agentgraphed

What it does

AgentGraphed automatically indexes every Claude Code and Codex session on your machine. With it you can:

• Find old work you forgot about — every session, searchable.
• Resume abandoned sessions with one click.
• Browse project history across every git repo you've touched.
• Generate context for a new chat from any past conversation.
• Track usage and costs across providers, models, and projects.

Everything stays local. AgentGraphed reads the JSONL logs your CLI tools were already writing (~/.claude/projects/, ~/.codex/sessions/) and turns them into a real dashboard.

Try it

npx agentgraphed

That's the whole install. No clone, no signup, no config file. The dashboard opens at http://localhost:3737. Re-run any time, or leave it running — it re-scans every 5 minutes.

Prefer a global install?

npm install -g agentgraphed
agentgraphed

Don't have npx / Node? Install Node 20+ from https://nodejs.org (the LTS version is fine).

What to expect on first run

1. npx agentgraphed downloads the package once (~15 MB, takes 30-60s on a typical connection). Subsequent runs start in seconds — npm caches the package.
2. You'll see one npm warn exec line and a couple of npm warn deprecated warnings from transitive dependencies. Harmless.
3. The dashboard server boots: › Starting AgentGraphed on http://localhost:3737 › Scanning local AI coding sessions… Found 142 Claude + 8 Codex sessions (14821 messages indexed in 1943ms) › Ready. Press Ctrl+C to stop.
4. Your browser opens to the dashboard automatically.
5. To stop the server, hit Ctrl+C. Your indexed data stays at ~/.agentgraphed/agentgraphed.sqlite for next time — even after Claude Code rotates the original log files off disk.

Troubleshooting: native better-sqlite3 errors

AgentGraphed uses better-sqlite3, a native module that compiles (or downloads a prebuilt binary) during install via an npm install script. If your npm is configured with ignore-scripts=true (common in hardened or corporate setups), that step is skipped and you'll hit errors on launch like:

Error: Could not locate the bindings file

or a NODE_MODULE_VERSION mismatch. The fix is to (re)install while allowing install scripts to run:

npm install -g agentgraphed --foreground-scripts --ignore-scripts=false
# …or, if you run it with npx:
npm_config_ignore_scripts=false npx agentgraphed

If it still fails, you likely need build tools for the native compile: Python 3 and a C++ toolchain (build-essential on Debian/Ubuntu, Xcode Command Line Tools on macOS, or the "Desktop development with C++" workload on Windows).

Screenshots

Find old work — every session, grouped by day. Search and filter by project, provider, or model.	Resume unfinished sessions — read the full conversation, copy a one-line resume command, or generate a primer to paste into a fresh chat.
Browse project history — every git repo you've worked in, ranked by activity. Click any project to see only its sessions.	See where your AI time goes — usage by day, model, project, and category, with retail-priced cost estimates from LiteLLM's 2700+ model catalog.

Features

• Timeline — every session grouped by day, with STARTED · SPANS Nd / CONTINUED / CLOSED badges so multi-day work doesn't hide on a single bucket.
• Sessions — read past conversations in a chat-bubble view; search by content, project, provider, or model.
• Resume — one click copies cd <cwd> && claude --resume <id> to your clipboard.
• Generate context (optional, BYO LLM key) — produces a primer to paste into a fresh chat so you don't lose context when resuming.
• Projects — auto-detected from git repo roots; per-project usage, model spend, and session history.
• Dashboard — usage chart, KPIs, top projects, work-type breakdown, per-model cost. Filter the whole dashboard by project or model family.
• Auto-classification (optional, BYO LLM key, on by default) — categorizes sessions as Feature / Debugging / Planning / Refactor / Styling / DevOps / Data / Payments / Docs / Content and writes a clean past-tense title for each. Typically a few cents for hundreds of sessions.
• Live quota probe (optional) — hover the sidebar widget to read your live Anthropic 5h/7d and OpenAI per-minute rate-limit utilization. Single-token probes (~$0.00006 each).
• Share — generate a stat-card PNG of your dashboard, project, or session view and copy it straight to your clipboard.
• Background ingest — re-scans local logs every 5 minutes; new sessions appear without you doing anything.
• Cost estimates — LiteLLM's auto-updating retail pricing for 2700+ models. Treat as directional.
• Range picker — 7d / 30d / 90d / all-time on every chart, with auto-suggested log scale when the data is long-tailed.

Privacy

AgentGraphed is local-first by default. Everything lives in ~/.agentgraphed/agentgraphed.sqlite on your machine.

