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.
I'm interested in the intent classification breakdown shown in the dashboard screenshot. That's cool, I'd love to know what's the breakdown b/w various types of tasks I accomplish with Claude code. Pretty cool!
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
Hey HN! OP here.
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.
Repo: https://github.com/manthan787/context-ledger
I'm interested in the intent classification breakdown shown in the dashboard screenshot. That's cool, I'd love to know what's the breakdown b/w various types of tasks I accomplish with Claude code. Pretty cool!
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