Privacy by architecture
Use an Anthropic-compatible endpoint you operate. Task execution and source stay on your machine.
Local-first AI software factory
Turn a project queue into reviewed pull requests.
Celestial coordinates local planning models, isolated coding agents, and deterministic checks—then hands you a pull request to review.
- Local model endpoints
- Human-reviewed output
- GitHub-native delivery
Pipeline
Ready
- 01 Select queued task
- 02 Plan locally
- 03 Build in isolation
- 04 Verify deterministically
- 05 Open pull request
Local-firstModels stay on infrastructure you control
IsolatedEvery task gets a dedicated worktree
VerifiedTests and coverage gate every result
ObservableGitHub and Telegram keep you informed
Built for real constraints
Automation without surrendering control.
Celestial is designed for teams that want agentic throughput while keeping source, execution, and the final decision close at hand.
Deliberate memory use
Sequential planning and implementation profiles make useful automation practical on Apple Silicon.
Pull requests, not surprises
Changes land as reviewable branches with tests, coverage evidence, and a clear audit trail.
One controlled loop
From intent to evidence.
-
01
Queue
Pick an eligible item from a GitHub Project.
-
02
Plan
A compact local model creates the implementation plan.
-
03
Implement
The coding agent works inside a dedicated Git worktree.
-
04
Verify
Deterministic tests, coverage, and path checks fail closed.
-
05
Review
Celestial opens a pull request; you remain the merge authority.
Boundaries are features
Designed to fail closed.
- Tracked paths are contained inside the task worktree.
- Build commands require an explicit sandbox policy.
- Memory and swap gates protect resource-constrained machines.
- Retries are bounded and failed worktrees remain inspectable.
Celestial is local-first, not air-gapped. GitHub and optional Telegram receive task metadata; model privacy depends on your configured endpoint.
Start with one task