or ChatGPT in a chat app? You don't need this package. Add the hosted SeaMeet connector — https://mcp.seameet.ai/mcp — and sign in with your SeaMeet account. No install, no API key. See Hosted connector.

What you can do

🎙️ Record & capture (desktop mode)

• Start a recording — microphone, screen, or screen + system audio — seameet_start_recording
• Stop — seameet_stop_recording · pause/resume audio-only — seameet_pause_recording / seameet_resume_recording
• Check recording status — seameet_recording_status · read the live transcript mid-recording — seameet_get_live_transcript
• Take a screenshot through SeaMeet — seameet_take_screenshot

📚 Work with recordings (desktop = local files · cloud = synced library)

• List recordings & screenshots — seameet_list_recordings · browse files in the save directory — seameet_list_files · app settings — seameet_get_settings
• Read generated artifacts (summary, transcript, SRT, chapters, action items, key decisions, screenshot OCR/description) — seameet_get_artifact
• Search across artifact text — seameet_search_text
• Fetch a recording's asset-bundle manifest — seameet_get_asset_bundle
• Rename a file — seameet_rename_file · save a new artifact — seameet_save_artifact · regenerate a summary from a template — seameet_regenerate_summary · list templates — seameet_list_templates

☁️ Cloud library (cloud mode)

• List recent synced recordings, including files you transcribed on app.seameet.ai — seameet_list_recent_recordings · fetch the summary, the speaker-labeled transcript (paged for long recordings) and chapters, plus action items and key decisions where the summary produced them; the transcript also comes as the same SRT or VTT file app.seameet.ai downloads — seameet_get_recording · one section (summary, action items, decisions, chapters) of up to 10 recordings at once — seameet_get_recordings
• Transcribe a file from this computer — upload a local audio or video file (mp3, m4a, mp4, mov, wav, flac, ogg, oga, opus, webm, aac, amr, spx; up to 512 MB and 5 hours) and start a speaker-labeled transcription; it returns an assetId right away and tells the agent when to poll seameet_get_recording — seameet_transcribe_file (runs in this local client, so it works in Claude Code, Codex, Cursor and Claude Desktop, but not from claude.ai on the web; uses your transcription allowance)
• Get a temporary media stream/download URL — seameet_get_media_url
• Search your recordings and the ones shared with you — seameet_search_recordings (full snippets need Sync Pro)
• Meeting Memory — find people, organizations and topics — seameet_search_memory · read a Memory page with the meetings it cites — seameet_get_memory_page
• Catch up across 2–5 related recordings, oldest first, with citations — seameet_whats_changed
• Rename speakers — seameet_rename_speakers · regenerate a recording's summary — seameet_regenerate_recording_summary · rename a recording — seameet_rename_recording
• Move up to 20 recordings to the Trash (restorable for 30 days) — seameet_trash_recordings · restore them — seameet_restore_recordings · list the Trash — seameet_list_recent_recordings with trashed: true (deleting forever stays on app.seameet.ai)
• Check storage usage — seameet_get_usage · how to record with the desktop app — seameet_desktop_recording_help
• Manage outbound webhooks (signed deliveries on ai.ready / recording.synced; creating one needs Sync Pro) — seameet_create_webhook · seameet_list_webhooks · seameet_update_webhook · seameet_delete_webhook

🔌 Connection (any mode)

• Which modes are connected — seameet_status · disconnect cloud / switch accounts — seameet_logout

That's 17 desktop tools + 19 cloud tools (fetched live, so new releases appear automatically), plus 3 tools that run in this client (seameet_status, seameet_logout, seameet_transcribe_file). A tool the current backend can't serve returns a clear, structured error instead of failing silently.

Hosted connector (no install)

SeaMeet also runs a hosted MCP server at https://mcp.seameet.ai/mcp. Add it to Claude or ChatGPT as a connector and sign in with your SeaMeet account (OAuth): no Node.js, no npx, no API key to copy. Any verified SeaMeet account works, Free included; Sync Pro unlocks full search snippets, full Meeting Memory pages, and webhooks.

Which one should I use?

	Hosted connector	@seameet/mcp (this package)
Install	none, add a URL	npx -y @seameet/mcp
Works in	Claude (web, desktop, mobile), ChatGPT (web, desktop), Claude Code, any client that speaks remote MCP with OAuth	Claude Code, Claude Desktop, Codex, Cursor, Antigravity, OpenCode, Copilot CLI, any stdio MCP client
Your cloud library (recordings, transcripts, search, Memory, webhooks)	✓ 19 tools	✓ the same 19, relayed
Record, screenshot, live transcript from the desktop app	—	✓ 17 desktop tools
Transcribe a file on this computer	—	✓ seameet_transcribe_file

In short: chat apps and cloud-only use → the hosted connector. A coding agent on the same machine as the SeaMeet desktop app, or uploading local files → this package.

