sana-mcp

Give any AI agent instant, local access to your Sana.AI meeting transcripts - it runs a background daemon that syncs them into a local database, unlocking search, read, summarize, and more through a single MCP tool (meeting_transcripts) and a CLI.

One static binary, about 14 MB. Nothing to install alongside it.

macOS / Linux:

curl -fsSL https://github.com/laelhalawani/sana-mcp/releases/latest/download/install.sh | sh

Windows (PowerShell):

irm https://github.com/laelhalawani/sana-mcp/releases/latest/download/install.ps1 | iex

The installer downloads the binary for your OS and CPU, finds the AI clients on your machine, asks which to connect, signs you in to Sana, and shows your meetings downloading. Then run sana-mcp.

Linux x64/ARM64, macOS x64/Apple Silicon, and Windows x64/ARM64 are published.

What it does

Sana has no public API for meetings. sana-mcp drives the same web backend the Sana app uses (tRPC at sana.ai/x-api) with your logged-in session, over HTTPS with a cookie jar. No browser is required, so it runs headless on macOS, Linux, Windows, and WSL.

• A background daemon downloads your meetings, transcripts, and metadata into a local SQLite database, then checks periodically for new ones.
• The agent tools read only from that database, so they respond instantly and never block on the network. The sole exception is recording, which fetches a short-lived link on demand.
• Everything - session, database, logs - stays on your machine.

Features

• One tool for agents - meeting_transcripts - with built-in discovery: help lists everything, help {tool} shows exact arguments.
• List, read, search transcripts, plus summaries, participants, and on-demand recording links.
• Search - line-level SQLite FTS5 with BM25 ranking, phrase and date-range filters, and sort options.
• Automatic sync - a background daemon starts itself whenever you use sana-mcp, polls for new meetings, retries incomplete downloads, and never re-downloads a transcript it already has.
• Correct transcription errors - and keep the original. See below.
• Works with your client - registers with Claude Desktop, Claude Code, Cursor, VS Code, Codex, Gemini CLI, Windsurf, Zed, Cline, Roo Code, Amazon Q, Continue, opencode.
• Local-first and private - no data leaves your machine except authenticated requests to Sana.

Agent tools

Agents call one tool, meeting_transcripts, with a tool name and an optional args object:

meeting_transcripts("<tool>", { ...args })

meeting_transcripts("search", {"query": "pricing", "sort": "newest"})
meeting_transcripts("read",   {"meeting_id": "v72HzzJDZx9WqTmF", "lines": [22, 26]})

Full schemas and semantics: docs/tool-contract.md.

Use it from the CLI

sana-mcp                       # the full-screen application
sana-mcp status                # sync progress
sana-mcp list                  # meetings, newest first
sana-mcp search pricing        # search every transcript
sana-mcp read <meeting-id> 20 40
sana-mcp configure             # change which clients are connected
sana-mcp daemon --stop

A bare sana-mcp opens the application on a terminal, and serves MCP without one - so a client that runs the binary with no arguments still gets a server.

Fixing transcription errors

Transcripts come from speech recognition, and it gets names wrong. A discussion about Fabrix can be transcribed as Fabrik, and then no search for "Fabrix" will ever find it.

You can correct any line - in the application (e to edit, ctrl+s to save, h for history) or through the tools. Corrections are safe by construction:

• Nothing is destroyed. What Sana delivered is kept forever, shown beside every change, and restoring is one key.
• A correction survives a re-sync. Edits re-attach by content, so they follow their line even when re-transcription shifts the numbering. One that no longer matches anything is kept and marked, never applied to the wrong line.
• Both spellings stay searchable. The corrected name finds the line; what was actually said still finds it too, ranked well below.

A warning that also sits in the tool description and on the edit screen: meeting transcripts are full of product, company and personal names that look like misspellings and are not - Zenolith, Fabrix, Vantik are real. An agent must never edit a transcript unless you asked for that specific correction.

Where things live

Everything is under ~/.sana-mcp: the SQLite database, your session, the daemon log, and the binary itself in bin/.

sana-mcp uninstall removes all of it. It lists what it is about to remove and asks first: the client registrations, that directory, every copy of the binary the installers have written (including ~/.local/bin, where an older version put it), and the PATH line they added to your shell profile. It stops any running daemon first, by asking each installed binary to stop its own - a daemon started by a different version only answers to that version.

Note that the database holds your corrections, which cannot be re-synced from Sana. It is not purely a cache.

Installing over an older version

An older version's database cannot be read by this one. The installer checks for that before it writes anything, says what continuing costs - your meetings are downloaded again, and you may have to sign in again - and asks. Nothing is deleted until the new setup succeeds, so cancelling or a failure part-way through puts the old state back exactly as it was.

Development

go vet ./... && go test ./... && go test -race ./...

Live tests exercise the real API with your stored session; they are read-only, write to temporary databases, and are skipped unless SANA_LIVE=1.

Releasing is bumping VERSION and merging to main.

See AGENTS.md for the constraints that are load-bearing, docs/tool-contract.md for the agent surface, and docs/app-design.md for the application.

License

GPL-3.0