Skip to main content
Korve offers three complementary agent surfaces: the native CLI, the remote OAuth MCP endpoint, and customer skills. Local setup never writes an organization API key or OAuth token into the project.

​
Connect MCP

mcp install is idempotent and manages only the korve entry. Other MCP servers and supported configuration comments survive. Invalid existing JSON or TOML fails closed instead of being replaced. Use --client <name> for one client and --global for user scope. Windsurf’s native location is user-scoped. The saved entry contains only https://mcp.korve.dev/mcp; the client performs browser OAuth on first connection. status --json exits 2 when an entry is missing, invalid, or points somewhere else.

​
Install project agent files

The CLI installs three focused, offline skills (korve, korve-cli, and korve-debug) into .claude/skills and .agents/skills. It owns only its korve* subdirectories. Re-running restores Korve-managed files without touching sibling skills. It also maintains a marked block in AGENTS.md and adds only .env.korve.local to .gitignore. The committed skill and AGENTS.md content contain non-secret project identity and public URLs; credentials remain in local CLI storage or injected runtime variables. The six broader skills and their installer are documented in Skills.

​
Pull local environment values

The default output is .env.korve.local. Korve writes it atomically with owner-only permissions on supported Unix systems, refuses symlink destinations, adds it to .gitignore, and will not replace it without --force. Output reports the path, environment, project, and count—not secret values. Environments do not inherit values. Pull or set each environment explicitly, and never commit the rendered local file.

​
Which surface to use

  • Use MCP for searched, typed, approval-aware agent operations.
  • Use the CLI for deterministic local scripts, repository setup, and parity with API-key operations.
  • Use skills to teach the agent workflow and safety judgment.
  • Use manifests for reviewable multi-resource desired state.