IDEs
IDE Integration Overview
An editor can connect to Verdictan in two separate ways. Pick the one that matches what you want, or use both.
- Model traffic. The editor sends its model requests to your Verdictan gateway, so your policies apply and usage is recorded. Start with Gateway Setup.
- MCP tools. The editor's agent gets Verdictan tools, such as searching history or linting a policy. This does not change where the editor sends model requests.
Two integration paths
| Path | Endpoint | What Verdictan governs |
|---|---|---|
| Model traffic | http://127.0.0.1:41002/v1, or your published model hostname | Model requests the editor sends to this base URL |
| MCP tools | https://<published-hostname>/mcp | Verdictan tool and resource calls from the editor's agent |
Only the model path governs prompts and answers. For that, the editor must let you set a custom OpenAI-compatible base URL. Chat Completions, Responses, and Messages requests all pass the same access, input, tool, output, and audit stages; use whichever the editor and your provider support. See Runtime Request Families.
Port 41002 serves an LLM API. It is not a forward HTTP proxy. Don't point HTTP_PROXY, HTTPS_PROXY, or an editor's network proxy setting at it. An assistant that has no custom-endpoint setting cannot send its model traffic through Verdictan.
Client support by path
Each guide covers the setup below. Check every feature you rely on: chat, inline completions, and agent mode can use different models or endpoints in the same editor.
| Client guide | Connection |
|---|---|
| Cursor | MCP tools |
| Claude Desktop | MCP tools, with request-header authentication |
| Windsurf | MCP tools in Cascade |
| VS Code and GitHub Copilot | Custom-endpoint chat, plus a separate MCP connection |
| JetBrains AI Assistant | Custom model endpoint, optional MCP |
| Zed | Custom model endpoint, optional MCP |
| VS Code and JetBrains with Continue | Model requests |
| Qwen Code | Model requests |
| VS Code and JetBrains custom clients | Model requests from a configurable extension or plugin |
| Emacs | Model requests |
| Neovim | Model requests |
| Xcode | Copilot Chat with a custom endpoint |
Authentication boundaries
An editor setup involves up to three credentials. Keep them apart:
| Credential | Held by | Sent to |
|---|---|---|
Gateway runtime token (VERDICTAN_API_TOKEN) | The verdictan gateway run process | The Verdictan control plane |
| Client API token | The editor's model client or MCP client | The Verdictan gateway, as a Bearer token |
Provider credential (secret_key_ref) | The gateway's provider target | The upstream model provider |
Never paste the gateway runtime token into an editor. The Console lists client tokens on the API Tokens page as an Access Key or a Gateway Key; both come from the same token system. See API Tokens for IDEs.
Recorded data
Every request through the gateway produces policy and audit events, and the gateway works out its cost from usage and price data. Request content appears in History only when capture is on. Traffic an editor sends straight to its vendor is invisible to Verdictan.
To confirm an editor goes through the gateway, send one request from the editor and then run:
verdictan events tail --since 10m --jsonA matching event proves the request passed through Verdictan. No event does not prove it bypassed the gateway. Check in order:
- The gateway mode delivers events.
- The CLI queries the same organization and region as the gateway.
- The time window and filters include the request.
- The editor's active model base URL points at the gateway.
Next steps
- Gateway MCP Surface: the MCP endpoint's connection details and tools.
- IDE Integration Troubleshooting: an editor that cannot connect, or events that don't appear.