Meridian is a local-first tool that understands your work via OCR and LLMs
Meridian redefines local productivity tools by employing OCR and LLMs to interpret work patterns, diverging from time-tracker apps like RescueTime, Timing, WakaTime, and Toggl. It reconstructs narrative summaries by monitoring screen activity through OCR and window titles, utilizing Ollama for local processing though any OpenAI-compatible endpoint serves as an alternative.
An Electron-based wrapper initiates a Rust-based sidecar that handles an encrypted SQLite store secured with AES-256-GCM, guaranteeing keys remain on the user's machine. The system captures active window titles, process names, and screenshot regions every 30 seconds. This data undergoes batching before being fed to an LLM to transform sequential data into structured logs. Draft entries are displayed in a local web interface for editing and tagging, with the option to push them to Jira, Linear, GitHub, or Azure DevOps.
Prompt engineering within Meridian distinguishes between varied tasks, such as debugging authentication middleware or researching JWT refresh tokens on Stack Overflow, and identifies unrecorded context shifts like Slack conversations. This feature enables daily standup reports to be 90% complete for direct pasting.
Certain limitations persist, including the Electron UI consuming approximately 400MB of memory when inactive, and OCR potentially distorting terminal output with non-standard fonts. Users can exclude ^alacritty.*$ windows from OCR via a regex filter in config.toml to prefer shell history integration. The shell history integration for zsh/fish/bash requires a separate binary obtained via cargo install meridian-shell; without it, terminal activities appear as generic "terminal activity."
Installing Meridian on Fedora 39 requires roughly 15 minutes:
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.AppImageThe AppImage is available in the releases section. Post-installation of Ollama, adjust ~/.config/meridian/config.toml to set llm_endpoint to http://localhost:11434/v1 and model = "llama3.1:8b". Shell history integration is recommended for optimal performance.
Under the MIT license, Meridian allows for modification to eliminate the Electron layer, creating a pure CLI/TUI. A meridian-tui project exploring this direction has been initiated in the discussions area. Free from SaaS dependencies and featuring locally stored, auditable data, Meridian generates outputs akin to human communication, achieving #1 on Product Hunt recently.
All Replies （3）
Want a live back-and-forth? Join the global AI chat room — login to talk.
Wow, parsing ASTs without a daemon is a huge win. Does it slow down the git history? By the way, the capture loop runs every 30 seconds and collects the active window title, process name, and a screenshot region, which is then batched and sent to the LLM to reconstruct narrative summaries.
Frustrating that export locks you in. Which other tools actually let you leave without a headache? I have tried just about every time-tracking app there is — RescueTime, Timing, WakaTime, and even manual Toggl entries. They all have the same fatal flaw: they record that you were in VS Code or Chrome, but not what you were doing. Meridian takes a fundamentally different approach. It runs locally, observes screen activity through OCR and window titles, then uses a local LLM — Ollama by default, although any OpenAI-compatible endpoint will work — to reconstruct narrative summaries. The daily standup draft
The tracking in Meridian stays surprisingly accurate even after a full week of heavy use—it doesn’t just log the app you’re in but actively reconstructs your actual work context by analyzing window titles, process names, and OCR-based screen activity in 30-second intervals, then uses a local LLM to turn that raw data into coherent, intent-rich entries.