A local-first journal for you and your agents

Your agents keep the record. You keep the meaning.

Lantai reads the sessions your agents leave on your machine and drafts daily, weekly and monthly reports, every point linked to its conversation. Beside them, you write what no session holds.

This is a live demo with made-up data. Switch views, click an anchor like [#S5.T2], reassign a session and press Update, or show only what you wrote.

Open sourcemacOS desktop app & CLIBeta

The problem

A day's work is scattered across a dozen sessions. Some of it is in none of them.

What changed, why it changed, where you got stuck: it's all in the conversations, and almost none of it is in your commits. By Friday it's gone.

And no session saw the talk with a colleague, the test on a real phone, or whether it all worked out. Agents can produce excellent results for us. They can't live the day for us.

Two hands, one journal

The agent drafts. You write.

Both hands share one page, one outline per project, and they stay distinct. Purple is yours, and regenerating a report never touches it.

The agent writes

  • What happened in the sessions: what was built, what changed, where it got stuck
  • Every point anchored to the turn it came from, checked by the program
  • A fresh draft whenever you regenerate or update

You write

  • What happened outside them, and whether it worked
  • Notes on the day and on projects, and comments on the agent's points
  • Background and plans on each project page
How it works

Gather, draft, write.

Writing the journal is meant to be a short evening ritual, not a one-click black box.

1

Gather

Lantai reads each agent's session files incrementally and splits long sessions by day. You put them into projects you created and ignore the noise; rules and the model suggest, you confirm.

2

Draft

Each session is condensed into a cached digest, and the model drafts the report by project: key points first, details and anchors below. Weekly and monthly reports build on the daily ones.

3

Write

Add what only you know, right beside the draft: notes on the day and on projects, comments on the agent's points. You can write before anything is generated, too.

Features

Built to be trusted, not just to be fast.

A report is only useful if you can check it. Lantai keeps a clear line from every sentence back to the work behind it.

Two hands, one journal

The agent's draft and your notes share one page. Purple marks what you wrote; switch between both, the agent's part, or only yours.

You sort, the model drafts

Projects are yours to define. Directory and remote rules or the model suggest where sessions belong, but nothing is filed until you accept.

Every point is traceable

Conclusions carry anchors like [#S12.T5] that open the exact turn they came from. Every anchor is validated, and any the model made up is dropped.

Your words survive rewrites

Update rewrites only the projects that changed. Your notes are never touched, and your edits to the agent's text stay put and stay marked.

Daily, weekly, monthly

Weekly reports cluster each project's week into a few threads of work; monthly reports build on them. The project view rolls any project up, notes included.

A dashboard of your work

Active hours with parallel sessions counted once, streaks, the tools your agents reach for, your hourly rhythm, and token usage by model.

Reads sessions from
Claude Code Codex CLI Antigravity DeepSeek Harness WorkBuddy OpenCode
Privacy

Your sessions stay on your machine.

Agent logs hold private code, file paths and sometimes secrets. Lantai treats them that way.

Read-only, always

Lantai never writes, moves or deletes anything in your agents' data folders. Where those folders also hold credentials, it reads only the session files and nothing else.

Redacted twice

Keys, tokens, private-key blocks and .env-style values are scrubbed before anything reaches the local database, and again before any text is sent to a model.

Keys in the Keychain

API keys live in the macOS Keychain, never in config files, the database or logs. Git remotes are stored as host/owner/repo, so tokens embedded in URLs are never kept.

Sealed generation

The model writes reports with all tools disabled and nothing saved to disk. Instructions hidden inside a session are treated as plain text, never as commands to follow.

Your model, your choice

Use any OpenAI-compatible endpoint, including local models through Ollama or LM Studio, so nothing has to leave your computer. Or reuse your existing Claude Code login.

No account, no telemetry

No sign-up, no cloud sync, no analytics. Everything lives in a local SQLite database, and the session data in it can be rebuilt from your agents' logs at any time.

Get started

Download, scan, write.

A free macOS app for Apple silicon and Intel, with the command-line tool inside. Lantai is in beta: until 1.0, an update may change the data format, and the changelog says what to do.

Download for macOS

Desktop app

Open the .dmg and drag Lantai into Applications. It isn't signed with an Apple Developer ID yet, so the first time, click Open Anyway in System Settings → Privacy & Security, or run:

xattr -dr com.apple.quarantine \ /Applications/Lantai.app

Then add a model in Settings → Models (Find tools picks up Claude Code), Scan, sort sessions into projects and Generate.

Command line

The same engine as a command-line tool, aj. Install it from Settings → System → Command-line tool. It links ~/.local/bin/aj, so that folder needs to be on your PATH:

export PATH="$HOME/.local/bin:$PATH" aj scan aj report daily # add the first line to ~/.zshrc to keep it

The first scan reads the last 90 days of sessions. Run aj --help for every command.

PlatformmacOS 11 or later, on Apple silicon or Intel.
A modelAny OpenAI-compatible API, a local model server, or a signed-in Claude Code.
From sourceRust (stable) and Node.js, then npm run build:app in app/. See the README.
Why I built it

Agents can't live the day for us

Much of our work now goes to agents: building products, chasing down problems, drafting emails, running experiments, writing docs. We explain the problem, and an agent gets it done. What we did outside those sessions, whether things really worked out, what we made of it all: only we know.

A record like this may be of little use to the agents, and of little use to the projects. But to us, as people, it matters.

Looking back over the days, weeks and months, you see how much you got done working alongside agents, including things you never thought you could do. That meaning is yours.

Lantai is also an experiment in how people and agents work together, and how they keep a record together. If you have thoughts on where the line should be, open an issue.

The name

Why “Lantai”?

Lantai (兰台) was where the Han dynasty court kept its books and archives. Its keepers collated texts and drafted documents; the historian Ban Gu worked there while compiling the history of his age. Over time the name came to stand for the historian's office itself.

Lan (兰) is the orchid, long a symbol in Chinese culture of quiet, understated character: not showy, but with a fragrance of its own.

Gathering each day's scattered work, putting it in order and filing it away is exactly what this app sets out to do.