meeting-ledger

Your meetings, on the record: who decided what, who promised what, and whether it happened.

Feed it a transcript, your notes or the recording itself. It pulls out the decisions, action items, promises and open questions, each one pinned to the transcript lines it came from. Then it keeps track of them: when a later meeting finishes, blocks or reverses something, the ledger updates. When a promise stops coming up even though the person who made it keeps showing up, it flags it as stale.

_{Fictional demo data: try it yourself with meeting-ledger demo, no API key needed.}

Why

Note-takers summarise one meeting at a time. The things that slip happen between meetings: "I'll send the invoice" is said once and never comes up again, a decision quietly gets reversed two weeks later, a task gets "assigned" to someone who never actually agreed to it.

meeting-ledger splits the work three ways:

There's also a plain check that needs no model at all: every quote must literally appear in the lines it cites, or the item is flagged.

What it does

• Any input: WebVTT/SRT exports from Zoom, Teams or Meet, plain text or Markdown notes with or without Name: speaker labels, or an audio/video file transcribed on your Mac with Parakeet, so the recording never leaves the machine.
• Cited items: every item lists its transcript lines (L12). In the terminal UI, moving to an item highlights and scrolls to those lines.
• Follow-through across meetings:
  • When a later meeting mentions an earlier item, Claude proposes a status and the decision model checks it. If both agree with ≥ 70 % confidence, the change is applied and logged with the line that justified it. Otherwise it goes to review.
  • Contradictions (a reversed decision, a withdrawn promise) always go to review.
  • Items whose owner attended but nobody mentioned count a miss. Two misses (configurable) and the item is stale.
• Ask your meetings: meeting-ledger ask "what did we decide about the online shop?" finds the matching lines across every meeting and gets Claude to answer with real citations (search-result blocks), so each sentence points to a meeting and its lines.
• Terminal UI: meetings, open items, the review queue and ask, in one keyboard-driven screen.
• MCP server: Claude Desktop, Claude Code or any MCP client can search your meetings and list or update open items.
• Obsidian: each meeting becomes a note with YAML frontmatter, a checklist, line references, the transcript, and [[wikilinks]] to people who already have a note in your vault.

Quickstart

pipx install meeting-ledger
meeting-ledger demo            # the terminal UI on fictional meetings, no key needed

With your own meetings:

echo 'ANTHROPIC_API_KEY=sk-ant-...' >> ~/.meeting-ledger/.env

meeting-ledger add "standup 2026-10-07.vtt"   # the date comes from the file name, or --date
meeting-ledger add notes.txt --title "Board call" -o board-call.md
pbpaste | meeting-ledger add -                # paste notes
meeting-ledger                                # open the terminal UI

Audio and video (Apple Silicon; needs brew install ffmpeg):

pipx install 'meeting-ledger[audio]'
meeting-ledger add call.m4a

Everything else is a command too:

Terminal UI keys: 1–4 switch tabs (Meetings, Open items, Review, Ask). On items: d done, b blocked, x dropped, o reopen, s stale only, Enter opens the item's meeting. In review: a accept, c use the checker's status, r reject.

The decision model: Jev, Kev, or neither

Jev is a System One model: you ask typed questions (Noul: how likely is this statement to be true; Choice: which of these labels), and each answer comes with a calibrated probability, typically in well under a second. That's the right tool for checking Claude: fast, cheap, and honest about when it isn't sure. meeting-ledger's questions are in decisions.py.

Pick a backend with environment variables (in ~/.meeting-ledger/.env):

Running Kev locally:

git clone https://github.com/jaredpalmer/kev && cd kev
uv sync --extra serve
uv run --extra serve python -m kev.serve --run jaredpalmer/kev-4b --port 8009   # kev-0.8b on a smaller Mac

Kev's own guidance: Kev-4B needs about 32 GB of memory and is much more accurate than Kev-0.8B, which fits any Apple Silicon Mac.

Measured on an 18 GB M3 Pro with Kev-0.8B (October 2026), running the demo meetings: follow-up checks took about 90 ms each and nothing was applied wrongly, but the model was rarely confident. Most follow-up checks scored below the 70 % threshold, and it flagged two correct items as unsupported. So with 0.8B the ledger still works, but most changes wait for you in review. For changes to be applied automatically, use Kev-4B or larger, or Jev.

What leaves your machine

Your ledger itself is one SQLite file, ~/.meeting-ledger/ledger.db.

MCP

Add to Claude Desktop's claude_desktop_config.json (or claude mcp add meeting-ledger -- meeting-ledger mcp in Claude Code):

{
  "mcpServers": {
    "meeting-ledger": { "command": "meeting-ledger", "args": ["mcp"] }
  }
}

Install with pipx install 'meeting-ledger[mcp]'. Tools: search, meetings, meeting, open_items (filter by owner or stale) and set_item_status. The server needs no API key: your MCP client's own model does the reasoning.

How it works

flowchart LR
  In[transcript · notes · audio] --> Ingest[ingest.py<br/>numbered lines]
  Audio[(Parakeet on your Mac)] -.-> Ingest
  Ingest --> Extract[extract.py<br/>Claude: items + proposed updates]
  Extract --> Ground[quote check]
  Ground --> Check[decisions.py<br/>Jev / Kev / adapter]
  Check -->|agree, confident| Ledger[(ledger.db)]
  Check -->|disagree, unsure, contradiction| Review[review queue]
  Review --> Ledger
  Ledger --> TUI[terminal UI] & MCP[MCP server] & Obsidian[Obsidian notes] & Ask[ask.py<br/>cited answers]

ledger.py holds every rule (grounding, flags, when an update is applied, staleness). Claude and the decision model are injected, so all of it is tested without a network.

Development

git clone https://github.com/owen-alderson/meeting-ledger && cd meeting-ledger
python -m venv .venv && . .venv/bin/activate
pip install -e '.[dev]'
pytest            # 160 tests, no network: Claude, Jev/Kev and Parakeet are faked

The Claude and TypeSafe tests run the real SDKs against a mocked HTTP transport, so the exact request shapes are checked. Releases: push a vX.Y.Z tag and GitHub Actions publishes to PyPI.

The original single-file CLI is kept at tag v1-cli.

License

MIT