Overview
Carina is the local-first runtime for coding agents — policy, hash-chained audit, and transactional rollback.
Carina is a local-first AI agent runtime. It sits where a coding agent leaves chat and starts acting on a real repository — under policy, with an audit trail you can verify, and patches you can roll back.
Why Carina
Section titled “Why Carina”Three promises — everything else hangs off these:
| Promise | Meaning |
|---|---|
| Policy before effect | Fourteen capabilities, profiles, approvals — the model never bypasses the kernel |
| Hash-chained audit | Append-only events; carina audit verify detects tampering |
| Transactional rollback | File mutations via patches with provenance, not silent overwrites |
Full power for the agent. Full control for you. Authority stays on your machine.
Get started in three steps
Section titled “Get started in three steps”Install
curl -fsSL https://carina.nebutra.com/install.sh | shMore options: Installation.
Run a task
carina daemon start && carina doctorcd /path/to/your/repo && carina run "Summarize this repo"Prove it was governed
carina audit SESSION && carina audit verify SESSION
Full tour: Quickstart · Recipes: Common workflows.
Who it is for
Section titled “Who it is for”- Engineers who want agents on local repos with audit and rollback
- Platform teams building an internal agent runner
- Security-minded teams watching files, shells, and secrets
- Builders embedding a runtime behind IDE, TUI, CI, or web UI
Not primarily: a finished editor chatbot, or a hosted cloud agent product.
How it fits together
Section titled “How it fits together”flowchart TB C["TUI / CLI / SDK / IDE"] -->|JSON-RPC| G[Go control plane] G --> K[Policy kernel] G --> A[Hash-chained audit] G --> P[Patch / rollback] K --> W[Workspace tools] G --> M[Models / BYOK]
Runtime layers, sessions, policy, audit.
Agents, tools, workflows, workers.
JSON-RPC, method catalog, playground.
Surfaces
Section titled “Surfaces”| Surface | Role |
|---|---|
| CLI / TUI | Local interactive coding and ops |
| JSON-RPC + Gateway | Embed in your product |
| SDKs | TS / Python / Go typed clients |
| Workers | Distributed execution with leases |
| MCP | External tools, still policy-gated |
Same engine: session + profile + audit wherever you attach.
What makes Carina different
Section titled “What makes Carina different”| Category | Optimized for | Carina |
|---|---|---|
| Editor assistants | In-IDE chat | Runtime governance behind any client |
| CLI coding agents | Conversational terminal UX | Durable sessions, audit, rollback, workers, embed |
| Cloud agent tasks | Hosted remote UX | Local-first; cloud is optional boundary |
| Homegrown stacks | One-off orchestration | Reusable policy / audit / patch primitives |
Non-goals
Section titled “Non-goals”- Full editor product
- Hosted managed agent SaaS as the core
- Replacing Git history or code review
- Pretending alpha packaging is finished enterprise distribution
Source of truth
Section titled “Source of truth”- Product framing:
docs/product.md· cloud boundary:docs/nebutra-cloud-boundary.md - Architecture map:
docs/architecture.md· docs map:apps/docs/FEATURE_MAP.md - Live CLI:
carina --help· CLI reference
| Goal | Page |
|---|---|
| Install on my machine | Installation |
| First end-to-end run | Quickstart |
| Day-to-day recipes | Common workflows |
| Terms | Glossary |
Was this page helpful?
What went wrong?
Thanks for the feedback.
One vote per page is recorded during this session.