Chronicle

Your coding agents solve problems all day, and then the lesson disappears: Claude Code deletes transcripts after 30 days, and what one agent worked out never reaches the next. Interlatch records what actually happened, every session from every agent, on your machine. It pulls out what was learned and gives it back to you in a dashboard and to all your agents through one MCP server. Each lesson stays latched to its source (the session, project and agent it came from) and earns trust as later work confirms it.

Interlatch was called Chronicle. The chronicle command still works, and an existing install moves over by itself when you update (Moving from Chronicle).

A session on the demo data. Download the one-minute tour (MP4, 6 MB): Home, a session and its transcript, ⌘K search, the glossary Map and a weekly review.

What it gives you

Interlatch reads each finished session and writes down what is worth keeping. A real example, taken word for word from the demo data:

Gotcha · billing-api Stripe webhook signatures need the raw request body Webhook.construct_event verifies the signature over the exact bytes Stripe sent. Parsing to JSON first, or posting json= in tests, fails with SignatureVerificationError. Read await request.body() and pass that. From the session "Stop double charges when Stripe retries invoice.paid", extracted automatically.

• Every session, kept for good. Raw transcripts are archived, so nothing is lost when an agent cleans up.
• Knowledge, extracted automatically. Fixes, gotchas, decisions, commands, project facts and preferences, merged into a knowledge base per project and a playbook across all of them.
• Every lesson keeps its source. Each one links to the session, project and agent it came from, so you can open the transcript behind it, and it gains standing as later sessions confirm it.
• Every agent can ask. Through the MCP server, each agent searches the same memory: what Codex worked out yesterday is there for Claude Code today. "Have we hit this error before?", "why did we put idempotency in Postgres?" Claude Desktop, Cursor, Windsurf and Gemini CLI can connect too.
• A dashboard to browse it all. Sessions with their full transcripts, statistics, a glossary of your own vocabulary drawn as a mindmap, and a weekly review you can take in at a glance. ⌘K jumps anywhere.
• A map of your systems. Every project as a system with its parts (UI, API, database, CI, where it is deployed) and the links between projects, drawn from your manifests and what sessions did, each with the evidence behind it.
• A mindmap for each project group. Put related projects in a group and press Analyze: the analyzer checks the links your glossary lists between the group's terms, keeps the real ones and names each ("uses as broker", "runs on"). It never adds one (Project groups).
• A career path, if you want a game. Turn on game mode and you are the hero of an RPG made from your own sessions: choose a path among 65 roadmap tracks, level up by answering questions on your past lessons, earn ranks at gates with your own work, and paste a job description to see which of its requirements you meet. Off until you turn it on (Innerlatch).
• What your agents made, in one place. The documents, pages, diagrams, decks, pull requests and commits from every session, each linked to the session that made it and marked when the file has changed or is gone.
• On your phone and your other computers. Open the dashboard on your phone through Tailscale, and keep every computer's sessions in one archive, analyzed once ().

Quick start

You need macOS 13 or later and something to analyze sessions: a logged-in Claude Code or Codex, IBM Bob with a Bob API key, or a model provider's API with your own key (Model providers).

1. Install.

   uv tool install --python 3.13 interlatch   # or: pipx install interlatch
   interlatch install
   

   interlatch install finds the coding agents on your Mac, asks which to record, imports their past sessions and asks whether to run Interlatch from login. For the app instead, download it from the latest release (Apple silicon) and choose Connect.

2. Use your agents as usual. Each session is recorded when it ends and analyzed in the background.

3. Explore. Open the dashboard at http://127.0.0.1:11524/ (or interlatch ui --open) and press ⌘K, or ask your agent what it learned last week.

Connect more agents later from Settings › Sources, interlatch connect <agent>, or by re-running interlatch install. Install covers what each step sets up and how to remove it.

Supported agents

Agent	Recorded from	Picked up	Search from the agent (MCP)
Claude Code	~/.claude/projects transcripts	as each session ends, plus every 15 min	✅
Codex	~/.codex/sessions rollouts	every 15 min, once idle	✅
Codex Cloud	tasks at chatgpt.com/codex (via the codex CLI: title, repo, diff)	every 15 min	via Codex
GitHub Copilot	Copilot CLI and agent sessions; Copilot Chat logs in VS Code	every 15 min	✅ VS Code and Copilot CLI
IBM Bob	~/.bob/db/bob.db, read-only	every 15 min	✅
Google Antigravity	~/.gemini/antigravity conversation logs, read-only	every 15 min	✅
OpenCode	~/.local/share/opencode/opencode.db, read-only	every 15 min	✅
claude.ai, ChatGPT	data export: interlatch import <zip>	when you import it	–

All of them share one dashboard, knowledge base, glossary and set of MCP tools; analysis runs through Claude Code or Codex, whichever you choose. Sources has the details for each.