There are exactly three things that can ever leave your computer, and every one of them is opt-in:

1. LLM session classification + "Summarize for new chat" — only fires when you click the button or enable auto-classify. Sends a sampled set of your own prompts to your configured LLM provider with your API key. Nothing reaches AgentGraphed servers.

2. The opt-in leaderboard. Off by default. When you turn it on, every six hours (or sooner if you finished a new session since the last submit) your local app posts a batch of session-level rows to https://agentgraphed.com/api/leaderboard/submit. Each row contains: the handle you picked, a random per-session UUID, start time, duration, provider, model, tokens by kind (input/output/cache_read/cache_write), est. cost, and message count. No prompts, no project names, no session content, no file paths, no git branches, no API keys. Full breakdown of what is and isn't sent: https://agentgraphed.com/privacy.

3. API keys are stored in plaintext in your local SQLite file — same threat model as ~/.aws/credentials or a .env. Don't commit your home folder to git.

To verify any of this, read the source. The leaderboard submitter is one function (maybeSubmitLeaderboard) in src/lib/ingest/auto.ts. The wire format is one schema — what's not in the schema isn't sent.

To audit or delete your leaderboard data:

# See everything we have for your handle
curl "https://agentgraphed.com/api/leaderboard/my-data?handle=YOUR_HANDLE"

# Delete it all
curl -X DELETE "https://agentgraphed.com/api/leaderboard/my-data?handle=YOUR_HANDLE"

Supported sources

• Claude Code — reads ~/.claude/projects/*/*.jsonl
• Codex CLI — reads ~/.codex/sessions/YYYY/MM/DD/*.jsonl

More adapters coming. If you want a specific one, open an issue.

Optional: LLM session titles & categories

Without an API key, sessions show the first line of the first prompt as the title. With a key:

• Anthropic — Haiku 4.5 (default) gives clean past-tense titles like "Fixed Stripe checkout double-decimal bug"
• OpenAI — GPT-5 mini works equally well, often cheaper

Click Settings → LLM provider, paste your key, then Classify uncategorized. Cost is fractions of a cent per session — typically $0.01–0.03 for a few hundred sessions.

Configuration

Environment variables:

Log directories can also be edited at Settings → Data sources without restarting.

Run as a service

To keep AgentGraphed running in the background so the dashboard is always at http://localhost:3737, run it with --no-open (or AGENTGRAPHED_NO_OPEN=1) under your OS service manager. The flag skips the browser launch, which would otherwise pop a window on every (re)start.

macOS (launchd)

Save as ~/Library/LaunchAgents/com.agentgraphed.plist (replace YOU with your username), then launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/com.agentgraphed.plist:

<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
<plist version="1.0">
<dict>
    <key>Label</key>
    <string>com.agentgraphed</string>
    <key>ProgramArguments</key>
    <array>
        <string>/usr/bin/env</string>
        <string>npx</string>
        <string>agentgraphed</string>
        <string>--no-open</string>
    </array>
    <key>EnvironmentVariables</key>
    <dict>
        <key>PATH</key>
        <string>/opt/homebrew/bin:/usr/local/bin:/usr/bin:/bin</string>
        <key>HOME</key>
        <string>/Users/YOU</string>
    </dict>
    <key>RunAtLoad</key><true/>
    <key>KeepAlive</key><true/>
    <key>StandardOutPath</key><string>/Users/YOU/Library/Logs/agentgraphed.log</string>
    <key>StandardErrorPath</key><string>/Users/YOU/Library/Logs/agentgraphed.log</string>
</dict>
</plist>

Linux (systemd user service)

Save as ~/.config/systemd/user/agentgraphed.service, then systemctl --user enable --now agentgraphed (and loginctl enable-linger $USER to keep it running after logout):

[Unit]
Description=AgentGraphed
After=network-online.target

[Service]
ExecStart=/usr/bin/env npx agentgraphed --no-open
Restart=always
RestartSec=5

[Install]
WantedBy=default.target

Node version note: the bundled better-sqlite3 ships prebuilt binaries for current LTS releases. If npx agentgraphed fails to install on a very new Node with a node-gyp/compile error, install and run under the latest LTS (e.g. Node 22).

Development

git clone https://github.com/sudomichael/agentgraphed.git
cd agentgraphed
npm install
npm run dev

Open http://localhost:3737. Hot reload, source under src/.

To produce a publishable build:

npm run build         # builds Next.js standalone bundle
npm pack              # creates agentgraphed-X.Y.Z.tgz

License

MIT © Michael Patrick

Built and maintained by @ushercakes.