AgentArchMate

• 

  Happy to discuss SpaceX

Elevator pitch: Meeting notes, fully local. A drop-in skill for WorkBuddy and any agentic tool — no cloud, no tokens, no leaks.

Inspiration

Every AI notetaker on the market works the same way: your meeting audio leaves your machine, gets processed on someone else's servers, and you pay per token for the privilege. For teams in public sector, financial services, or healthcare, that's a non-starter — frameworks like PDPA and internal security policies often prohibit sending meeting content to external clouds. We kept seeing the same pattern: organizations either ban AI notetakers outright or people quietly break policy to use them. Meanwhile, the agentic ecosystem is exploding with tools like WorkBuddy that can consume skills built by others. The question wrote itself: what if meeting notes were a fully local capability that any agent could simply call?

What it does

AgentArchMate is a meeting-notes mate that runs entirely on your machine. Feed it a recording or a raw transcript and it produces structured meeting notes — a concise summary, decisions made, action items with owners and deadlines, and open questions — as clean markdown for humans and JSON for machines. It's packaged as a portable skill, so WorkBuddy and other agentic tools can invoke it inside larger workflows ("summarize today's meetings, then draft the follow-up emails"). Nothing ever leaves the device: no external API calls, no cloud inference, no token bills. Privacy isn't a policy promise here — it's the architecture.

How we built it

The pipeline is local end to end. Speech-to-text runs on-device with whisper.cpp for fast, accurate transcription. The language layer is a quantized open-weight model served through Ollama, handling summarization and structured extraction against a strict JSON schema. Long meetings are chunked with overlapping windows, extracted per chunk, then merged in a consolidation pass so decisions and action items don't fall through the cracks. The whole pipeline is wrapped in a skill manifest that declares its inputs, outputs, and invocation contract, so agent frameworks can discover and call it like any other tool — WorkBuddy first, with the interface kept framework-agnostic by design.

Challenges we ran into

Small local models are humbling. A 7–8B model will happily invent an action item if you let it, so we built guardrails: strict output schemas, extraction grounded in transcript evidence, and a verification pass that rejects any item without supporting text. Speaker attribution without cloud diarization APIs was another fight. Making one skill definition portable across different agent hosts meant designing for a common contract without dumbing down the interface. And keeping everything responsive on ordinary laptop hardware — no GPU cluster — forced constant trade-offs between model size, quantization, and output quality.

Accomplishments that we're proud of

It works with the wifi off. That single demo — airplane mode on, meeting in, structured notes out — proves everything we set out to build. We're proud that the skill is genuinely portable rather than welded to one host tool, that processing costs exactly zero tokens no matter how many meetings you run, and that the structured output is reliable enough for other agents to build on without human cleanup in between.

What we learned

Scaffolding beats scale. With the right chunking, schemas, and verification passes, a small local model can deliver notes good enough to act on — at zero marginal cost. We also learned that the skill is the right unit of value in the agentic era: building one composable capability that many agents can call creates more leverage than shipping yet another standalone app. And privacy-by-architecture resonates in a way policies never do — people trust "it physically cannot leave your machine" far more than any terms-of-service paragraph.

What's next for AgentArchMate

Real-time mode, so notes materialize while the meeting is still running. Multilingual support is close behind — starting with English, Mandarin, and Cantonese, including the code-switching that's normal in Asian workplaces. On the integration side, we're expanding the skill packaging so AgentArchMate is a drop-in for more agent frameworks, and adding local-only downstream actions like task creation and follow-up drafting. The long-term goal is simple: take it from hackathon demo to production grade — hardened, auditable, and ready for the regulated industries that need local-first AI the most.

Built With

• agent
• agentarch
• agentarchmate
• local
• localmeetingnote
• meeting
• meetingnote
• zerotoken