Route, teach, measure
Checking your session...
Tap an example, check which tools the graph picks, then teach it a correction and route again to see it change.
Use it from your own agent
Or as an MCP server for Claude Code or Claude Desktop:
claude mcp add desire-path -- python3 /path/to/desire-path/scripts/dp_mcp.py
It calls the shared demo graph. Private workspaces are not reachable from it yet. Setup in the docs.
Describe a task. The graph picks the tools.
Were these the right tools?
Uses the task from the Route tab. Tap the tools it really needed, in order, then teach. Route again to see the change.
Picked: none
Everything you have taught your private graph.
Held-out tasks the graph never learned from. Compares the graph, the graph without sequence edges, and a keyword baseline.
100 requests each at 1, 5 and 20 concurrent. Every request walks the graph, no cache. One run at a time.
Use Desire Path from your own agent. Both ways call this site.
Three tools: route_tools, list_catalog, and teach_tools (off by default, because the demo graph is shared). Get scripts/dp_mcp.py from the repo, then run:
claude mcp add desire-path -- python3 /path/to/desire-path/scripts/dp_mcp.py
Claude Desktop config:
{"mcpServers": {"desire-path": {"command": "python3", "args": ["/path/to/desire-path/scripts/dp_mcp.py"]}}}Limits: tested from the command line, not yet verified inside Claude Desktop or Claude Code. A client still loads its other tools as usual; the saving comes when the agent calls route_tools first. Both ways use the shared demo graph, so private workspaces are not reachable yet. Setup in the README.
Cookie choice: