MCP Integration Reference – Agent Tool Specification
Homegrown Intelligence workflows expose a Model Context Protocol server on port 8100 by default. The MCP server enables AI agents to query workflow status, initiate inference jobs, retrieve results, and adjust runtime parameters through typed tool calls. This document describes every available tool, its input schema, and the expected response format.
Connection Details
- Transport
- HTTP SSE with JSON-RPC 2.0 framing
- Default endpoint
http://localhost:8100/mcp/v1- Authentication
- Bearer token passed via
Authorizationheader - Protocol version
- 2025-03-26
Tool: flux2_generate
| Parameter | Type | Required | Default | Description |
|---|---|---|---|---|
| prompt | string | yes | — | Text description for image generation |
| width | integer | no | 1024 | Output width in pixels, 512–2048 |
| height | integer | no | 1024 | Output height in pixels, 512–2048 |
| steps | integer | no | 28 | Inference steps, 4–50 |
| guidance_scale | number | no | 3.5 | CFG scale, 1.0–10.0 |
| seed | integer | no | random | Reproducibility seed |
Returns: {image_id, width, height, seed, generation_time_ms}
Tool: whisperx_transcribe
| Parameter | Type | Required | Default | Description |
|---|---|---|---|---|
| audio_path | string | yes | — | Absolute path to audio file on server |
| language | string | no | auto | ISO 639-1 language code |
| output_format | string | no | srt | srt, vtt, json, txt |
| diarize | boolean | no | false | Enable speaker diarization |
Returns: {job_id, segments_count, duration_seconds, output_path}
Tool: llm_complete
| Parameter | Type | Required | Default | Description |
|---|---|---|---|---|
| messages | array | yes | — | Chat completion messages array |
| max_tokens | integer | no | 2048 | Maximum output tokens |
| temperature | number | no | 0.7 | Sampling temperature |
| stream | boolean | no | false | Enable SSE streaming |
Returns: {id, choices, usage} matching OpenAI chat completions format
Error Handling
All tools return JSON-RPC error codes for common failure modes. Code -32000 indicates a parameter validation failure with a descriptive message in the data field. Code -32001 signals a resource exhaustion condition such as insufficient VRAM. Code -32002 reports a model-loading failure when the weight file is corrupt or missing. Clients should implement exponential backoff with jitter for transient errors and surface permanent failures through the agent framework's native error reporting.