Add it:

• Claude on the web — open claude.ai/new#customize/connectors/yours (or Settings → Connectors) → Add custom connector → name SeaMeet, URL https://mcp.seameet.ai/mcp → Add → Connect → sign in on app.seameet.ai → Allow. A connector added on claude.ai also shows up in the Claude desktop and mobile apps.
• Claude desktop app — Settings → Connectors → Add custom connector, same URL, same sign-in.
• ChatGPT on the web — open chatgpt.com/plugins → add an MCP server with the URL above → Authenticate → sign in → Allow.
• ChatGPT desktop app — Plugins → MCP → add a server, transport Streamable HTTP, the URL above, token/header fields left empty → Authenticate (no restart needed). Your browser opens app.seameet.ai/connect → sign in → Allow. The consent page notes that this app finishes signing in on your own computer (a local 127.0.0.1 address); that is expected for desktop apps.
• Claude Code — claude mcp add --transport http seameet https://mcp.seameet.ai/mcp, then run /mcp to sign in. (Should work; not yet verified end to end.)

Claude asks for read access; the ChatGPT desktop app also asks for write. Read-only tools carry readOnlyHint, so clients that honor it can tell them apart from the write tools.

Hosted tools (19):

Tool	Access	What it does
seameet_list_recent_recordings	read	Your most recent synced recordings
seameet_get_recording	read	A recording's summary, transcript (also as SRT or VTT), notes, chapters and more
seameet_get_recordings	read	One section of up to 10 recordings at once, e.g. this week's action items
seameet_search_recordings	read	Full-text search over your recordings and the ones shared with you (snippets need Sync Pro)
seameet_search_memory	read	Find people, organizations and topics in Meeting Memory
seameet_get_memory_page	read	A Meeting Memory page with the meetings it cites
seameet_whats_changed	read	Catch up across 2–5 related recordings, oldest first, with citations
seameet_get_media_url	read	A temporary stream/download link for a recording
seameet_get_usage	read	Storage used versus your quota
seameet_desktop_recording_help	read	How to record with the SeaMeet desktop app
seameet_list_webhooks	read	Your webhooks, their delivery health and signing secret
seameet_rename_speakers	write	Rename the speakers in a transcript
seameet_regenerate_recording_summary	write	Regenerate a recording's summary
seameet_rename_recording	write	Rename a recording
seameet_trash_recordings	write	Move up to 20 recordings to the Trash (restorable for 30 days)
seameet_restore_recordings	write	Restore recordings from the Trash
seameet_create_webhook	write	Create a webhook (Sync Pro)
seameet_update_webhook	write	Change a webhook
seameet_delete_webhook	write	Delete a webhook

Disconnect: open app.seameet.ai/connectors (profile menu → Connect with Claude/ChatGPT), click Disconnect next to the app, then remove the connector in the app.

Install

Every client runs the same command — npx -y @seameet/mcp — only the config format differs. Full per-tool recipes: INSTALL.md (mirrored at app.seameet.ai/mcp/install.md).

Let your agent do it — already inside a coding agent? Paste:

Install the SeaMeet MCP server — fetch https://raw.githubusercontent.com/seameet-ai/seameet-mcp/main/INSTALL.md, apply the section for whichever tool you're running in, and tell me how to reload.

Or pick your tool:

Tool	Fastest install
Claude Code	claude mcp add seameet -- npx -y @seameet/mcp
Codex CLI / IDE	codex mcp add seameet -- npx -y @seameet/mcp
GitHub Copilot CLI	copilot mcp add seameet -- npx -y @seameet/mcp
Cursor	Add to Cursor (one-click)
Claude Desktop	one-click .mcpb bundle — see below
Antigravity · OpenCode · anything else	see INSTALL.md

The generic block, accepted by most MCP clients:

{
  "mcpServers": {
    "seameet": { "command": "npx", "args": ["-y", "@seameet/mcp"] }
  }
}

Requirements

• Node.js ≥ 18.
• For desktop mode: the SeaMeet desktop app installed and running — brew install --cask seameet-ai/tap/seameet (macOS) or winget install seameet (Windows). Needs v3.2.0+; an older running app reports app_outdated.
• Cloud mode is opt-in and needs no app — the first cloud tool call starts a one-time authorization (below).

Cloud authorization (one-time, no key copy/paste)

Cloud mode never activates unless you provide SEAMEET_API_KEY or complete the device flow. On the first cloud tool call with no key, the agent shows a short code + https://app.seameet.ai/link; you open it (signed in), click Authorize, and a read+write key is minted and cached at ~/.seameet/credentials.json. Silent thereafter; revoke any time under API keys on your account. To disconnect or switch accounts, call seameet_logout (or rm ~/.seameet/credentials.json).

Transcribing large files

seameet_transcribe_file uploads in 8 MB parts and sends stage notifications (notifications/progress: reading the file, checking your allowance, each part starting and finishing, finishing the upload, starting transcription), at most one per second. Hosts that reset their tool timeout on progress keep waiting while parts move, but a stalled transfer sends nothing, and hosts that don't reset on progress may cut off a long upload either way: in Claude Code, raise the tool timeout for big files, e.g. MCP_TOOL_TIMEOUT=600000 claude (10 minutes). Transcription itself runs in the cloud after the tool returns, so only the upload counts.

The file is opened once, before anything is sent, and every part is read from that open file. If it is rewritten, grows, or is replaced at its path during the upload, the upload is stopped with file_changed rather than sending a mix of old and new bytes. A recording whose header says it is longer than 5 hours is refused before anything is sent.

Claude Desktop one-click bundle

manifest.json is an MCPB bundle manifest, so the server ships as a one-click Claude Desktop extension:

npm install && npm run build:mcpb   # → dist/seameet.mcpb

Install via Settings → Extensions → Advanced → Install extension… and pick dist/seameet.mcpb. Claude Desktop runs it with its own Node.js — no terminal needed. (The desktop app still needs to be installed + running for desktop mode.)

Claude Code plugin (adds a /seameet skill)

This repo doubles as a Claude Code plugin marketplace. The plugin registers the MCP server and a /seameet skill that teaches the core workflows (record + transcribe, capture a bug report, mine meetings for action items):

claude plugin marketplace add seameet-ai/seameet-mcp
claude plugin install seameet@seameet

Try it

Just ask in plain language — the agent picks the right tool (and the right backend). A few starters:

Record my screen for 2 minutes, then give me the transcript.

What were the action items from my last meeting?

Take a screenshot and describe what's on it.

Transcribe ~/Downloads/interview.m4a

Desktop mode — a full recording session (app running)

• "What are my SeaMeet settings — where does it save recordings?"
• "List my summary templates."
• "List my local recordings." · "List the files in my SeaMeet save folder."
• "What's my current recording status?"
• "Start recording my microphone." (or: "record my whole screen")
• "Check the recording status now."
• "Show me the live transcript so far."
• "Pause the recording." · "Resume it."
• "Take a screenshot."
• "Stop the recording."
• "List my recordings again — the new one should be on top."
• "Give me the summary and transcript of that recording."
• "Regenerate its summary using one of my templates."
• "Search my recordings for the word ''."
• "Rename that recording to 'MCP dual-mode test'."
• "Get the full asset bundle for it."
• "Save this text as a note/artifact on that recording: ."

Cloud mode — your synced library (no app needed)

• "List my 5 most recent SeaMeet recordings." · "What did I record in the last 7 days?"
• "Give me the summary of the top recording."
• "Now show its action items and key decisions."
• "Pull the full speaker-labeled transcript for it, and its chapters."
• "Get me a download/stream link for that recording's media."
• "Search my meetings for 'pricing'." · "What does Meeting Memory know about Acme?"
• "What changed across my last three standups?"
• "Transcribe ~/Downloads/interview.m4a" (uploads it, then checks back until the transcript is ready)
• "How much SeaMeet storage am I using versus my quota?"
• "Create a webhook to https://webhook.site/<your-id> that fires when AI results are ready."
• "List my webhooks with their delivery health and signing secret."
• "Update that webhook to also fire on recording.synced."
• "Delete that webhook."
• "What's my SeaMeet connection status?"
• "Disconnect my SeaMeet cloud account." (forgets the cached key; the next cloud call re-authorizes)

How it works

A thin stdio↔HTTP proxy. The SeaMeet desktop app hosts a localhost-only bridge (127.0.0.1:3741, fallbacks 3742/3743) authenticated with a per-launch random secret; on startup it writes credentials to $TMPDIR/seameet-mcp-bridge-<username>.json (mode 0600), which this server reads and proxies (tools/list → GET /mcp-bridge/tools, tools/call → POST /mcp-bridge/call-tool). Nothing leaves your machine in desktop mode. Cloud mode talks to the hosted SeaMeet worker (the same server as the hosted connector) over HTTPS with your minted key. Every request to SeaMeet's cloud carries User-Agent: seameet-mcp/<version>.

