📒 Agent Ledger

One daily log for all your AI coding agents.

You ask Claude, Codex and Cursor to do the work, and by evening you can't remember what got done. Agent Ledger fixes that: each agent writes a short note when it finishes a task, and one command turns those notes into a clean daily digest.

• ✅ Works with Claude Code, Codex, Cursor, and any agent that reads AGENTS.md
• 📄 Plain Markdown files. No server, no account, no tracking
• 🐍 One Python file, zero dependencies

How it works

1. Agents log. After each task, an agent adds a short entry to its own file for the day (claude.md, codex.md, ...).
2. You compile. Run one command (or ask any agent) and get DIGEST.md: what's done, what's blocked, follow-ups, and the day's git commits.

Pick a mode

Not sure? Choose Global. You can use both at the same time.

Global mode                          Project mode
~/agent-ledger/                      your-project/
  2026-11-05/                          ledger/
    claude.md                            2026-11-05/
    codex.md                               claude.md
    DIGEST.md                              DIGEST.md

Quick start

1. Get it

pipx install ai-agent-ledger      # or: pip install ai-agent-ledger

That gives you the agent-ledger command. Prefer no install? Clone the repo and use python3 Agent-Ledger/ledger.py wherever this README says agent-ledger:

git clone https://github.com/naveedali/Agent-Ledger.git

2. Set it up

agent-ledger init

It asks which mode you want. Or skip the question:

agent-ledger init --global                            # one ledger for everything
agent-ledger init --project path/to/project           # ledger inside a project
agent-ledger init --project path/to/project --private # same, hidden from git

💡 In project mode, point --project at the project your agents work in, not the Agent-Ledger folder. . means the folder you're in right now.

What init does:

• Global: adds the logging rule to your personal agent settings (~/.claude/CLAUDE.md, ~/.codex/AGENTS.md) and adds /digest to Claude Code. For Cursor, paste the rule from ~/agent-ledger/CURSOR_RULE.md into Settings → Rules → User Rules.

• Project: adds the rule to the project's AGENTS.md and CLAUDE.md and creates ledger/. With --private, everything Agent Ledger creates is hidden from git on your machine only. (If AGENTS.md was already in git, the added rule still shows in git diff; use global mode to avoid that.)

• Both: in Claude Code, adds a small log reminder hook. If Claude changed files but didn't log the work, it's asked to log before it finishes. Skip it with --no-hook.

Your existing content is always kept, and running init again is safe.

3. Restart your agents. Agents read their rules when a session starts, so close and reopen any Claude Code, Codex or Cursor session that was already open.

4. Work as usual. Give your agents tasks. They log as they go.

5. See your day

python3 ~/agent-ledger/ledger.py today    # global mode
python3 ledger/ledger.py today            # project mode

Or just ask any agent: "give me today's digest". In Claude Code, type /digest.

What an entry looks like

Agents log with one command, so the date, time and format are always right:

python3 ~/agent-ledger/ledger.py log --agent codex --task "Add onboarding screen with skip flow" \
  --files "OnboardingScreen.kt, NavGraph.kt" --follow-up "Needs analytics event"

That adds this to today's codex.md (the project name comes from the git repo you're in):

## 14:32 · Atfal Compass · DONE
- Task: Add onboarding screen with skip flow
- Files: OnboardingScreen.kt, NavGraph.kt
- Repo: ~/code/atfal-compass
- Follow-up: Needs analytics event

Status is DONE (default), PARTIAL or BLOCKED. The Repo: line is used in global mode so the digest can include that repo's commits. Agents that can't run commands write the same entry by hand.

What the digest looks like

# Daily Digest — 2026-11-05

**3 tasks** · ✅ 1 done · 🟡 1 partial · 🔴 1 blocked · 2 commits
Agents: claude, codex
Projects: Atfal Compass, Client App

## ✅ Done
- **14:32** · Atfal Compass · _codex_ — Add onboarding screen with skip flow

## 🟡 Partial / 🔴 Blocked
- 🟡 **09:10** · Client App · _claude_ — Write README
- 🔴 **16:00** · Client App · _claude_ — Fix payment crash

## 📌 Follow-ups
- [ ] Needs analytics event _(Atfal Compass, codex)_

## 🔀 Commits
- `610cfdf` [atfal-compass] 14:40 feat: onboarding screen
- `a91c2e4` [client-app] 16:20 docs: add README

Use it as an MCP server (optional)

Agents that support MCP can log through a real tool instead of a shell command. It's more reliable, and works even where an agent can't run commands. No extra install: the server is built into ledger.py.

agent-ledger mcp-install

This adds the server to Cursor (~/.cursor/mcp.json) and Codex (~/.codex/config.toml), keeping what's already there, and prints the one command to add it to Claude Code:

claude mcp add --scope user agent-ledger -- python3 "/path/to/ledger.py" mcp

Restart your agents and they get five tools:

So you can also just ask an agent "what did we do this week?" and it answers from the ledger.

Commit safety net (optional)

Turn it on in any repo, and a commit nobody logged still makes it into your ledger:

cd your-project
agent-ledger git-hook install        # agent-ledger git-hook uninstall to remove

After each commit, if no agent logged work for that repo in the last 30 minutes, an entry is added for agent git with the commit message and files. It never blocks or slows a commit, keeps any post-commit hook you already have, and skips merge commits. Set AGENT_LEDGER_COMMIT_WINDOW (minutes) to change the window.

Recover work nobody logged

Claude Code and Codex keep a record of every session on your machine. recover reads those records and finds sessions that changed files but never made it into your ledger:

agent-ledger recover                   # preview today's missing work
agent-ledger recover --date yesterday  # or any day
agent-ledger recover --apply           # add it to your ledger

Each recovered entry uses the agent's own end-of-task summary, the files it changed (or the files in commits made during the session), and a note that it was recovered. Sessions that already logged, read-only sessions, and anything recovered before are skipped, so it's safe to run any time. Nothing leaves your machine.

