Spacedock
Turn multi-stage agent development work into a reviewable, traceable queue of decisions.
- Source repo
- spacedock-dev/spacedock
- Stars
- ★ 86
- Last updated
- 1d ago
- License
- Apache-2.0
- Primary language
- Go
- FA score
- 64/100 · Some gaps
At a glance
- How it runs
- Works with
- Portable with changesCodex · Claude Code
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- A developer trying to map an existing project's implicit agent workflow and open decisions can run
/spacedock:survey. - Not a fit if
- Teams using none of Claude Code, Codex, Pi, or compatible skill systems
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-19latest release v0.28.1 (2026-11-18)
- Source review
- 64/100 · Some gaps
What does this agent do, and when should you use it?
Spacedock is a multi-agent orchestrator that runs within coding agent harnesses such as Claude Code, Codex, and Pi. It divides work into stages, advances multiple queued items, and surfaces decisions at each stage boundary. Each decision comes with a stage report covering findings, verdicts, artifacts, and anomalies; a person can approve, send back, or escalate it, or delegate the call to an agent. Making and reviewing happen in separate stages, with the reviewer using fresh context and no access to the maker's reasoning. After installation through Homebrew or an install script, users launch it from a terminal and can survey an existing project.
Run /spacedock:survey to have Spacedock read an existing project's agent history and show its implicit workflow, definition of done, and open decisions. The orchestrator splits work into stages and advances queued items; each stage works to its written criteria and produces a report of findings, verdicts, artifacts, and anomalies. A separate review stage checks evidence in fresh context and pushes back on thin support. At decision gates, a person can approve, send back, or escalate work, or delegate the call; /spacedock:debrief captures session learnings, and an agent can propose edits to written criteria for approval. Stages that touch shared state run in their own Git worktree, while lighter stages run inline; when context runs out, a successor agent carries forward work in flight.
- A developer trying to map an existing project's implicit agent workflow and open decisions can run
/spacedock:survey. - A team handling several coding items at once can queue them and review decisions as stages finish.
- An engineering team that wants evidence checked in a separate context before accepting a stage can use the independent review stage.
- Maintainers who want stage criteria and session learnings to carry into future work can use criteria proposals and
/spacedock:debrief. - Developers who need isolation for shared-state work but want lighter stages to run inline can use its worktree and inline execution modes.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Install / run
brew install spacedock
- Install notes cover
- macOS
How do you install or deploy this agent?
A coding agent harness such as Claude Code, Codex, or Pi is a prerequisite. Install the stable channel with Homebrew:
brew tap spacedock-dev/tap
brew install spacedockFor the edge channel, which follows prereleases, install the spacedock@next cask or run the install script with SPACEDOCK_CHANNEL=edge. The README does not provide the install script's exact command.
How do you use this agent?
On first launch, Spacedock sets up the plugin. To survey an existing project with Claude Code, run:
spacedock claude "/spacedock:survey"Examples for Codex and Pi:
spacedock codex "/spacedock:survey"
spacedock pi "/spacedock:survey"Use /spacedock:debrief to capture session learnings. The documentation also includes an install walkthrough, concepts, and a command reference.
What are this agent's strengths and limitations?
- The separate review stage uses fresh context without access to the maker's reasoning and challenges unsupported claims.
- Each stage decision includes a report of findings, verdicts, artifacts, and anomalies for later traceability.
- Users can queue work in batches, and stages that touch shared state can run in isolated Git worktrees.
- Agents can propose edits to stage criteria for approval, while
/spacedock:debriefcarries session learnings forward.
- It requires an existing coding agent harness; the README names Claude Code, Codex, and Pi, while other harnesses depend on skill-system compatibility.
- Installation and first use involve terminal commands and access to a project, so it is not a direct fit for teams that avoid the command line.
- The README does not state software pricing, model charges, or required credentials, so total cost cannot be determined from it.
- The README does not specify supported sandbox profiles, permission scopes, or detailed recovery behavior for failure cases.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Spacedock This agent | 64 · Some gaps | CLI | ★ 86 | 1d ago | Go | Codex · Claude Code |
| Maestro: Agent Orchestration Command Center | 61 · Some gaps | Desktop appFree + model costs | ★ 3.6k | today | TypeScript | Codex · Claude Code |
| Friday Agent Workbench | 73 · Some gaps | Desktop appFreemium | ★ 83 | today | TypeScript | Codex · Claude Code |
| Open Dynamic Workflows | 69 · Some gaps | CLIFree + model costs | ★ 115 | 1mo ago | TypeScript | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
The docs describe stage isolation, independent review, approve/return/escalate gates, and binary-owned path resolution and mutation guards. The extension install contract also specifies checksum verification and reporting/skipping failures, supporting bounded permissions and user control (least_privilege 2, user_confirmation 2, external_effects 2). Compass says it reads local session files without uploading, and debrief limits extraction to local Git and files; these cover only some data paths, with no visible sensitive-data handling or retention policy (data_flow_transparency 2, sensitive_data_handling 1). Go dependencies are version-pinned and extension archives are checked with SHA256, but there is no evidence of dependency auditing or vulnerability response (dependency_security 2). Worktree isolation and failure behavior offer limited recovery, but no general rollback mechanism is documented (rollback 1). README, plugin metadata, and contribution guidance identify an author, repository, and issue channel; release maintenance is described, while publisher identity remains unverified (source_attribution 2).
