AI Tools Integration Overview
Cloud Harness MCP is compatible with all leading AI agent environments, IDEs, and coding tools. Depending on the client's authentication capabilities, you will connect using either Managed OAuth or a Static API Key.
Compatibility Matrix
| AI Tool / Client | Connection Mode | Target Endpoint | Configuration Surface |
|---|---|---|---|
| ChatGPT | Managed OAuth | https://harness.zuey.me/mcp | Web Settings → Apps → Create Custom MCP |
| Claude Desktop | Managed OAuth (recommended) or Static API Key | https://harness.zuey.me/mcp (OAuth) or https://api.harness.zuey.me/mcp (key) | Connectors UI (OAuth) or claude_desktop_config.json (key) |
| Claude Code | Static API Key | https://api.harness.zuey.me/mcp | claude mcp add / env var |
| Cursor IDE | Static API Key | https://api.harness.zuey.me/mcp | Settings → Features → MCP |
| OpenAI Codex | Managed OAuth (recommended) or Static API Key | https://harness.zuey.me/mcp (OAuth) or https://api.harness.zuey.me/mcp (key) | ~/.codex/config.toml |
| Gemini CLI | Static API Key | https://api.harness.zuey.me/mcp | Environment header / CLI flags |
| Google Antigravity | Static API Key | https://api.harness.zuey.me/mcp | Antigravity MCP Settings |
| Grok / xAI | Static API Key | https://api.harness.zuey.me/mcp | Responses API tool definitions |
To connect a static-key client to the MCP Gateway instead, use the same key with https://api.harness.zuey.me/mcp-gateway.
Staying Signed In (Managed OAuth)
Managed OAuth clients sign in once in the browser and then refresh short-lived access tokens silently. They sign in again only when the OAuth grant expires. An operator sets that duration in Dashboard → Settings → MCP OAuth sessions; Stay signed in uses the provider maximum. Static API key clients are unaffected: their key's own expiry applies. If a client keeps asking you to sign in, see Troubleshooting.
Native Skills Support
Clients that implement the MCP Skills extension (io.modelcontextprotocol/skills, protocol revision 2026-07-28) can discover and read your built-in and owner skills directly from the HTTP /mcp endpoint. Serving a skill does not activate it: the client decides whether to load it, and any script still runs only through skills_run. Other clients keep using the skills_list, skills_read, and skills_run tools. See Skills over MCP.
Model Setup from Your AI Tool
Any connected client can prepare the model profiles your subagents use, so you can ask it to "pick a cheap model for reviews and set up a profile". The model_* tools list providers and your connected keys, search the catalog, check a connection, dry-run and save profiles, and read estimated spend. Three things stay with you on the Dashboard Models page:
- Keys. Add provider keys on the Dashboard. No tool accepts or returns a key.
- Activation. A profile your AI tool creates arrives Disabled. Review it and click Activate before
agent_spawncan use it. - Escalations. Raising a limit or lowering a price, widening an agent's proxy tools, or pointing a Custom key at a new URL returns
OWNER_APPROVAL_REQUIRED; make that change on the Dashboard. See Troubleshooting.
model_profile_update is marked destructive, and model_connection_verify spends a few provider tokens, so a client that asks before such tools will prompt for them. The full flow is in Model Providers.
Security Best Practices for Clients
- Keep Secrets Local: Never paste API keys into prompts, git commits, or shared project configs.
- Use Environment Variables: Reference tokens via environment variable substitution (e.g.
$CLOUD_HARNESS_MCP_TOKEN) in tool config files whenever supported. - Approval Modes: For tools modifying the filesystem or Git (
files_write,git_push), configure your AI client to require interactive confirmation before execution.