@superdesign/loopbase
v0.2.0
Published
Cross-session agent memory — discover, read, search, and hand off work across local AI agent sessions (Claude Code, Codex, pi).
Readme
loopbase
Cross-session agent memory. Discover, read, and hand off work across local AI
agent sessions — Claude Code, Codex, and pi — on one machine.
agent-facing usage in SKILL.md.
What it does
Every agent dumps its session transcript to disk. loopbase indexes all of them
(lazily, no daemon) and gives one CLI to:
**loopbase list**— see the sessions in your project, across agents, each with itsworklog as a table of contents.
**loopbase search "<text>"**— find sessions/turns by content across agents; eachhit carries a handle to jump straight to it.
**loopbase show <id>**— understand a session: a structural map by default(worklog + turn outline), then drill into a turn, a worklog span, a tool
result, or a subagent.
**loopbase log "<what I did>"**— leave a worklog entry on your current session;it auto-spans the messages since your last log.
**loopbase cost**— token + USD cost per session (also--summaryby model,or
loopbase cost <id>for one session's per-model breakdown).**loopbase insights**— ranked automation candidates mined from everysession's tool calls: repeated/expensive patterns, recurring call sequences, and tools that keep failing. Ranked by real attributed USD, tagged with the dominant repo, and nested with sub-clusters (
composio run→ the Intercom tools it's really running). Each row links toshow --turnexamples.**loopbase serve**— a local web dashboard over everything above (see below).
Output is JSON by default (--text for humans). Session id is auto-resolved from
the environment — agents pass nothing.
Web dashboard (loopbase serve)
loopbase serve # → http://localhost:4178
loopbase serve --port 8080 # pick another portA local, read-only dashboard over your indexed sessions — no account, nothing leaves your machine, and it re-indexes on each load so it's always fresh. What you get:
- Sessions by cost — every session with its token + USD spend; sortable, filterable by agent and time window, with resizable columns.
- Group by working dir — fold sessions under their project (git root), each with a cost subtotal; the git branch is shown per session.
- Drill into a session — per-model cost breakdown with a token-burn sparkline, cost per log batch (when you've logged more than once), the worklog, and a paged conversation viewer loaded on demand.
- Insights tab — automation candidates across all sessions: the most repeated/expensive tool patterns, recurring call sequences, and recurring errors, each with example sessions to open.
Cost is a list-price estimate computed from a built-in price catalog
(subscription billing differs). Refresh rates from upstream any time with
loopbase cost --refresh.
Install
loopbase runs on Bun ≥ 1.3 (it uses bun:sqlite).
bun add -g @superdesign/loopbase # installs `loopbase` (+ `lb` alias) globally
bunx @superdesign/loopbase list # or run without installingPrefer a standalone binary (no runtime needed)? See INSTALL.md.
Develop
bun install
bun run loopbase -- list # run the CLI from source
bun run serve:dev # dashboard with hot reload (edits in src/web/ live-reload)
bun run check # lint + typecheck + tests
bun run build # compile a single binary → dist/loopbaseRequires Bun ≥ 1.3.
Status
v0 (the JSONL family): Claude (read + write), Codex (read), pi (read). SQLite
(hermes) and openclaw adapters, hooks, LLM summaries, and team sync are planned
License
Mostly AGPLv3; files marked /* @license Enterprise */ fall under the
loopbase Enterprise Commercial License. See LICENSE.
