Most agents preload every tool definition into the prompt. Desire Path learns which tools tasks actually use and finds them by walking a graph, so prompts stay small and the agent gets sharper with use.
Try it liveSee the proofPark designers know the pattern: people ignore the paved route and wear a dirt track along the shortcut they actually want. These are called desire paths, and some campuses now wait to see where people walk before laying any stone.
Agent tooling is still paved in advance, with every tool listed in every prompt. We do it the other way round. Let the agent walk, record which tools real tasks needed, and make those trails the map.
Agent tasks per day
Price per 1M input tokens ($), edit to match your model
Assumes each call would otherwise carry the whole tool catalog.
Gemini agents using real MCP tools, success checked by code. Ten held-out tasks, one run each. The graph row is from a run where 100 extra synthetic tools were also on offer; the same run on the real 42 tools also passed 10 of 10, at 3,241 prompt tokens.
Desire Path ships an MCP server. Your agent asks which tools a task needs and loads only those. Setup in the docs.
Successful runs become edges in HydraDB: task words point to the tools that were needed, and tools point to the ones that usually come next.
A new task walks those edges in milliseconds. No model call, and only the matching tool definitions go into the prompt.
Every confirmed run strengthens the trail. Teach it in the app and watch the next routing change.
Recall of the tools the agent really needed on held-out tasks, as it learns more traces. Loaded live from the saved results.
Proposed early-access plans. We are not charging yet.
$49 / month (proposed)
Hosted graph, usage dashboard, shared learning across your agents.
Request early accessA routing layer for AI agents. It learns which tools tasks really use and finds them with a graph walk in HydraDB, so your agent loads a handful of tool definitions instead of all of them.
No. Routing is graph traversal in HydraDB with no model call. The only model calls are in your own agent.
In real Gemini agent runs on 10 held-out tasks, the graph with 5 tools passed 10 of 10 using 3,241 prompt tokens per task, against 29,352 when loading all 42 tools (also 10 of 10). Keyword search with 5 tools passed 4 of 10. That is a small test, so treat it as early evidence.
Yes. Use Try the demo on the sign-in page for a private, temporary workspace. No email needed.
Yes. Lessons you teach inside your workspace only change your own routing. The public app page is a shared demo, so lessons taught there are visible to everyone.
Your email, a password hash (never the password itself) and your lessons. We run no analytics. Cookies: one login session cookie, plus your own display settings.
Not yet. The demo uses a fixed catalog from four MCP servers. Custom catalogs are the next thing we plan to build.
It is an early demo on a free host. The first visit after a quiet spell can take about a minute to wake, and there is no uptime guarantee. Prices shown are proposed.
Yes. The code is open source and uses the open-source HydraDB engine.