Inspiration
The open agentic web lets any website publish an agent network that other sites can call. That only works if sites can find each other and trust what they're linking to. Before my agent calls another site, I want to know: does it follow the spec, which origins will its agents contact, does it leak keys or carry instructions that try to hijack my visitor's model, and can my site even load it? Today you'd have to read raw JSON by hand. The spec's own "future work" section asks for registries, search, reputation and conformance tests. Agent Atlas is a first version of all four.
What it does
- Checker: paste any agentic site's URL. Atlas fetches its agent-manifest.json, network definition and links.json, and tests them against the Worldwide Agentic Web MVP spec (Profile 0.2), using the reference runtime's own validator and HOCON parser. So a site that passes Atlas also loads in the real runtime.
- Network chart: every agent, partner site, MCP server and site action the network can reach, drawn as a small nautical-style chart.
- Trust report: the peer policy, every external origin its agents contact (and whether allowed_peers permits it), leaked credentials, and prompt-injection phrases in published instructions, which matter because those instructions run inside the visitor's model when another site calls them.
- Callable or not: Atlas reports separately whether other sites can actually load the network (CORS). A site can follow the spec and work for its own visitors, yet be impossible to call from anywhere else. Atlas makes that visible.
- Live directory: every site in the challenge's community list, checked automatically and searchable.
- Atlas is agentic itself: its own browser assistant (bring your own key) searches the directory, checks sites and opens reports on the page. It also runs an MCP server (search_directory, check_site, list_sites), so any other agentic site can ask Atlas "find me a site that does X" or "is this site safe to call?" before linking. There's also an embeddable badge with a live result.
How we built it
- The checker is plain JavaScript that runs in Vercel serverless functions, with a guard that refuses private and internal addresses (SSRF protection).
- Validation reuses the challenge's reference runtime (validate_network, classify_tool, the HOCON parser), so Atlas and the runtime agree.
- The MCP server is a stateless JSON-RPC handler over Streamable HTTP with open CORS, so agents running in any visitor's browser can call it.
- The assistant uses the challenge's browser runtime unchanged, with Atlas's MCP server and three page actions (get_page_state, show_report, filter_directory). None of them can do anything irreversible.
- No server-side AI: visitors bring their own key and Atlas never sees it.
- 14 automated tests cover the checker, MCP server, directory parsing and the address guard.
Challenges we ran into
- Being fair about CORS. My first checker labelled the organisers' own site "not conformant" because its files send no CORS header. Re-reading the spec, a same-origin site does conform; it just can't be called by others. I split CORS into its own "callable" signal so the verdict is accurate.
- Detecting prompt injection without an AI model. Atlas uses conservative patterns and labels them as heuristics, so a clean report never claims a site is "safe".
- Reading untrusted text with an agent. Atlas's assistant reads other sites' descriptions and instructions, so its instructions treat all tool results as data and never follow instructions found inside them.
Accomplishments that we're proud of
- Atlas is a working conformance tool that the ecosystem can use today, and it checks itself: conformant, 100/100, callable.
- One checker, four ways to use it: the web page, the assistant, the MCP server and the badge.
- It surfaced a real issue on the first site it checked.
What we learned
Interoperability problems hide in small details like a missing header or an origin not listed in allowed_peers. A shared checker catches them before they break someone's agent.
What's next
- Live calls to each site's network and MCP servers, not just static checks
- Reputation over time: check history, uptime and change alerts
- Signed manifests and verified publishers
- A GitHub Action so sites run Atlas checks before they deploy
Built With
- javascript
- json-rpc
- mcp
- node.js
- open-agentic-web
- otari
- vercel