thatyou were in VS Code or Chrome, but not
whatyou were actually doing. Meridian takes a fundamentally different approach. It runs locally, watches your screen activity via OCR and window titles, then uses a local LLM (Ollama by default, though you can point it at any OpenAI-compatible endpoint) to reconstruct narrative summaries.
The architecture is surprisingly clean. A lightweight Electron wrapper spawns a Rust sidecar that handles the encrypted SQLite store — AES-256-GCM, keys never leave the machine. The capture loop runs every 30 seconds by default, pulling active window title, process name, and a screenshot region. That raw stream gets batched and fed to the LLM with a prompt that essentially says "turn this chronological noise into a coherent work log." The output lands as draft entries in a local web UI where you can edit, tag, and push to Jira, Linear, GitHub, or Azure DevOps.
What impressed me most during a week of daily driving it: the prompt engineering is good. It correctly distinguished between "debugging the auth middleware" and "reading Stack Overflow about JWT refresh tokens" — same files, totally different intent. It also caught context switches I'd forgotten: that 20-minute Slack thread about database indexing that turned into a real schema change. The daily standup draft it generated was 90% ready to paste.
A few rough edges worth knowing. The Electron UI feels a bit heavy on memory (~400MB idle). The OCR occasionally garbles terminal output with weird fonts — I ended up adding a regex filter in config.toml
to skip windows matching ^alacritty.*$
and just rely on the shell history integration instead. Speaking of which, the shell history hook (zsh/fish/bash) is a separate binary you install via cargo install meridian-shell
— not documented prominently enough. Without it, terminal work shows up as generic "terminal activity."
Setup took me about 15 minutes on Fedora 39:
wget https://github.com/meridian-dev/meridian/releases/download/v0.4.2/meridian-0.4.2.AppImage
chmod +x meridian-0.4.2.AppImage
curl -fsSL https://ollama.com/install.sh | sh
ollama pull llama3.1:8b
./meridian-0.4.2.AppImage --init
cargo install meridian-shell
meridian-shell install --shell zsh
./meridian-0.4.2.AppImage
The MIT license means you can fork and strip the Electron layer entirely if you want a pure CLI/TUI — someone's already started a meridian-tui
experiment in the discussions tab.
For teams drowning in status-update theater, this hits a sweet spot: zero SaaS dependency, auditable local data, and output that actually resembles human communication. The fact that it's #1 on Product Hunt today suggests I'm not the only one tired of performing productivity for Jira.
Agile development is basically just a fragmented waterfall now 4d ago
Stop tracking "AI adoption rates" because they are essentially 10d ago
Next Huzzah flips the agent loop: you write pseudocode →