Spacedock
Spacedock is a multi-agent orchestrator where nothing ships without a decision.
It lives within your existing harness: Claude Code, Codex, or Pi. It breaks work
into stages and surfaces the decisions each stage needs, batched for you. Each
decision arrives with evidence measured against a predefined bar for what good
looks like. You approve, send back, or escalate. Or you delegate the call to an
agent. Either way, the decision is recorded with its evidence and reason.
Why?
- You're the human, and agents from a dozen sessions ping you all day: a design
call, a one-line approval, "ship without full coverage?", none of it on a
schedule you can plan around. You stopped deciding the work and started just
answering the agent.
- You're the agent, and you stall every few steps waiting on a human who isn't
watching, with no clear scope and no clear bar for done, so you ask and then
you wait.
- Generation got cheap. Your attention and judgment are now the bottleneck.
Every task still ends in a decision, and no one owns it, so it falls on
whoever's around.
Start with what you already built. Point Spacedock at a project you
vibe-coded into spaghetti and run /spacedock:survey. It reads your own agent
history and shows you three things: the workflow you've been running without
naming it, how you've been calling work done, and the decisions still open and
waiting on you.
What's different
- The agent doesn't get to judge its own work. Review runs as a separate
stage with fresh context, no access to the maker's reasoning. It pushes back on
thin evidence and work that looks busy without proving its claim.
- Every decision leaves a trail. Each gate, the decision point at the end of
a stage, carries a stage report: findings, verdicts, artifacts, anomalies. You
decide on evidence, not the transcript.
The record outlives the reviewer, so you can trace a bad result back to the
call that caused it.
- The bar sharpens as you use it. Each stage declares what good means, and
the agent works to that line on its own. When a standard turns out fuzzy in
practice, the agent proposes an edit to the stage's written criteria for your
approval.
/spacedock:debrief captures each session's learnings so the next
one starts from them.
- Batch the work; decide as it flows back. Queue many items at once. Agents
advance each through its stages. You handle gates as they surface, not one
session at a time.
- Isolation when it matters. Stages that touch shared state run in their own
git worktree. Lighter stages run inline.
- Native sandbox integration. Drop a profile in the project and Spacedock
runs the agent sandboxed.
- Work survives the context limit. When an agent runs out of context, a
successor carries forward what's in flight.
Install
Prerequisite: a coding agent harness. Claude Code, Codex, and Pi are tier-1
supported; through skill systems it also runs in most other harnesses, including
Hermes-class agents.
Install with Homebrew:
brew tap spacedock-dev/homebrew-tap
brew install spacedock
Then launch. The first launch sets up the plugin for you, so a single line gets
you a working session. Point it at a project you already have and let it survey:
spacedock claude "/spacedock:survey"
Using Codex or Pi instead? Swap the subcommand: spacedock codex "/spacedock:survey"
or spacedock pi "/spacedock:survey".
Full docs — the install walkthrough, the Codex and Pi paths, concepts, and the
command reference — live at spacedock.md/docs.
Browsing the repo on GitHub? The same install guide is at
docs/site/get-started/install.md.
License
Spacedock is released under the Apache License 2.0.
Contributing
Spacedock is early; we welcome proposals as GitHub issues.
See CONTRIBUTING.md.
