Thanks, glad that caught your eye! The summarizer tags each session with a hierarchical intent label (e.g. `coding/frontend`, `research/tech-qna`, `deploy`) and a confidence score. You can run it fully local with Ollama so nothing leaves your machine. Would love to hear any feedback when you try it out
I built ContextLedger because I kept losing context between AI coding sessions. I'd wrap up a session with Claude or Codex, and the next day I'd spend the first 10 minutes re-explaining what I was working on, what failed, and what was left to do.
ContextLedger is a local-first CLI that captures events from your AI coding sessions (prompts, tool use, errors, files touched), summarizes them into structured "capsules," and generates resume packs / handoff documents so your next session starts with the right context already loaded.
Key design choices:
Local-first, SQLite-backed – everything stays on your machine. No cloud, no accounts.
Multi-agent – works with Claude Code (via hook events), Codex (rollout JSONL ingestion), and Gemini (coming soon!).
Privacy controls – prompt capture is off by default, secrets are redacted, and you can block prompt content from being sent to remote summarizers.
Summarizer-agnostic – supports local models via Ollama, or OpenAI/Anthropic APIs for generating session capsules and intent labels.
Handoff workflow – ctx-ledger handoff --agent claude --from latest generates context and can launch the agent directly with it injected.
It also ships with a web dashboard (ctx-ledger dashboard) that gives you a live view of your active agents and detailed analysis of your sessions — planning vs. execution time breakdowns, project-level time tracking, intent distribution, and tool usage stats. It auto-refreshes every few seconds and keeps syncing enabled integrations in the background, so it doubles as a monitoring panel while you work.
The onboard command gets you set up in one step – it enables integrations, runs an initial sync, and starts the dashboard.
I've been using it daily for last few days and it's made a noticeable difference in how quickly I can pick up where I left off, especially on multi-day projects. The intent labeling and planning-vs-execution time breakdowns have also been interesting for understanding how I actually use these tools.
Would love feedback.
Great product! Love the design. Minor feedback: "create your first graph" button seems to be broken for me. Thanks for making it really easy to try it out!
Interesting perspective. After reading this article, if you want to use FP in your python projects, definitely check out the toolz library. It's got great API and really good documentation.
I am not really a hardcore fan of Julian. I found out about his encounter with Schmidt only a few weeks back. and the views he expressed in the article manifested when the whole firing thing happened. So I just wanted to share that.
Oh and I've been a long time Hacker News reader, but never posted comments. So I made a new account. There's nothing else to it.
Yes, I was amazed too. He also published the whole 2 hours+ long audio recording of the whole thing. So you can actually listen to the whole thing, the things he is referring to in the article.