Allyn — your accessibility ally in Slack

• 

  allyn-architecture

• 

• 

• 

Inspiration

Slack runs on fast, dense, jargon-heavy text. For a lot of people that's a barrier, not a convenience: non-native speakers, neurodivergent teammates, people with low vision, or anyone new to a team and drowning in acronyms. Existing "summary" bots give everyone the same wall of text. We wanted something that meets each person where they are — and that helps them participate, not just keep up.

What it does

Allyn is a personalised, agentic accessibility layer for Slack.

• /allyn setup captures each person's needs: language, reading level, summary length, jargon handling, and free-text access notes. Everything Allyn shows that person is adapted to it.
• /allyn catchup gives a private, plain-language, translated digest of a channel — the same channel renders differently for each teammate.
• “Simplify with Allyn” (message shortcut) rewrites any single message at your reading level and language, expanding team acronyms it has learned.
• /allyn triage reads a conversation, extracts the action items, and turns each into a task — then /allyn tasks lets you complete them with one tap.
• /allyn remind schedules reminders Allyn DMs you when due.

How we built it

• Slack Bolt (Python) + Socket Mode for a no-hosting, zero-cost runtime.
• A custom MCP server (allyn.mcp_server) exposes Allyn's action tools (create_task, set_reminder, glossary, list/complete) over the Model Context Protocol. The agent connects to it as an MCP client over stdio and routes every action through it — that's the challenge's "MCP server integration", used for real.
• A free-tier LLM brain: Google Gemini primary, Groq automatic fallback.
• SQLite for profiles, tasks, reminders, and the team glossary — inspectable and portable.
• The accessibility logic lives in pure, unit-tested functions (22 tests, no network); MCP, the LLM, and Slack are thin adapters around that core.

Challenges we ran into

• Bridging Slack's synchronous handlers with the async MCP client SDK — solved with a persistent MCP session on a dedicated event-loop thread.
• Reliable action-item extraction on free models — solved by making extraction a deterministic two-step (one structured LLM call → explicit MCP tool calls) with tolerant JSON parsing.

Accomplishments we're proud of

• Personalisation that's genuinely per-user, not one-size-fits-all.
• A real MCP integration, not a checkbox.
• Runs entirely free, locally, with a tested core.

What we learned

Accessibility is most powerful when it's the default and when it's active — helping people act, not just read.

What's next

Voice/audio digests, scheduled morning briefings per profile, more MCP tools (calendar, docs), and publishing the action tools as a reusable MCP server.

Built With

• block-kit
• fastmcp
• google-gemini
• groq
• mcp
• model-context-protocol
• python
• slack-bolt
• socket-mode
• sqlite