PortwayMCP Chat
MCP Chat connects an AI model to the Portway MCP tools. The model selects tools; each call goes through the regular API with authentication, rate limiting and environment scoping. Responses are streamed.
INFO
MCP Chat requires Mcp:Enabled: true and at least one endpoint with "Mcp": { "Exposed": true }.
Enable chat
Setting Mcp:ChatEnabled: true in appsettings.json enables the Chat UI and the /ui/api/mcp/chat SSE endpoint.
"Mcp": {
"Enabled": true,
"ChatEnabled": true
}Provider, model and API key are set in the setup wizard and stored encrypted in mcp.db, not in appsettings.json.
Configure a provider
The setup wizard opens on the first visit to /ui/mcp/chat and under Settings → AI. It sets the provider, the model and the API key. The API key is encrypted (PWENC:) before it is written to mcp.db.
Supply the API key via environment variable
The environment variable PORTWAY_CHAT_API_KEY overrides the key stored in mcp.db:
With the variable set, the wizard and the status endpoint report api_key_source: environment. PWENC encryption: Secret Encryption.
Supported providers
| Provider | API | Example model |
|---|---|---|
| Anthropic | Messages API with streaming and tool use | claude-sonnet-5 |
| OpenAI | Chat Completions API with streaming function calling | gpt-5.5 |
| Gemini | Generative Language API (streamGenerateContent) with function declarations |
gemini-3.5-flash |
| Mistral | Chat Completions API; codestral-* models use codestral.mistral.ai and a Codestral API key |
codestral-latest |
How tool calls work
Each chat turn runs up to five tool rounds:
- The conversation and the exposed tools are sent to the model.
- The model returns text or tool calls.
- Each tool call requests
{baseUrl}/api/{environment}/{endpoint}with the model's method, query, body and tenant values, under the user's token. - Tool results are added to the conversation for the next round.
- The loop ends with a text response or after five rounds.
Environment selector
The environment selector lists the environments from GET /ui/api/environments. A tool call uses the environment the model passes, or the selected environment. An environment outside the endpoint's AllowedEnvironments is replaced by the endpoint's first allowed environment.