Founder & CEO at https://teapaw.org/t.co/JZfyrlFmNA | Open Source Real-Time Meeting Bots API
Lisbon, Portugal
Joined March 2010
- Tweets381
- Following20
- Followers155
- Likes239
Pinned Tweet
I flipped the switch: Vexa Open Source APIs - real-time Google Meet transcription is live on @Vultr for builders.
Use Hosted to build features, not manage GPUs. Same Apache-2.0 code as self-host.
Subscribe to support open-source work if Vexa’s useful. Comment for $50 credit.
2
5
386
Sunday 8 Nov, the eve of #WebSummit: AI Tinkerers Lisbon agent-memory demo night at HOOD (Marvila), 18:00. Four live demos, working code, free. With @SageOx @AdrianPatziF @BeneBagus. Every registration is reviewed: lisbon.aitinkerers.org/p/ai-…
![AI Tinkerers Lisbon: Agent Memory Demo Night (Web Summit Eve) [AI Tinkerers - Lisbon]](https://pbs.twimg.com/card_img/2118367037764374528/gDywWw8e?format=jpg&name=420x420_2)
100% with the thesis — keep your meetings on your own machine. Worth adding for teams: meetily runs on one person's laptop and has no API. If you want a bot that joins the call itself and gives your whole team (or your product) speaker-labeled transcripts over a self-hosted API — that's Vexa. Apache-2.0, your models. github.com/Vexa-ai/vexa
This quoted post is unavailable.
1
76
v0.10.6 shipped last week.
Recording assembly moved out of the bot, into a separate service that reads chunks already in MinIO. Bot can OOM, SIGKILL, get evicted. Recording still finalizes.
Failure class → recoverable event.
Same release deleted ~2,300 lines of platform-duplicated audio code. One module now handles Meet, Teams, Zoom. The dedup is what keeps the guarantee from drifting back into per-platform failure within a quarter.
Apache 2.0. Self-host. Audit trail is the commit history.
github.com/Vexa-ai/vexa/rele…
1
2
47
My GitHub account was compromised in the TeamPCP supply chain attack — attackers used ~73 stolen tokens to spam a security issue on @LiteLLM (BerriAI/litellm#24512).
GitHub flagged my account (the victim) and my open-source org has been returning 404 for 2 days.
Vexa (github.com/Vexa-ai/vexa) is an Apache 2.0 project with active users (1.8k stars) who are now blocked.
No response on support ticket #4199834. @github @githubsupport @ishaan_jaff @krrish_dh
Can someone please escalate?
@GitHubSecurity still no response - a popular open source project is blocked fo no good reason, github stays silent
62
this is why we open-sourced our MCP server from day one. Apache 2.0, not "open" until the platform decides otherwise.
we build meeting bot infrastructure — same pattern. competitors ship closed-source MCP servers for their APIs. when the platform changes the rules, their users have zero leverage.
self-host it, fork it, own it. that's the only MCP server nobody can kill.
github.com/Vexa-ai/vexa
Figma shipped a silent patch specifically to kill figma-use — my open-source tool that did what they wouldn't: an MCP server that creates and modifies designs, JSX export, design linting. Then they scrambled to catch up with their own MCP server.
So I spent the weekend recreating @Figma from scratch.
OpenPencil: reads and writes .fig files, AI chat with full design tools, P2P collaboration with zero servers, ~7 MB app. No account, no subscription.
Three days, one developer, MIT license.
openpencil.dev
110
sachin's /save-meeting skill is the one that caught my attention. he downloads meetings from granola into local markdown so claude code can reference what was discussed.
we built vexa to make that pipeline automatic — bot joins the meeting, transcribes with speaker labels, serves it via MCP so any agent can query "what did the team decide about X?"
the 1,500 PMs learning claude code skills will eventually hit the same wall: the agent knows your codebase but not what you decided in yesterday's standup. meeting context is the missing skill category.
Yesterday 1,500 product managers joined my live webinar on Claude Code. I went deep into:
- Why I believe Claude Code is the most productive AI platform for PMs
- Showed off the 13 skills I've built to automate workflows across product strategy, design, and execution
- Walked through exactly how to get started by installing Claude Code as well as picking the right set of editors, terminals, and voice tools
- Give away my detailed playbook to build your own skills to automate your own product workflows
If you missed the session, the video is now up on YouTube.
My favorite comment after the session:
"I just finished digesting what Sachin shared, and… wow. Awe beat angst. The thing that excites me most about AI is its potential to free humans up for higher-order work — and how that extra capacity and access could help level the playing field (if we do this responsibly). Huge kudos to Sachin for making that end goal feel so approachable."
