SuperAPI AI Directory
Partner with us中文AI newsFavorites
All toolsAI codingLians

Lians

Lians is an open-source "completion guard" tool for AI coding agents, supporting Claude Code, Codex, Cursor, and others. It can recover interrupted agent sessions, reject stale task states, and block "completed" conclusions before a task is truly ready for human review. Its core is binding the results of checks that actually ran to the current Git state, distinguishing "measured evidence" from "agent self-claims"; only the three evidence types measured_local, measured_ci, and human_confirmed can pass the review gate. Lians also offers an optional local control layer, Mission Control (Beta), which lets a mission follow a bounded Codex route and requires a real verifier to pass. The free recovery path requires no account or API Key, data is stored by default in ~/.lians/mcp.db, and it is licensed under Apache 2.0.

Open website

About this tool

Lians is a "current state and completion guard" for AI coding agents, aimed at tools such as Claude Code, Codex, and Cursor. The core problem it solves is: an agent can forget chat context, but it should not forget what has already been completed, what was changed, and what must still pass.

Purpose

  • Recover interrupted agent work.
  • Reject stale task states, preventing old evidence from being silently reused.
  • Block "completion" conclusions before a task is truly ready for human review.

Main capabilities

  1. Recover: recover a bounded current task across supported Claude Code and Codex sessions.
  2. Reject stale state: bind checkpoints to the current repository and task state so old evidence is not reused as current evidence.
  3. Guard completion: separate measured evidence from the agent's own claims; keep the gate closed when work is missing, unknown, failed, or blocked.
  4. Require review: READY FOR HUMAN REVIEW is a signal for handoff to a human, and never means the work is correct, approved, or safe to deploy.
  5. Stay local: the free recovery path does not require a Lians account, AI password, or provider API Key.

Trust model

  • measured_local, measured_ci, and human_confirmed evidence can satisfy a criterion.
  • agent_attested and inferred_activity records can serve as context, but cannot open the review gate.
  • The agent cannot promote its own checkpoints to a trusted category.
  • Trusted CI evidence requires exact GitHub proof and commit matching, as well as interactive check-to-criterion mapping; human evidence requires interactive confirmation.

Two-layer structure

  • The open-source Lians Guard layer: exists within Codex through MCP, skills, hooks, and repository state, recording corrections and exposing evidence, but does not take over the model.
  • The optional Lians Mission Control beta: a local control layer; give it a mission, let it choose a bounded Codex route, and require a real verifier to pass before returning PASS. On supported Codex versions, a Mission Control mission uses a Codex app-server thread instead of starting a new process and replaying summaries at each stage. Codex still handles authentication, tools, sandboxing, and execution; Lians handles the persistent mission, usage policy, and receipts. App-server is currently experimental, so the single-shot adapter remains available.

How to use

  • Choose the AI tool you already use, for example install uv, connect Codex, restart Codex, then save a safe project fact and restore it in a new session.
  • Local memory is stored by default at ~/.lians/mcp.db. This is a usable free recovery path; the full Guard workflow is currently a developer preview.
  • The full quick start covers installation, recovery, correction, deletion, and Guard preview boundaries.
  • The local Python SDK can be used in-app; the installation guide covers TypeScript, Go, Java, C, framework integrations, and self-hosting.

When to use it

  • Continuity recovery for AI coding agents across sessions and across tools.
  • Evidence-backed verification of task completion and handoff to human review.
  • Teams with governance needs using advanced capabilities such as project-level agent handoff, signed selections and review receipts, local research and browser summaries, temporal reconstruction, lineage, information barriers, confirmed deletion, and bounded formal checks; these are not required for the introductory memory workflow.
  • A student and community toolkit for classrooms, clubs, hackathons, or campus developer groups.

Limitations and boundaries

  • Lians is a current-state and completeness guard, not a replacement for your model, Git, CI, repository instructions, or human review.
  • It does not compete with native Claude Code, Codex, Cursor, GitHub Copilot, Entire, Factory, and AI review workflows on all capabilities, nor does it claim that every project needs a separate memory layer.
  • The project is under active development; recovery features and preview Guard features are distinguished, and the repository does not imply production guarantees that do not yet exist.
  • The macOS and Windows desktop builds are still release candidates, pending platform signing and notarization.
  • The attached evidence is limited test results and does not mean every real coding session is extracted perfectly.
  • ContinuityBench v0.1 is still under development, and its current Lians fixture serves only as evidence for that fixture, not as a completed competitor leaderboard.
  • The Guard correctness benchmark is a local deterministic test of configurationpolicy, not proof of semantic correctness or production outcomes.

License and contributions Apache 2.0, see LICENSE. Please read CONTRIBUTING.md before contributing. Feature ideas, integration requests, and reproducible bugs are welcome in GitHub Issues.

Related tools

Browse category