CLEO is a harness-agnostic core for AI coding agents: project management,
evidence, memory and docs, plus the coordination that lets agents in different
harnesses work on one project. It has no agent loop of its own. Agents run in
their own harness (Claude Code, Codex, OpenCode, Kimi, Gemini and others) and
call one CLI, cleo <command>; every call returns a single JSON envelope on
stdout, so any harness can read it. Any agent can call it and any project can
integrate it.
CLEO keeps each change traceable from task to pull request to release, demands
evidence before a task counts as done, remembers what was learned across
sessions, and keeps documentation from rotting. Its data lives in one SQLite
store per scope: <project>/.cleo/cleo.db for the project, and a global
cleo.db in the CLEO home directory for what spans projects.
| Area | What you get | Start with |
|---|---|---|
| Tasks | Sagas, epics and tasks with acceptance criteria, dependencies and ready waves; sessions that survive interruptions | cleo add, cleo next, cleo focus <id> |
| Evidence | A task is done only when its gates are proven: commits and files, CI or test runs, lint and typecheck, linked to each acceptance criterion | cleo done <id> --plan |
| Pull requests | A PR names its tasks (Closes: T#### AC1,AC2); an independent reviewer records a verdict per criterion; merging on green CI closes the task |
cleo done --from-merge <pr> --dry-run, cleo review packet |
| Docs | One canonical home per doc, stable names, published mirrors and drift checks (see below) | cleo docs query, cleo docs update |
| Memory (BRAIN) | Observations, decisions and lessons kept across sessions and searchable | cleo memory observe, cleo memory find, cleo briefing |
| Code intelligence (Nexus) | A symbol graph of your repo: callers, blast radius, impact before you edit | cleo nexus analyze, cleo nexus impact <symbol> |
| Release | One command tags the newest commit on main with green CI and publishes from changesets | cleo release cut <version> --dry-run |
| Multi-agent work | A lead splits an epic into waves; cleo orchestrate spawn prepares a worker's prompt for any harness, and cleo orchestrate spawn-execute can start a harness CLI through a provider adapter. Each worker gets its own worktree and session binding. AgentMBX is the messaging layer between agents: --lead-mbx names the lead's mailbox in the worker's prompt |
cleo orchestrate start <epic>, cleo orchestrate spawn <task> |
| Autonomous runs (CLEO Goal) | CLEO Goal (@cleocode/goal, bin cleo-goal) is the one autonomous driver: it keeps a harness working toward a goal through that harness's own hook. CLEO drafts the goal from task criteria and closes tasks from the run's evidence |
cleo goal draft, cleo goal install, cleo goal status |
Run cleo <command> --help for the full flags of any command.
npm install -g @cleocode/cleo
cd your-project
cleo init
cleo session start --scope global --name "first session"
# Work is a tree: a saga at the root, epics under it, tasks under those.
# By default every task needs at least 3 acceptance criteria, and an epic
# needs 5 plus a description.
cleo saga create --title "Ship the auth hardening" \
--acceptance "login is rate limited|sessions expire|both are documented"
cleo add "Auth hardening" --type epic --parent T001 \
--description "Rate limiting and session expiry are live and tested" \
--acceptance "rate limit shipped|session expiry shipped|tests cover both|docs updated|no new lint errors"
cleo add "Rate-limit the login endpoint" --parent T002 \
--acceptance "returns 429 after 5 failed attempts|the limit resets after 15 minutes|a test covers the limit"
cleo next # what to work on
cleo focus T003 # identity, blockers, docs and memory for one taskThen do the work and close it.
- With a pull request (the normal path): work on a branch, open a PR whose
body says
Closes: T003 AC1,AC2,AC3(every criterion the task has), check it withcleo done --from-merge <pr> --dry-run, get a per-criterion review (cleo review packet <pr> --task T003, thencleo review verdict), and merge on green CI. The merge sweep records the evidence and closes the task. - Without a pull request:
cleo done T003 --plannames the evidence each gate needs; record it withcleo verify, thencleo complete T003.
Finish with cleo memory observe "<what you learned>" --title "<short title>"
and cleo session end --note "<handoff>". The next session starts with
cleo briefing.
AI agents write a lot of documentation, and most of it rots. An agent that cannot find the existing doc writes a new one; names carry task ids and dates; two copies drift apart; nobody can tell which one is current. After a few months the docs folder is a pile of near-duplicates that agents keep extending.
CLEO Docs fixes this at the source:
- One canonical home per doc. The store is the source of truth; files in
docs/are published mirrors. Find before you write (cleo docs query "<topic>"), and update the existing doc (cleo docs update <slug>) instead of adding a second one.cleo docs addis for genuinely new docs. - Stable names. Every doc name and slug is lowercase kebab-case with no
task id, session id or date; task and decision links live in metadata. A CI
gate (
cleo check canon docs) refuses new violators. - One kind per folder.
docs/adr,docs/spec,docs/plan,docs/research,docs/guide,docs/concept,docs/reference,docs/migration,docs/releaseanddocs/note. - Drift is visible.
cleo docs statuscompares each published mirror with the store,cleo docs rotflags docs whose covered code changed after them, andcleo docs doctorrepairs the store. - Messy folders can be migrated.
cleo docs modernize <folder>plans a kind, a stable name and keep/rename/merge/delete for every file, and applies it with a receipt you can roll back.
The how-to is docs/guide/cleo-documentation-sop.md.
The CLI is @cleocode/cleo; it depends on the rest. The main packages:
| Package | Purpose |
|---|---|
@cleocode/cleo |
The cleo command-line interface |
@cleocode/core |
The SDK: tasks, sessions, evidence, docs, memory, orchestration, release |
@cleocode/contracts |
Shared types and operation contracts |
@cleocode/lafs |
The JSON envelope format every call returns |
@cleocode/nexus |
Code intelligence and the cross-project registry |
@cleocode/caamp |
Provider registry and agent package management |
@cleocode/adapters |
Provider adapters (Claude Code, Codex, Cursor, Gemini, Kimi, OpenCode and more) |
@cleocode/skills |
The ct-* skills that teach agents how to use CLEO |
@cleocode/cant |
The CANT agent-definition language |
Requires Node.js >= 24.16.0 and pnpm.
git clone https://github.com/kryptobaseddev/cleo.git
cd cleo
pnpm install
pnpm run buildRepository rules (package boundaries, architecture gates, evidence) are in
AGENTS.md; the system overview is
docs/concept/cleo-architecture-guide.md;
see CONTRIBUTING.md to contribute.
Every change lands through a pull request that names its tasks; releases are
cut from main with cleo release cut.
If CLEO helps your workflow, consider supporting development:
MIT License - see LICENSE for details.
