Measure twice, code once.
Cadre is a human-governed, Git-aware delivery harness for OpenAI Codex, Claude Code, and the native Zed Agent. Codex, Claude Code, and native Zed Agent integrations are stable as of Cadre 3.9.1. Cadre turns approved project context into resumable feature and bug tracks, carries learning forward between phases, and records implementation provenance in Git.
Cadre is installed as a user integration: stable native plugins for
Codex/Claude and stable global skills plus a custom MCP server for Zed. Its bundled TypeScript MCP
server provides immutable versioned templates and narrow state/Git operations.
A target project keeps only approved mutable delivery state under .cadre/;
runtime code, workflow skills, worker definitions, and templates stay in the
installed payload.
Cadre 4.0.0
The 4.0.0 release consolidates state mutations around immutable operation receipts and ships only template v6 with 22 MCP tools. All ten workflows and the existing approval/scheduling modes remain available. Compatible 3.9.0/3.9.1/3.9.2 projects using schema-2 state and v5 need no migration or version-only refresh. Older formats receive an explicit Cadre 3.9.0 bridge route.
See release notes for breaking API changes, verification results and publication status, and getting started for installation and upgrade instructions.
The upcoming roadmap adds deeper GitHub / GitLab integration for multi-person coordination, a knowledge base built during work and reused as context, and polyrepo support. Each will be optional, independently opt-in and disabled by default.
What Cadre Solves
| Need | Cadre's approach |
|---|---|
| Project context | Human-approved product, engineering, stack, workflow, styleguide, and pattern artifacts. |
| Delivery planning | Feature or bug specifications followed by a validated phase/task dependency DAG. |
| Implementation | Parallel by default when safe, explicitly sequential when requested, with isolated worktrees. |
| Human control | Every artifact and lifecycle transition is presented before mutation. |
| Recovery | Durable setup, operation, and execution journals reconcile files and Git after interruption. |
| Quality | Per-phase and track-level manual verification followed by evidence-backed review. |
| Learning | Dependency-aware phase learning and durable pattern distillation during archive. |
Lifecycle
refresh updates project context, revert prepares additive Git reversals,
status validates without mutation, and wisp explores without entering the
tracked lifecycle.
Ten Workflow Skills
| Skill | Purpose |
|---|---|
create | Initialize or resume an approved .cadre/ project. |
track | Specify and plan a feature or bug. |
implement | Execute the approved dependency DAG. |
review | Record approved findings or complete a clean review. |
revise | Change active approved scope or propose a successor. |
archive | Archive completed tracks and distill patterns. |
refresh | Reconcile project context with current evidence. |
revert | Additively reverse a task, phase, or track. |
status | Validate and summarize current project state. |
wisp | Perform lightweight untracked exploration. |
Start Here
- Installation covers the
cadre-aiCLI, client installation, permissions, update, and uninstall behavior. - Quickstart walks through the first create-to-archive cycle.
- How Cadre Works explains approval, journals, state, and the MCP boundary.
- Workflows describes when to use each skill.
- Parallel Execution explains workers and worktrees.
- Complex Parallel Execution Walkthrough follows worker fan-out and phase integration end to end.
- Troubleshooting covers common recovery paths.