- Mathew Kahansky
youtube.com/watch?v=zsAAaY8a…
1
2
163
we've collected 135 phrases whisper generates when nobody is talking.
it doesn't output silence. it outputs "Thanks for watching!" and "Subtitles by the Amara.org community."
one time it generated "I'm going to be a bad person" on repeat for a full paragraph. from an empty room.
1
71
why: whisper's decoder is a language model trained on 680K hours of youtube. silence → most probable youtube completion → "thanks for watching."
the fix isn't no_speech_threshold (unreliable). it's:
- silero VAD as a pre-gate
- condition_on_previous_text=False
- exact-string blocklist from production data
1
41
Symphony, CrewAI, AutoGen - every agent framework assumes your agent reads code, docs, and tickets.
None of them assume your agent was in the meeting where the decision was made.
That's the gap. Vexa is an open-source meeting bot that joins your calls, transcribes, and serves everything back via MCP. Self-hosted, plugs into any agent.
Your agent doesn't need more tools. It needs ears.
1
1
46
Everyone's excited about Ultrathink making Claude reason harder about code.
But your agent still forgets what was decided in yesterday's standup.
Code memory is solved. Meeting memory isn't.
Vexa is open-source meeting infrastructure - bot joins your call, transcribes, and serves it back via MCP. Your agent queries past meetings the same way it queries your codebase.
1
51
Recall.ai raised $38M in September to build meeting bot infrastructure.
We open-sourced the same thing.
Vexa - open-source meeting bot API:
- Drop a bot into Meet, Teams, or Zoom
- Real-time transcription + speaker diarization
- REST API + native MCP server
- Self-host on your infra or use cloud ($0.45/hr)
1,700+ GitHub stars. Apache 2.0.
github.com/Vexa-ai/vexa
1
56
AI agents are getting memory for code changes, task history, and file edits.
But they still have zero memory of what humans discuss in meetings.
Standup decisions. Architecture reviews. Client feedback. All invisible to your agent.
That's the gap I'm filling with Vexa - open-source meeting bot API with a native MCP server. Your agent can now search and reference actual conversations.
github.com/nicepkg/vexa
1
30
AI agents are getting memory for code changes, task history, and file edits.
But they still have zero memory of what humans discuss in meetings.
Standup decisions. Architecture reviews. Client feedback. All invisible to your agent.
That's the gap I'm filling with Vexa - open-source meeting bot API with an MCP server. Your agent can query "what did we decide about the migration?" and get the actual conversation back.
Task memory + meeting memory = agents that actually know what's going on.
1
20
Claude Code can now talk to you via voice.
But can it remember what happened in your last standup?
I built an open-source MCP server that gives Claude access to your meeting transcripts. Speaker labels, timestamps, searchable across months of meetings.
Your AI coding assistant should know what the team decided yesterday - not just what's in the codebase.
github.com/Vexa-ai/vexa
1
67
Things nobody tells you about building a meeting bot:
- Google Meet, Teams, and Zoom all have different join flows. They change without warning.
- Your bot gets kicked for "suspicious activity"
- Multi-speaker diarization on compressed WebRTC audio is still unsolved cleanly
- Real-time transcription is a completely different architecture than batch
Two years of this. Still shipping.
51
1,700+ GitHub stars on Vexa.
That's more than all Recall.ai repos combined — and they raised $38M.
Open source compounds differently. Every star, every fork, every self-hosted deployment is a moat a VC-funded competitor can't replicate.
github.com/Vexa-ai/vexa
35
Most meeting bots just sit there silently recording.
Vexa bots can actually speak in meetings via TTS. They can share their screen. They can listen and respond in real-time.
Building voice AI agents that participate in meetings, not just transcribe them. Open source.
github.com/Vexa-ai/vexa
1
32
Replaced $200/yr in meeting transcription subscriptions with a free n8n pipeline:
1. Vexa bot joins call, transcribes
2. Webhook fires when meeting ends
3. n8n sends transcript to LLM for summary
4. Summary drops in Slack
Took 3 hours to set up. Runs for $0.
github.com/Vexa-ai/vexa
2
49
"Where does our meeting data go?"
If your security team can't answer this, you have a problem.
I built Vexa so that the answer is always "our own servers." Docker Compose, ~4GB RAM, no GPU needed. Nothing leaves your network.
The only meeting bot API you can fully self-host.
16
Meeting transcription APIs in 2026:
Recall.ai — closed source, cloud only, $38M VC
Fireflies — closed source, cloud only
Otter.ai — closed source, cloud only, facing class action
Vexa — open source, self-hostable, Apache 2.0
When your meetings contain strategy and financials, "trust us" isn't enough.
github.com/Vexa-ai/vexa
29