ToolDNS: обнаружение инструментов для AI-агентов через DNS сокращает поиск на 95%

In July 2026, researchers published the ToolDNS framework on arXiv, which moves tool discovery for AI agents into the Domain Name System (DNS) and reduces the search space per query by 95.26% while keeping accuracy on par with the best solutions.

Why tool search got bolted onto DNS

ToolDNS addresses a scale problem: autonomous AI agents will soon need to navigate among millions of tools, while existing registries operate with O(N) complexity and require centralized management. The authors propose not building yet another fragile add-on layer, but using the most resilient infrastructure on the internet — DNS. A tool's functional purpose and an organization's trust level are encoded into a hierarchical namespace, turning expensive semantic search into a series of lightweight name resolutions with O(log N) complexity.

Key facts of the work:

• Benchmark of 33,688 real-world tools — four protocols: MCP, A2A, RESTful, and Skill
• Reduction of the search space per query — 95.26%
• Search complexity — O(log N) instead of O(N) for centralized registries
• Three protocol-compatible extensions: partially unfolded names, intent payload in EDNS0, and logical subdomains
• Transport — native UDP instead of HTTP

Three extensions on top of DNS

ToolDNS introduces three protocol-compatible extensions that provide decentralized management and semantic pruning. The first is partially unfolded names: part of the tool's hierarchy is revealed directly in the domain name, allowing irrelevant branches to be pruned without querying an external index. The second is an intent payload in EDNS0: the agent conveys its functional intent in the extended field of the DNS query. The third is logical subdomains encoding organizational trust: which organization a tool belongs to is already visible from the structure of the name.

What the benchmark showed

On the assembled dataset of 33,688 tools, ToolDNS reduced the search space per query by 95.26%, while retrieval accuracy stayed on par with state-of-the-art solutions. The dataset was deliberately assembled to be heterogeneous — drawn from the fragmented tool landscape and covering the MCP, A2A, RESTful, and Skill protocols; the authors released it in open access alongside the paper.

The key result is that the 95.26% reduction was achieved without any loss of accuracy: on retrieval metrics, ToolDNS keeps pace with today's best semantic search engines, even though it examines only a small fraction of the tool space.

How much faster is it

ToolDNS's UDP-based design cuts discovery latency by orders of magnitude compared to registries built on top of HTTP. Name resolution fits into lightweight UDP packets instead of full HTTP requests, so an agent doesn't need to establish a connection and wait for a response from a heavy registry for every single tool. It is the shift to native UDP — not an extra middleware layer — that the authors call the key to scalability.

"Scalable AI interoperability requires not more layers of middleware,

but smarter use of the infrastructure that's already under our feet," the paper's abstract on arXiv states.

What this means

ToolDNS shows that infrastructure for the AI agent ecosystem doesn't need to be invented from scratch — it can be built on top of DNS, a protocol that has scaled across the entire internet for decades. For AI agent developers, this means a potentially standard, decentralized way to publish and discover tools — without a single intermediary registry that would become a bottleneck and a point of control. If the approach catches on, finding the right tool among millions could become as cheap as an ordinary DNS query.

FAQ

What is ToolDNS?

ToolDNS is a framework proposed on arXiv in July 2026 that embeds semantic tool discovery into the Domain Name System (DNS). It turns the search across millions of tools from an O(N) complexity operation into a chain of name resolutions with O(log N) complexity.

Which protocols does the benchmark cover?

The ToolDNS benchmark consists of 33,688 real-world tools and covers four protocols: MCP, A2A, RESTful, and Skill. The dataset is published in open access alongside the paper.