The README's approval gates, review stages, and stage reports broadly align with the extension contract and skill flows. However, the plugin and marketplace manifests show different versions, and some descriptions in this partial packet are incomplete or difficult to reconcile, so points are deducted (self_consistency 1). Go module dependencies are version-pinned, and installation guidance covers Homebrew, scripts, and host selection; external services and host availability remain environmental dependencies (dependency_availability 2). Code and contract material provide explicit error text, exit codes, and stop-or-skip behavior on failures (failure_messages 2).
The docs target Claude Code, Codex, and Pi, and describe survey, workflow design, batching, and session handoff scenarios; the audience and scenarios are clear but not exhaustive (audience_and_scenarios 2). Skills define responsibilities, write scopes, capability checks, and host adapters, though this packet contains only a sample of skills (capability_boundaries 2). Trigger descriptions are concrete, including skill phrases, commands, Pi launch markers, and trigger conditions covered by the included test source (trigger_precision 2). Installation guidance covers multiple hosts, stable/edge channels, and native sandbox integration, while full environment requirements mostly reside in linked documentation (environment_fit 2).
The repository separates CLI, internal packages, skills, specifications, and development docs; README points to installation, docs, and contribution guidance (information_architecture 2). README provides install commands, channel differences, and launch examples (install_notes 2). Product, plugin, and command naming is generally stable and conventions are clear, though the plugin and marketplace version labels differ (naming_stability 2). The packet includes survey and commission examples and detailed skill procedures, but no FAQ is visible (examples_and_faq 2). The extension contract describes compatibility and failure cases and README says the project is early; there is no complete, unified known-limitations list in this sample (known_limitations 1). LICENSE explicitly grants Apache-2.0 and metadata agrees (license 3). Stable/edge versions, release tags, and release procedures are described, but the packet does not provide a complete, inspectable changelog (versioning_changelog 2). Author, GitHub issues, CONTRIBUTING, and release docs give maintenance and update paths, but actual responsibility and maintenance state are not independently confirmed (maintenance_responsibility 2).
Stage reports, evidence-based gates, batching, and context handoff are designed to make decisions actionable, and skills provide concrete procedures and commands for ordinary workflows (output_usability 3). Independent review and decision records address the coordination and ownership problem described in the README, but the benefit claim is mainly supported by the project's own account (marginal_value 2). The zero-dependency workflow description, defined install paths, and staged automation suggest a cost-saving design; there is no evidence about runtime cost, measured savings, or scale (cost_benefit 2).
The README's main product claims map to mechanisms in skills, the extension contract, source, or workflow configuration, though the packet is partial and some behavior is documented rather than demonstrated (claim_traceability 2). README, Go source, plugin metadata, skills, install contract, and CI workflow provide partial corroboration across sources; not every product claim has evidence from multiple sources (cross_source_corroboration 2). The docs distinguish design intent from specific mechanisms reasonably well, but effectiveness and reliability claims remain product claims and are not treated as verified outcomes (fact_inference_separation 2).
- This assessment uses only the supplied static sample and has low confidence; no execution or independent testing is claimed.
- The packet contains only selected skills and source files; unseen files are unknown.
- The README calls the workflows zero-dependency, while go.mod lists several dependencies; the claim may concern workflow configuration or a particular layer and needs the full install docs for context.
- The plugin and marketplace manifests show different versions, so the release state needs clarification.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
Spacedock 0.28.1 makes the first install and the first workflow work without surprises. - The installer and the install guide now use the stable channel, and the installer's closing check no longer reports Review and Compass as missing on a fresh install. - spacedock update checks every host that holds a Spacedock piece. It fails with a clear message when a plugin or a Pi extension package is missing, and it says how to change a marketplace entry that Claude Code refuses. …View on GitHub ↗ -
Release notes excerpt
Spacedock 0.28.0 — one install brings Spacedock Review and Compass along, Pi becomes a host you can start on from the docs, and finished work is only marked finished once its delivery is proven. - Install the whole kit in one step: the installer now sets up Spacedock together with Spacedock Review (Subspace) and Compass, and spacedock update keeps the programs and plugins on one release — on Claude Code, Codex, and Pi. A failed install attempt no longer silences the offer in later sessions. …View on GitHub ↗
FAQ
Which coding harnesses does Spacedock support?
Does it automatically approve or ship work?
What goes into a stage report?
What does it cost, and which model credentials are needed?
What happens if an agent runs out of context?
Is Spacedock open source?
Is Spacedock safe to use?
Is Spacedock still maintained?
Who makes Spacedock?
Thanks — we use this to improve the page.