asistonto-mcp

An MCP server that teaches any AI assistant to build question packs for asistonto, a live meeting copilot.

asistonto listens to a meeting and, the moment a question it was told to watch for is actually being discussed, it surfaces that question. What it watches is a pack. Writing a good one used to be expert work: the questions have to be stated in a particular four-part format, and each of those parts belongs in a different field because of what was measured about retrieval — get that wrong and the pack loads fine and then stays dark all meeting.

This server hands that expertise to Claude, ChatGPT, Claude Code, Codex, or anything else that speaks MCP. You paste your meeting notes; the assistant writes the pack; the pack loads into asistonto unchanged.

Install

Nothing to install. Every client below runs it through uvx, which fetches and launches it on demand. You need uv and Python 3.12+.

uvx asistonto-mcp        # runs the server on stdio; Ctrl-D to exit

Until the package is published to PyPI, point uvx at the source instead — uvx --from git+https://github.com/<you>/asistonto-mcp asistonto-mcp, or uvx --from /path/to/asistonto-mcp asistonto-mcp from a local checkout — and use the same --from form in the client configs below.

Claude Code

claude mcp add asistonto -- uvx asistonto-mcp

Or, for everything on this machine rather than one project:

claude mcp add --scope user asistonto -- uvx asistonto-mcp

Claude Desktop

claude_desktop_config.json — macOS: ~/Library/Application Support/Claude/claude_desktop_config.json, Windows: %APPDATA%\Claude\claude_desktop_config.json.

{
  "mcpServers": {
    "asistonto": {
      "command": "uvx",
      "args": ["asistonto-mcp"]
    }
  }
}

Restart Claude Desktop afterwards.

Codex, Cursor, Windsurf, and anything else taking the generic block

The same mcpServers object, in whichever file that client reads (~/.codex/config.json, .cursor/mcp.json, …):

{
  "mcpServers": {
    "asistonto": {
      "command": "uvx",
      "args": ["asistonto-mcp"],
      "env": {}
    }
  }
}

Codex's TOML config takes the same three fields:

[mcp_servers.asistonto]
command = "uvx"
args = ["asistonto-mcp"]

From a checkout

git clone https://github.com/<you>/asistonto-mcp.git
uvx --from ./asistonto-mcp asistonto-mcp

Related MCP server: Office-PowerPoint-MCP

What it gives the assistant

Tools

Prompts — build_pack_for_meeting, write_paraphrases, formulate_question, review_pack.

Resources — asistonto://guide/pack-format, asistonto://guide/four-part-questions, asistonto://guide/paraphrases, asistonto://guide/calibration, asistonto://guide/workflow, asistonto://guide/labels, asistonto://example/pack, asistonto://schema/brief.json.

The resources matter as much as the tools. An assistant that reads the paraphrase guidance before writing phrasings produces a materially better pack than one that only calls functions — the quality lives in judgement, and the judgement is written down there.

Worked example, end to end

1. Ask. In any client with the server registered:

Build me an asistonto pack for Tuesday's meeting with the warehouse lead about the new label. Here are my notes: (paste). The spec says the label carries the lot code, but the picking screen we saw last week only shows the pallet id, and nobody has said which one the scanner reads.

2. The assistant states each question in four parts and calls compose_question:

which becomes one entry of brief.json:

{
  "id": "QC-7",
  "title": "what the scanner reads at picking",
  "priority": "alta",
  "source": "ETIQUETAS_v3.md",
  "line": 41,
  "probes": ["When the operator scans at picking, does the code identify the pallet or the lot?"],
  "people": ["warehouse lead"],
  "rationale": "the traceability chain hangs off which of the two is scanned",
  "blocking": true,
  "context": "ETIQUETAS_v3.md line 41 says the label carries the lot code. The picking screen shown on 12 May displays only the pallet id."
}

The provenance sits in context, which is never indexed — indexing it costs 15 points of recall@1. The question sits in probes, the label in title, and both of those are indexed.

3. Ten paraphrases per question, written from the question alone:

{
  "QC-7": [
    "when they scan at picking, what are they actually reading",
    "the guy with the gun scans the pallet or the box",
    "does the scan tell you which lot it is or just where it is",
    "if two lots sit on the same pallet, how does the scan tell them apart",
    "what comes up on the screen after the beep",
    "can you tell from the scan which batch went out",
    "is the code on the label the same one the system stores",
    "do they scan once per pallet or once per unit",
    "what happens when the label is torn and it does not scan",
    "who decided which code goes on that label"
  ]
}

4. validate_pack, then write_pack. You get a directory:

packs/almacen/
├── brief.json         ← the only file the live loop reads
├── paraphrases.json   ← the editable source, folded into the brief
└── background.txt     ← a PAST meeting, for calibration

5. Load it into asistonto. Copy that directory to packs/<name>/ on the machine running the copilot, then pick it by name when you start the session at https://asistonto.escorpia.com. Pack names must be letters, digits, dot, dash or underscore, starting with a letter or digit — the server rejects anything else.

The three things worth knowing before you start

Provenance is never a paraphrase. Parts 1 and 2 describe documents; paraphrases describe speech. Measured with paraphrases already indexed, adding the provenance took recall@1 from 71.1% down to 55.6%. Worse, writing paraphrases while reading the provenance pulls document vocabulary — codes, annex numbers, file names — into phrasings that nobody ever says.

Paraphrases must separate siblings, not reword them. Six ways of saying the same thing all resemble the neighbouring question equally, and 70% of measured top-1 errors landed on a sibling of the same block. check_paraphrases flags phrasings that fit the neighbour better than their own question.

A small pack cannot fire at all. The live gate is a z-score over the pack's own scores, so its ceiling is (n-1)/√n. Against the default gate of 4.75 that means at least 25 questions — below it the panel stays dark all meeting and looks exactly like a quiet room.

The measurements behind all of this are in the guide resources, which the assistant can read directly.

Development

git clone https://github.com/<you>/asistonto-mcp.git
cd asistonto-mcp
uv venv && uv pip install -e '.[dev]'
uv run pytest

The test suite runs the MCP server over a real client session and asserts that what it writes round-trips through a copy of the product's own loader (tests/test_product_compatibility.py). No network access, no product code.

License

MIT.