AI engineering portfolio project

Meeting transcripts in. Approved, tracked tasks out.

MCP Task Assistant lets an AI assistant turn a meeting transcript into structured action items, then waits for a human to approve before anything is saved to Google Sheets or announced in Slack.

The problem

The meeting ends. The real work starts.

After every meeting someone has to reread the transcript, pull out who agreed to do what, chase deadlines, update the tracker and tell the team. It is slow, easy to get wrong, and easy to forget.

Manual follow-up

Copying action items into spreadsheets and chat by hand costs time after every call, and the details drift.

Accountability gets fuzzy

Owners and deadlines are guessed or lost. A proposal that was only discussed gets recorded as a decision.

AI output needs a gate

Language models can invent names, dates and approvals. Automation should propose; a person should decide.

This project turns the tedious part into a tool an AI assistant can call, while keeping a person in charge of what becomes real.

Workflow

From conversation to accountable execution

  1. Submit transcript

    An assistant sends the meeting transcript, and optionally the meeting date.

  2. AI extracts action items

    Structured tasks with owner, deadline, priority, confidence and a quoted piece of evidence.

  3. Human reviews and approves

    Every task starts as pending. Only tasks a person explicitly approves move on.

  4. Saved to Google Sheets

    Approved tasks are appended to the team sheet. Duplicates are skipped.

  5. Sent to Slack

    The meeting summary and the approved tasks are posted once to the channel.

MCP tools

Three tools, one safe path

No side effects

extract_action_items

Analyzes a transcript with an LLM and returns proposed tasks. Nothing is saved or sent.

  • Owner and deadline only if the transcript states them
  • Verbatim evidence quote, checked against the transcript
  • Proposals and "maybe" ideas are not turned into tasks
Writes to Google Sheets

save_approved_tasks

Appends the tasks a human approved to the Google Sheet, following the sheet's own columns.

  • Needs explicit task ids and a human confirmation flag
  • Unverified evidence needs a separate acknowledgement
  • Re-submitting never creates duplicate rows
Posts to Slack

notify_slack

Posts the meeting summary and the approved, saved tasks to the team channel.

  • Unapproved tasks can never be announced
  • The same set of tasks is announced once
  • Mentions in task text are escaped
Architecture

A small, layered system

Clients reach one hardened container through TLS. The server validates everything, keeps each client's data in its own session, and only talks to three external services.

MCP clientClaude Code, others HTTPS ApacheTLS 1.2+, only /mcp proxy MCP server (Docker)non-root, read-only, no capabilities Bearer auth + rate limits Per-client sessions Zod + approval gate 3 MCP toolsextract · save · notify OpenRouterLLM extraction Google Sheetsapproved tasks Slackwebhook notification The same server also runs locally over stdio for desktop clients.
  1. MCP client
    Claude Code or any client that can send a header
  2. Apache, TLS 1.2+
    Only /mcp is reachable
  3. MCP server in Docker
    Bearer auth, rate limits, sessions, Zod validation, approval gate, 3 tools
  4. OpenRouter, Google Sheets, Slack
    The only three external services
Security

Built to be safe by default

Authentication on every request

A bearer token is checked in constant time on initialize, follow-ups, streams and session deletion. The server refuses to start without one.

Human approval gate

Tasks start as pending. Saving needs explicit ids and a confirmation; unapproved tasks can never be saved or announced.

No invented facts

Owners must appear in the transcript, deadlines need a reliable meeting date, and every evidence quote is checked against the text.

Isolated, expiring sessions

Each client gets its own session bound to its token: max 20 at once, 30 minutes idle, 8 hours absolute.

Abuse limits

120 requests/min per IP, 60/min per token, and a 15-minute block after 10 failed logins. Host header and body size are validated.

Hardened container

Runs as uid 10001 with a read-only filesystem, all capabilities dropped, no-new-privileges, resource limits, and loopback-only publishing.

Secrets stay secret

Keys live in the environment or read-only mounts, never in the image or Git. Logs are redacted and a scanner guards the repository.

Safe external writes

Sheet rows are de-duplicated, Slack is notified once per task set, mentions are escaped, and timeouts are reported honestly.

Encrypted in transit

Let's Encrypt certificate, TLS 1.2 and 1.3 only, HTTP redirects to HTTPS, and HSTS enabled.

Demo workflow

A sample run, step by step

Output from a real run on a fictional transcript, with the meeting date supplied as 2026-11-18.

1

Submit the transcript

An excerpt of the fictional meeting:

Dana: Sarah, can you dig into the underperforming campaigns and report back by tomorrow?
Sarah: Yes, I'll investigate the underperforming campaigns and have findings by tomorrow.
John: ... I'll update the financial tracker by Friday.
Dana: I'd like to propose reducing the worst campaign's budget by 10%...
Dana: Let's keep the 10% reduction on the table. Nobody needs to action it yet.
2

AI extracts structured action items

extract_action_items returns proposals only, all pending_approval:

TaskOwnerDeadlinePriorityConfidence
Investigate underperforming Meta Ads campaignsSarah2026-11-19 ("tomorrow")high0.97
Update the financial tracker with the last two weeks of ad spendJohn2026-11-20 ("by Friday")medium0.97

Not extracted, on purpose: the 10% budget reduction was only proposed and never approved, so it is not a task.

3

A human reviews and approves

The person tells the assistant which tasks to approve. Only then does the assistant call:

save_approved_tasks({ extractionId, approvedTaskIds: [ ... ], humanConfirmation: true })
4

Saved to Sheets, announced in Slack

Rows appended in the sheet's own column order:

MeetingTaskOwnerDeadlinePriorityStatus
Weekly Meta Ads reviewInvestigate underperforming Meta Ads campaignsSarah2026-11-19highapproved
Weekly Meta Ads reviewUpdate the financial tracker...John2026-11-20mediumapproved

Then notify_slack posts one message:

2 approved task(s) from Weekly Meta Ads review
Meeting summary: Conversions fell 15% on a flat $8,000 monthly spend...
• Investigate underperforming Meta Ads campaigns, owner: Sarah, due: 2026-11-19, priority: high
• Update the financial tracker..., owner: John, due: 2026-11-20, priority: medium
Tech stack

Built with

Node.js 22TypeScript strictMCP SDK stdio + Streamable HTTP OpenRouter LLM gatewayZod validationGoogle Sheets API Slack WebhooksDockerExpressApache + Let's EncryptVitest 65 tests
Status

Endpoint status

An automatic check runs every few minutes. It confirms the endpoint answers and the tools are discoverable. It does not run the tools.

Live endpoint check

Checking status…

Last check
…
Container
…
TLS certificate
…
MCP initialize
…
Tools discoverable
…

End-to-end tool test

Exercising the tools against live OpenRouter, Google Sheets and Slack is done by a manual test, not continuously. Its most recent result is shown here.

Last passed
…
Most recent run
…
Tools exercised
…

Want to try it?

The endpoint is protected by a bearer token. The connection guide shows how to add it to an MCP client.