It reads ~/.codex/sessions and ~/.claude/projects (or CODEX_HOME and CLAUDE_CONFIG_DIR if you set them). Both formats belong to those tools and can change between versions, so treat this as a best-effort backup to logging, not a replacement.

Standup in one command

agent-ledger standup

Standup · Mon 09 Nov 2026

Last working day (Fri 06 Nov)
• Added a Contact & Support section to Settings (project-xyz)
• Migrate auth screens to Compose, in progress (client-app)

Today
• Done: Committed the profile editor with change notes (project-xyz)
• Finish signup validation (client-app)
• Continue: Migrate auth screens to Compose (client-app)

Blockers
• Payment integration (needs: production keys) (client-app)

Plain text, so it pastes cleanly into Slack, Teams or a meeting chat. "Yesterday" is your last day with logged work, so Monday shows Friday. Today collects what's already done today, plus follow-ups and unfinished work carried over. Filter with --project or --agent, add --redact before sharing, and use --date for another day. MCP agents get it as the standup tool.

Reports for any date range

For a weekly review, a client update or an invoice, build a report across several days:

python3 ~/agent-ledger/ledger.py week                                    # last 7 days
python3 ~/agent-ledger/ledger.py report --from 2026-11-01 --to 2026-11-15
python3 ~/agent-ledger/ledger.py report --from 2026-11-01 --project client-app --redact

The report groups work by project and day, with a summary table, follow-ups and commits. It's saved in reports/ inside your ledger (add --print to also show it, or --out file.md to save it elsewhere).

• --project NAME / --agent NAME: only that project or agent
• --redact: hides emails, API keys and tokens, and home-folder paths. Use it before sending a report to anyone.

Clients and time (for freelancers)

Tag work with a client, and get the time spent per client and project, ready for a client update or an invoice.

agent-ledger client client-app "Acme Corp"      # link a project to a client once
agent-ledger report --from 2026-11-01 --client "acme corp" --redact --open

• Time: an entry logged with --time 1h30m (or 45m, 1.5h) uses that. Otherwise the time is estimated from the gap since the same agent's previous entry that day, up to 2 hours, and marked with ~. The first entry of a day gets no estimate.
• Clients: an entry's own --client wins; otherwise the project's linked client is used. agent-ledger client lists the links; client PROJECT --remove drops one.
• Reports show time per task and a Time table (client, project, logged, estimated, total). The HTML view adds a total-time tile.

Estimates are a guide, not a timesheet: when you bill by the hour, log real time with --time.

View it as a web page

Add --html to any digest or report to also get a clean, formatted page, or --open to open it in your browser straight away:

python3 ~/agent-ledger/ledger.py today --open                 # today's digest
python3 ~/agent-ledger/ledger.py week --redact --open         # last 7 days

The page has summary tiles, a status bar, an Activity chart (tasks per day by status, and tasks per agent; hover a bar for the numbers), every project's work grouped by day, follow-ups and commits. You can filter by status or search tasks, files and agents. It's a single file with no internet needed, works in light and dark mode, and prints cleanly, so Print → Save as PDF gives you a report to attach to an email or invoice.

Commands

Updating

init puts a copy of ledger.py in your ledger (~/agent-ledger/ or a project's ledger/), so pulling new code doesn't change those copies. After you upgrade, refresh them in one step:

pipx upgrade ai-agent-ledger                # if you installed with pipx
agent-ledger update                          # updates every copy it knows about
agent-ledger update path/to/project          # or name a project explicitly

Using a clone instead? git pull, then python3 ledger.py update.

update also refreshes the agent rules, the /digest command and the Claude Code hook, and keeps your logs and settings. status tells you when your copy is out of date, and ledger.py --version shows which version you're running.

Auto-compile at the end of the day (optional)

Add a cron job (crontab -e) to build the digest every night at 11 PM:

0 23 * * * python3 ~/agent-ledger/ledger.py compile                      # global
0 23 * * * cd /path/to/project && python3 ledger/ledger.py compile      # project

Troubleshooting: "my agent did work but nothing was logged"

Run this first. It checks the setup and tells you what's wrong:

python3 ledger/ledger.py status            # project mode
python3 ~/agent-ledger/ledger.py status    # global mode

The usual causes:

1. The session was already open. Agents read the rules when a session starts. Restart the agent and try again.
2. That project isn't set up. Project mode covers one project only. Run init --project in each project, or use init --global to cover all of them.
3. The agent skipped it. It happens now and then. Remind it once: "log what you just did in the ledger". The day's git commits still show up in the digest either way. In Claude Code, the log reminder hook catches this for you; status shows whether it's on.

FAQ

It says "Unknown command" for a command in this README. You're running an older copy. Run agent-ledger update (after pipx upgrade ai-agent-ledger), or python3 ledger.py update from your clone after git pull.

Can I change where the global ledger lives? Yes. Set AGENT_LEDGER_HOME (for example to a synced folder) before running init --global.

Can I use another agent, like Gemini or Copilot? Yes. Any agent that reads AGENTS.md follows the same rule and logs to its own file.

Related tools

Agent Ledger is deliberately small. If you need more, these are worth a look: SpecStory (captures full AI chat history) and vibe-log (session analytics and standups).

Contributing

Ideas, issues and PRs are welcome. Keep it simple: one file, no dependencies.

Run the tests before opening a PR (standard library only, nothing to install):

python3 -m unittest discover -s tests -v

License

MIT © 2026 Naveed Ali