Spacedock is the judgement layer for agent work.
It works inside Claude Code and Codex, so there is no new agent to learn. It runs on macOS and Linux.
| Feature | What it does | Start it | Code |
|---|---|---|---|
| Spacedock Review | Highlight what needs changing, add your comments, and submit your decision. The requesting agent receives it all. | /spacedock:review plan.md decision |
spacedock-dev/subspace |
| Spacedock Compass | See where agent work departed from your goal, with evidence to help you course-correct. | /spacedock:compass |
spacedock-dev/compass |
| Spacedock Workflow | Define what agents must check, which decisions they can make, and when they need you. Workflow records each decision with its evidence and reason. | /spacedock:commission |
This repository |
The commands are for Claude Code. In Codex, Review and Compass start with $
in place of /, for example $spacedock:review plan.md decision.
This repository holds Spacedock Workflow, the spacedock command and its
installer. One install brings Review and Compass with it.
# macOS with Homebrew
brew trust spacedock-dev/tap
brew install --cask spacedock-dev/tap/spacedock
# Linux, or macOS without Homebrew
curl -fsSL https://raw.githubusercontent.com/spacedock-dev/spacedock/main/install.sh | sh
export PATH="${SPACEDOCK_INSTALL_DIR:-$HOME/.local/bin}:$PATH"Then connect your agent with spacedock install --host claude or
spacedock install --host codex. Run spacedock doctor to check the setup
(spacedock doctor --host codex for Codex). The
installation guide has the
full steps.
Interactive launches check once a day for a newer Spacedock. The check asks
spacedock.md, which redirects to GitHub's releases API, so both see the
request. Add --no-update-check to a launch command, or set
SPACEDOCK_NO_UPDATE_CHECK=1, to skip it.
Paste this into Claude Code or Codex:
Install Spacedock using https://spacedock.md/agents.md. Help me choose a plan to review, then open it in Spacedock Review.
Your agent installs Spacedock and opens the plan in Review. Comment on the text, then approve, request changes or hold. The agent receives your decision and comments.
Read spacedock.md/agents.md for the install steps and how to bring a first artifact to Review.
A workflow gives recurring work a set of stages, such as design, implementation and review. You set the standards for each stage. You choose where work pauses for your decision; these pauses are called gates. The stages and standards live in plain text in your repository.
To start from the work you already did, run /spacedock:survey in Claude
Code. It reads your local agent history, without changes, and reports repeated
work, current workstreams and unresolved decisions. Your agent can then
propose a workflow from the findings.
To describe a workflow yourself:
spacedock claude "/spacedock:commission ship features through design, implementation, and review"Run /spacedock:debrief before you end a session. Run spacedock claude to
resume.
- The agent does not judge its own work. Review runs as a separate stage with fresh context and no access to the maker's reasoning.
- Every decision leaves a trail. Each gate carries a stage report with findings, verdicts and artifacts. You decide on evidence, not the transcript.
- The standards improve as you use them. Each stage states what good means. When a standard is unclear in practice, the agent proposes an edit for your approval.
- Work keeps moving. Queue many items at once and handle gates as they arrive. When an agent runs out of context, a successor continues the work.
- Documentation
- Release notes
- Discord
- spacedock-dev/subspace: Spacedock Review
- spacedock-dev/compass: Spacedock Compass
- Contributing
- License: Apache License 2.0