Full tool reference (for LLMs): seameet.ai/llms.txt — also served locally at http://localhost:3741/llms.txt while the app runs.

Errors are machine-readable

Tool failures return structured JSON your agent can branch on:

{ "success": false, "error": { "code": "app_not_running", "tool": "seameet_start_recording",
  "message": "The SeaMeet desktop app is not running.",
  "install": { "macos": "brew install --cask seameet-ai/tap/seameet", "windows": "winget install seameet" },
  "hint": "Ask the user to install and launch it, then retry." } }

Code	Meaning
app_not_running	Desktop app is closed — launch it (payload carries install commands + downloadUrl)
app_outdated	App is running but too old — update to requiredVersion (don't reinstall the same build)
auth_required	Cloud tool needs authorization — the payload has a user_code + the /link URL
insufficient_scope	The API key is read-only — authorize a read+write key
unsupported_type / too_large / empty_file	seameet_transcribe_file refused the file before uploading anything (a known length over 5 hours is session_too_long, also before uploading)
file_changed	The file changed during the upload; the upload was stopped — call again once the file is finished being written
insufficient_allowance / free_exhausted / import_cap_reached / daily_import_limit / queue_full / session_too_long	Not enough transcription allowance or room right now — the hint says what to tell the user
not_entitled	This account can't transcribe files — check the plan at app.seameet.ai
disabled	Transcription is turned off for this account — contact support
upload_failed	The upload stopped; step says where (upsert-asset, create, part <n>, complete) and the partial upload is aborted (the asset entry may remain in your library). cleanup: "failed" means the abort itself failed; SeaMeet clears stale partial uploads on its own
upload_unknown	The upload was sent but its completion wasn't confirmed (network drop at complete) — open webUrl to check before uploading again
service_unavailable	SeaMeet is briefly unavailable — retry in a minute. During the upload step says where it stopped; once the upload completed, the payload carries webUrl to start transcription there
job_start_failed	The file is in the library (assetId, webUrl) but transcription could not start — open webUrl to start it; don't re-run the tool, which would upload a duplicate
upload_incomplete	SeaMeet was still finishing the upload after one retry — open webUrl in a minute to start transcription
cancelled	The host cancelled the call before the transcription request was sent; the partial upload is aborted and no transcription starts
outcome_unknown	The transcription request was sent but no answer came back (cancel, timeout or network drop) — it may have started, so check asset.transcriptionJob with seameet_get_recording before doing anything else
app_not_ready	App is starting up — retry in a few seconds
invalid_request	A required parameter is missing/invalid — re-check the tool schema
path_forbidden	filePath must be inside the SeaMeet save directory
not_found	File or artifact doesn't exist — discover it with a list/search tool first
timeout	App didn't respond in time — check status, retry once
unknown_tool	Includes did_you_mean suggestions

Configuration (rarely needed)

Env var	Purpose
SEAMEET_API_KEY	Cloud API key (smk_…) — skips the device authorization flow
SEAMEET_MCP_CREDENTIALS_FILE	Explicit path to the desktop-bridge credentials file
SEAMEET_BRIDGE_PORT + SEAMEET_BRIDGE_SECRET	Bypass the bridge credentials file entirely
SEAMEET_CLOUD_CREDENTIALS_FILE	Where the minted cloud key is cached (default ~/.seameet/credentials.json)
SEAMEET_REMOTE_URL / SEAMEET_DEVICE_URL	Override the cloud endpoints (default: production)
SEAMEET_SUPABASE_URL / SEAMEET_SUPABASE_ANON_KEY / SEAMEET_STT_PROXY_URL / SEAMEET_WEB_URL	Override the endpoints seameet_transcribe_file uploads to and links (default: production)

Development

npm install
npm test             # wrapper tests (fake bridge + fake cloud) and upload tests (mocked fetch + temp files)
npm run build:mcpb   # validate manifest.json + pack the Claude Desktop extension

Releasing is automated — see PUBLISHING.md.

Legal

• Terms: https://seameet.ai/en/terms/
• Privacy: https://seameet.ai/en/privacy/

License

MIT © SeaMeet.ai