- Category
- Developer Tools
- Rank
- No. 819Tools index
Previous survey · No. 814 ·
- Pricing
- Open Source
- Platform
- cli
- Type
- TOOL
- Builder
- huangruiteng
- GitHub
- 5.9k stars
- Latest release
- desktop-main-35521617063-1
- Date
About
LoopX is a lightweight, agent-agnostic state kernel that acts as a local control plane for long-running AI agent work, tracking objectives, gates, todos, evidence, and quota across bounded turns handled by Codex, Claude Code, Cursor, or custom runtimes. It aims to keep multi-day agent workflows reviewable, restartable, and handoff-safe without replacing the underlying agent runtime.
What it does
LoopX gives long-running agent work a persistent operating record outside the chat session. Each turn reads the current objective, permissions, work items, evidence, handoff state, and compute allowance. The agent performs one bounded segment, validates it, then records the result and next action. Human decisions remain explicit gates, while status views and a local dashboard project the underlying state for operators.
Why it's ranked here
LoopX addresses a real failure mode: timers can restart agents, but they cannot preserve authority, evidence quality, ownership, or spending policy. Its strongest trait is disciplined separation between durable state, replaceable executors, provider integrations, and human judgment. The design is unusually explicit about private data and production authority. The cost is substantial operational vocabulary and policy machinery for a tool described as lightweight.
What's good
The package requires only Python’s standard library at runtime. Goals survive thread reloads and executor changes through registries, active state, compact run history, and an event-backed spend ledger. Quota follows health, operator, evidence, and focus gates instead of acting as another permission system. Peer agents use claims, leases, scoped work, and typed continuation rather than depending on a permanent leader.
Tradeoffs
Adoption means maintaining structured goal state, todo classes, gates, evidence, quota, and writeback discipline. The local dashboard is a projection, not an authoritative workspace, and its write paths remain disabled unless explicitly enabled. Probe commands are stored as free-form text, so read-only safety depends on execution policy rather than registration-time validation. The project also states that it is not an autonomous production controller.
How to use it well
Use LoopX for multi-day engineering, research, benchmark, experiment, monitoring, or operations work where agents must resume safely and humans need a legible audit trail. Start with one connected goal, explicit private-data boundaries, a clear write scope, and the smallest validation that proves progress. It complements Codex, Claude Code, Cursor, OpenCode, Pi, shells, and custom runners. It does not supply the underlying agent runtime or authorize publishing, destructive actions, or production writes.
Technical notes+
pyproject.toml defines a Python 3.11+ setuptools package with no required dependencies and four console scripts, led by loopx = "loopx.entrypoint:main". docs/architecture.md describes six durable layers: registry, goal state, run log, run history, status or attention queue, and compute quota. docs/status-data-contract.md specifies a JSON status export plus a loopback HTTP server, with dashboard writes opt-in. docs/quota-allocation.md documents event-derived spend accounting and gate ordering. docs/public-private-boundary.md requires live state, raw evidence, credentials, paths, and traces to remain in ignored local storage.
Observed
- License
- MIT
- Primary language
- Python 3.11 or newer
- Packaging
- setuptools build backend with a Python package named loopx
- Runtime dependencies
- No required dependencies outside the Python standard library
- Install surface
- Hosted shell installer or local checkout installer, exposing commands through a user-local bin directory
- Interfaces
- Command-line interface, JSON status export, and loopback HTTP status server
- Console scripts
- loopx plus Lark and OpenViking extension entry points
- Platform support
- macOS and Linux shells
Read from README.md, pyproject.toml, docs/index.md, docs/README.md, docs/integration.md, docs/architecture.md, docs/quota-allocation.md, docs/status-data-contract.md, docs/public-private-boundary.md, docs/state-interaction-model.md, docs/heartbeat-automation-prompt.md.
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.
