AI integration compatibility
Which external tools have been tried with Fabric Desktop by KratosBI, how, and what didn't work. Only rows marked Tested are claimed to work. Architecture: ai-integration.md.
Protocol#
| Item | Supported |
|---|---|
| MCP protocol versions | 2025-06-18 and 2025-11-25 (handshake era) and 2026-07-28, negotiated by the MCP Python SDK |
| Transports | stdio (default); Streamable HTTP on 127.0.0.1 when switched on |
| Server features | Tools, resources and resource templates, prompts, tool and resource list-changed notifications (on the connection at 2025-11-25, on subscriptions/listen at 2026-07-28), cancellation |
| Not supported | SSE transport (deprecated by MCP), remote (non-loopback) HTTP, OAuth authorization server, progress notifications |
Clients#
| Client | Transport | Status | Notes |
|---|---|---|---|
| MCP Python SDK client 2.3.0 | stdio, in-memory, Streamable HTTP | Tested (automated, every build) | Protocol 2025-11-25 and 2026-07-28; over HTTP with a bearer token against a real server on 127.0.0.1 |
| Claude Code | stdio | Untested | |
| Claude Desktop | stdio | Untested | |
| OpenAI Codex CLI | stdio | Untested | |
| Cursor | stdio | Untested | |
| VS Code with GitHub Copilot (agent mode) | stdio | Untested | |
| httpx (Python) | REST /api/v1 |
Tested (automated, every build) | OpenAPI 3.1 at /api/v1/openapi.json; any HTTP client should work but only this one is tested |
To test a tool, follow the AI tool test checklist and record the result here.
Untested means a configuration example exists (setup for each tool) but nobody has run it with that client yet. MCP prompts (workflows) are tested with the SDK client only; whether a client shows them, and how, is part of each client's test. To report a result, open an issue with the client name and version, the transport, and what happened.
Conformance#
Every build runs the same protocol checks through all three ways a tool can connect (in memory, the stdio process, Streamable HTTP) on both protocol eras, and checks the REST API against its OpenAPI document: sidecar/tests/test_integrations_conformance.py. The MCP project's own suite (@modelcontextprotocol/conformance 0.1.16) was run by hand on 2026-10-17: every scenario that applies to this server passed except one that is stricter on purpose (requests with any Origin header are refused). Details, and how to run it again, are in the checklist.
| Not supported | Why |
|---|---|
| Sampling and elicitation (the server asking the tool's model or user for something) | The server never asks an AI model for anything, and approvals happen in Fabric Desktop by KratosBI, not in the tool |
logging/setLevel, completion/complete, resource subscriptions |
Not needed by any workflow yet; list-changed notifications cover capability changes |
Requests from web pages (any Origin header) |
Only programs on this PC may call the local API |
Need a hand?
Find a fix in troubleshooting, or share a reproducible issue with the community.
Get support ↗Improve this guide on GitHub ↗