Free tools Windows power users keep installed
One-click scans. No signup required.
iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Mycelium is an open-source routing layer that maps a natural-language request to an agent endpoint using a local vector index, so the routing step does not need an LLM call. Its headline figures, including about 9.56 ms cold discovery and 70.7% Top-1 intent accuracy, are the project’s own measurements on a synthetic benchmark. They have not been independently reproduced in the sources available for this article, so treat them as claims to test rather than settled performance.
What Mycelium does
When an agent has dozens or hundreds of tools, it must decide which one fits a request such as “convert this invoice to French.” The usual approach sends the tool descriptions and the request to a language model, which adds latency and token cost to every routing decision. Mycelium proposes a different path: embed the request, search a local index of agent descriptions, and return the endpoint that best matches.
![Claude AI for Beginners Bible: [5 in 1] The Ultimate Guide to Automate Your Work, Save Hours Every Week, and Use AI for Real-World Results](https://m.media-amazon.com/images/I/415+fSJacsL._SL160_.jpg)
The project’s GitHub README describes the stack as a local ChromaDB vector store, all-MiniLM-L6-v2 sentence embeddings, and a FastAPI service. It presents itself as an open-source semantic registry and routing protocol for agentic workflows. SDKs are advertised for Python and JavaScript:
- Python:
pip install mycelium-agents - JavaScript:
npm install mycelium-js
These installation details come from project documentation. They confirm what the project says it ships, not that the packages have been independently audited or widely deployed.
#1 Best Overall
The reported numbers, with their conditions
The project’s September 27, 2026 announcement describes an evaluation on a synthetic corpus of 100,000 agents, using 441 task-oriented queries. Cold-cache latency was measured with embedding time included, on commodity CPU hardware. The README later publishes a separate benchmark table labelled v0.3.0. The two sources are not identical, and the table below keeps each figure tied to its origin.
| Metric | Reported value | Source | Conditions stated |
|---|---|---|---|
| Top-1 intent accuracy | 70.7% for Mycelium vs 40.4% for BM25 (30.3 percentage-point gain) | Announcement, September 27, 2026 | Synthetic corpus of 100,000 agents; 441 task-oriented queries |
| Cold discovery latency | 9.56 ms for Mycelium vs 194.0 ms for BM25 | Announcement, September 27, 2026 | Cold cache; embedding time included; commodity CPU |
| Two-hop chain, end to end | 37.6 ms (weather-to-translation) | Announcement, September 27, 2026 | Two hops; further hardware detail not stated |
| Three-hop native chain | 36.25 ms | README performance table, v0.3.0 | Three hops; hardware and load detail not stated |
| P95 latency | 11.4 ms | README performance table, v0.3.0 | Test conditions not stated |
| Throughput | 130+ requests per second, 0.0% errors under 100 concurrent workers | Announcement, September 27, 2026 | 100 concurrent workers; the README gives single-node throughput above 130 requests per second without a load profile |
Two points deserve care. First, the two chain figures are not the same test: one is a two-hop chain and the other a three-hop native chain, so they should not be compared as one improvement. Second, the BM25 baseline is the project’s comparison point. The accessible sources do not describe how that baseline was configured, so the size of the accuracy gap depends on a comparison the reader cannot yet inspect.
What is not established
The material available for this article supports a narrower conclusion than the headline suggests:
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →- Independent replication. No third-party reproduction of the accuracy or latency figures was found.
- Production behavior. A synthetic 100,000-agent corpus does not show how a real catalog with overlapping, poorly described, or frequently updated tools will behave.
- Methodological detail. The sources do not give enough detail on query construction, relevance labels, hardware specifications, or load profiles to rebuild the tests.
- Wrong-route and ambiguity handling. Top-1 accuracy says nothing about how often a confident wrong match is returned, or how the system behaves when no tool clearly fits.
Independent context: the LatentGate paper
The ACL Anthology records LatentGate: Low-Latency Semantic Routing via Frozen-Backbone Probing of Small Language Models, a 2026 ACL Industry Track paper by Shivam Ratnakar, Abhiroop Talasila, and Vinayak K Doifode. The paper uses a different method, probing a frozen small language model rather than relying on a sentence-embedding index. It reports 98.8% accuracy on in-domain natural queries and 80.0% on out-of-domain queries across 100 enterprise agents, with a runtime of about 28 ms on a T4 GPU.
The paper is also useful for a warning it raises: embedding-based routers can collapse semantically similar agents that do different things. That is a general failure mode of the approach Mycelium uses, and it is a reason to test with agents that look alike on the surface. The LatentGate results do not validate or rank Mycelium. The two systems were not evaluated head to head, and their benchmarks differ.
The MCP bridge and the human-on-the-loop guard
The announcement says Mycelium includes a bridge for Anthropic’s Model Context Protocol. It also describes a “Human-On-The-Loop” guard with a specific rule: read-only intents may execute automatically, while mutating intents are intercepted and held until a human provides cryptographic authorization.
This is a control design described by the project. The sources reviewed do not include a third-party security audit, a published threat model, formal verification, or independent testing of the guard. Whether the approval mechanism holds up under misuse, misclassified intents, or compromised clients is unknown from the available material.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
How to evaluate a fast semantic router before you rely on it
Speed matters only if the router picks the right endpoint and fails safely when it does not. A useful evaluation uses your own tools and traffic:
Best Value
- Build a held-out query set. Collect real requests that were not used to write tool descriptions, and record the correct endpoint for each one.
- Measure accuracy at the task level. Top-1 is a start, but also check whether the correct tool appears in the top three and whether the chosen tool completes the task.
- Measure latency on your hardware. Include query embedding, cold and warm cache runs, and your own network path. Report P95 and P99, not only averages.
- Test ambiguous and out-of-scope requests. Confirm that the router asks for clarification, returns nothing, or hands off to a fallback instead of forcing a match.
- Stress the catalog. Add near-duplicate tools and measure how accuracy changes as the number of overlapping tools grows.
- Check authorization and rollback. For any tool that writes data or moves money, confirm where approval happens, what is logged, and how a wrong action is reversed.
Adjacent options exist in the same category. StackOne published a vendor engineering article on May 12, 2026 describing semantic discovery for SaaS connector actions. It is a vendor’s own account and does not establish how that approach compares with Mycelium.
Mycelium fits best where tools are numerous, descriptions are reasonably distinct, and wrong actions are low-cost or can be gated. It fits poorly where the cost of a misrouted write is high and no independent test has been run.
Reader phrasing
The project itself uses the phrases “The Tool Routing Bottleneck” and “sub-10ms semantic tool routing,” with “No LLM overhead” as a tagline. These are the project’s own framing, and they describe the design goal rather than a proven result.
”
The Bottom Line
Mycelium is a credible design for fast, local tool routing, and its numbers are specific enough to test. Its published speed and accuracy figures, however, come from the project’s own synthetic benchmark and remain unreplicated. Run a held-out evaluation on your own tools, with ambiguity handling and gated write actions included, before you put it in front of anything that changes data or moves money.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