A closer look

Home. Active time, sessions, tokens and estimated cost, day by day.	Map. Your glossary as a mindmap; each term opens into the knowledge and sessions behind it.	⌘K. One search over sessions, knowledge, projects, terms and commands.
Group mindmap. A project group's terms and the links between them, each checked and named by the analyzer.	Career path. In game mode, you are the hero: your level, your path and your rank in each of its modules.	Job goal. A pasted job description as a castle; a tower lights when your rank at a gate meets what it asks.

Screenshots use made-up demo data.

How it works

1. A hook (or the 15-minute sync) hands each finished session to Interlatch, which archives the raw transcript and parses it: prompts, replies, tool calls, files, tokens and cost.
2. Once the session is idle, a condensed digest with secrets redacted goes to whatever you chose for analysis (Claude Code's claude -p, Codex's codex exec, IBM Bob or a model provider's API), which returns a summary and knowledge items. The call runs sandboxed: no tools, hooks or MCP servers.
3. New knowledge is merged into the project's knowledge base, the glossary is refreshed, and each finished week gets a written review.
4. Everything is served to you (dashboard, app, vault, CLI) and to your agents (MCP).

What leaves your machine: only that redacted digest, sent to Anthropic or OpenAI through your own Claude Code or Codex login, to IBM with your Bob API key, or to the model provider you set up with your own key (with Ollama on your computer, not even that). No telemetry, nothing sent to anyone else. Data and privacy lists what is stored where.

What it costs: through Claude Code or Codex, analysis draws from your Claude or ChatGPT plan like any other use of the agent; Bob and model providers bill your own account, and Ollama costs nothing. With Claude, in API terms it averages about $0.38 per session with Sonnet; interlatch analyze --pending --dry-run sizes a backlog before you spend anything. How analysis works has the details.

FAQ

Will it slow down my agent? No. The session-end hook hands off to a detached process and returns in milliseconds; analysis runs later in the background.

Do I need Claude Code? No. Analysis runs through Claude Code, Codex or IBM Bob, or a model provider's API with your own key: Anthropic, Amazon Bedrock, OpenAI, Azure OpenAI, OpenRouter, any OpenAI-compatible server, or Ollama on your own computer. Pick one when you run interlatch install, in Status › Analysis, or with interlatch config set analysis.backend <name> (Model providers). Recording and browsing work either way; with none set up, sessions are archived and wait in the analysis queue.

Windows or Linux? The desktop app is macOS only. On Windows, uv tool install interlatch and interlatch install work as on a Mac, with Task Scheduler running the sync and the dashboard (Windows); a Windows computer joining a team hub is untested yet. On Linux, interlatch install runs the sync and the dashboard as systemd user units, so a Linux box can be the hub for your other computers. A team's hub also runs in Docker.

Can I keep a project or a session out? Add the project to sources.exclude_projects in the configuration, or remove a session for good with interlatch forget <id>.

How do I remove it? interlatch uninstall removes the hooks, background agents and MCP registrations and keeps your data; add --purge to delete the data too.

More questions, on privacy, cost, teams and updating, are answered in the FAQ. Something not working? See Troubleshooting.

Documentation

Install · Sources · Dashboard, glossary and Map · Command line · MCP server · VS Code extension · Phone and other computers · A hub in Docker · Joining your team's hub · A copy in Postgres · What gets recorded and how analysis works · Configuration · Data and privacy · FAQ · Troubleshooting · Moving from Chronicle · Development

Contributing

Issues and pull requests are welcome. CONTRIBUTING.md explains how to run the tests and work on the dashboard with demo data instead of your own sessions.

License

MIT, except the ee/ directory: the features a company needs to run Interlatch across its teams (single sign-on, policies, audit export) are under the Interlatch Enterprise License and need a subscription in production. The interlatch package on PyPI is MIT only.