# CodeMarine

> Let AI build. CodeMarine stands guard.

CodeMarine is the independent security control plane for AI-assisted development. It works alongside Claude, Codex, Cursor and other coding agents to secure what they write, verify what they bring in and control high-impact actions on supported surfaces.

## Product model

### Secure what AI writes

CodeMarine continuously evaluates resulting code regardless of whether a frontier model, editor, automation or person produced it. The high-frequency path is deterministic and local-first. It does not require a model call for every change.

### Verify what AI brings in

CodeMarine examines dependencies, manifests, lockfiles, install behavior, suspicious packages, MCP configuration, plugins, skills, agent rules and prompt-injection signals.

### Control what AI does

CodeMarine evaluates defined destructive action classes. Provider adapters can return blocking decisions on supported shell or write paths after setup. Live-host proof does not exist yet, so no adapter surface is represented as Guarded. Native, remote and provider-hosted actions remain outside coverage unless a separate control path exists.

## Current capability

- Local-first deterministic code scanning
- Structural confirmation and graph context for supported finding classes
- Supply-chain scanning across several dependency ecosystems and repository artifacts
- AI artifact scanning for selected prompts, MCP descriptions, skills and rules
- Destructive command and sensitive action evaluators with sanitized verdicts
- Workspace monitoring and provider integration assets

## Beta capability

- Provider adapters for defined shell or write paths
- CodeMarine-mediated MCP operations

Setup, host behavior and proof limitations apply.

## Planned capability

- Per-surface Protection Center with live canaries and proof records
- Governed MCP and API gateway
- Scoped credential brokerage and remote sidecars
- Local specialist security model tier
- Frontier escalation, investigation and deterministic patch validation
- Organization SSO, SCIM, RBAC and compliance reporting

## Frontier-model collaboration

Frontier models are strong at novel vulnerability discovery, business-logic reasoning and contextual remediation. CodeMarine supplies continuous coverage, supply-chain verification, cross-provider policy and reproducible evidence. The planned workflow gives an approved model bounded context, then returns any patch through an independent CodeMarine gate.

## Trust standard

CodeMarine uses Current, Beta, Planned and Not covered labels on this site. The target standard for Guarded is a supported pre-execution mechanism plus current live proof tied to the host, workspace, configuration and policy revision.

## Pages

- Platform: https://codemarine.ai/platform/
- Runtime Safety: https://codemarine.ai/runtime-safety/
- Supply Chain and Agent Trust: https://codemarine.ai/supply-chain/
- Frontier Collaboration: https://codemarine.ai/frontier-collaboration/
- Integrations: https://codemarine.ai/integrations/
- Trust and Coverage: https://codemarine.ai/trust/
- Product FAQ: https://codemarine.ai/faq/
- CodeMarine Journal: https://codemarine.ai/blog/
- Documentation: https://codemarine.mintlify.app/get-started/introduction
- Contact: mailto:hello@codemarine.ai