I wanted Claude Code but for my Meetings & Calls, so I spent 12.5 months building it 🧑💻
retained text
I'm a software dev by day, and for about a year my nights and weekends have gone into **Samurai Scribe⚔**️. In plain English: it records your meetings (Zoom, Teams, Meet, or in person), transcribes them, figures out who said what, and writes your notes and summaries. Ask it "what did I promise Marta this week?" and it answers with a link to the exact moment. All on your own computer (Mac + Windows), with no meeting bots and no account. For the nerds: it's basically Claude Code + Obsidian + Granola, for your meetings. It started as a little tool for my mom, so it still has to pass the mom test: hit record, get notes 🙇 Claude made me *way* faster at writing code. Everything around the code is still on me: deciding what to build, testing on real Macs and PCs, chasing weird Windows GPU-driver crashes, answering users. That's most of the work, and why this took a year. # The idea: your meetings, as a repo Everything is plain Markdown in a folder you own (open it in Obsidian 💜). The database is just a cache of the files: Vault/ ├── AGENTS.md ├── Meetings/ │ ├── 2026-11-03 Design Review/ │ │ ├── 2026-11-03 Design Review.md # tags, links │ │ ├── Transcript.md # speakers are \[\[links\]\] │ │ ├── Summary.md │ │ ├── Roadmap.pdf # attachments feed the summary │ │ └── \_playback.m4a │ └── 2026-11-02 Customer Call/ ├── People/Marta.md ├── Notebooks/Launch/ # a project across meetings └── Skills/brief/SKILL.md # run it with / * **The assistant works like Claude Code.** It can edit any meeting or note, but edits show up as diffs you keep or discard. Same modes too (Ask, Review, Accept edits, Auto). * **Claude Code plugs in** through the built-in MCP server. # How Claude helps (code is maybe half of it) Daily driver since May: \~1,500 commits (Claude co-authored 1,086) and 72 versions shipped. * **R&D on local models.** I hand-label recordings, and Claude turns them into a test pipeline and dashboard (word error rate, speaker accuracy over time). Models get picked on numbers, not vibes. * **Running the backlog.** 150+ tickets and one of me. A Claude Code plugin files reports as GitHub issues, works tickets in their own worktrees, and at every release drafts a reply to everyone whose bug got fixed. * **Setup:** Pro → Max-level usage (best money I spend on this). Terminal → the desktop app, for parallel sessions and notifications. # What went wrong * **Wrong backend language.** Python first, way too slow. Spent Christmas break porting it all to Rust. * **Wrong storage.** JSON, then SQLite, before I realized the Markdown files should *be* the data. No AI was going to tell me I'd picked the wrong foundation. * **"Fixed" on my branch ≠ fixed for the user.** Early reply drafts told people bugs were fixed when the fix only existed on my dev branch 🤦 Now every claim gets checked against the release they actually have. * **Usage limits are real.** Hit 99% of my weekly limit last week. The culprit wasn't my agents, it was my own giant never-ending chats. Start fresh sessions, people. # Closing stuff Download [https://samuraiscribe.com/](https://samuraiscribe.com/) (Mac + Windows). It runs self-hosted models, and runs completely offline. It uses open-source transcription models (Parakeet, and Whisper) and even open source LLMs to summarize (Qwen, or BYO key) **Free to use forever:** no account, no subscriptoin. Pro is $19 once during the beta ($39 after Nov 16) **Question for y'all:** if Claude Code could read every meeting you've ever had, what would you ask it to do? AMA about the setup and I'll try my best to answer. Cheers! 🍻
document fetched 2026-11-10 00:22 · canonical https://reddit.com/r/ClaudeAI/comments/1xrpzz0/i_wanted_claude_code_but_for_my_meetings_calls_so