Reference benchmark replay and recovery simulation. No live database is killed. Current product results: 700 mean tokens over seven fixture tasks.
Fewer tokens.
Right order.
Recorded recovery demonstration, not a promise for this implementation.
tokens to hand an agent all 420 tools from 20 MCP servers, before it reads a word of the task.
Counted with the cl100k tokenizer over the real tool schemas.
Top-left wins.
Each dot is a way of picking tools. Further left means fewer tokens in context. Higher means more of the right tools, in an order that runs.
Reference replay only. Current implementation is measured on seven local fixture tasks; no live latency is being reported here.
Reference comparison data from the original demo. These values are not current HydraDB measurements.
| Method | Chain recall | Order | Disambiguation | Calls saved | Round trips | Tokens |
|---|---|---|---|---|---|---|
| preload-all | 100.0% | 0.0% | 20.0% | 0.00 | 0.0 | 110,238 |
| BM25 top-5 | 31.0% | 0.0% | 0.0% | 0.00 | 1.0 | 1,430 |
| BM25 top-10 | 62.0% | 0.0% | 20.0% | 0.00 | 1.0 | 2,471 |
| Switchboard | 100.0% | 100.0% | 100.0% | 1.00 | 1.0 | 469 |
No current latency data is connected. The grid and kill flow below are an illustrative replay, not a live HydraDB benchmark.

Pull the plug. Watch the line.
Reference replay only. Current implementation is measured on seven local fixture tasks; no live latency is being reported here.
Reference replay only. Current implementation is measured on seven local fixture tasks; no live latency is being reported here.
Recorded run, 140 plans: p50 269 ms, p95 295 ms, p99 307 ms.
Kill the database
Local demo only.
- Last recorded kill: plans worked again after 3.3 s, full recovery took 41 s, and the graph came back with the same counts.
What's in the graph.
Every tool comes from a real server's
tools/list
response.
