404.directory: one-call official docs search and a read-only Agent trust gateway #1564
Replies: 2 comments
|
v0.8.1 adds an installable cross-client Agent Skill so Agents can move from registry discovery to a useful first call without hand-written prompting. Install in Codex, Claude Code, Cursor, Cline, and other Agent Skills clients: npx skills add MM-sheng/404-directory --skill use-404-directory -g -ySkill page: https://www.skills.sh/mm-sheng/404-directory/use-404-directory The Skill routes official technical questions to Official Registry v0.8.1: https://registry.modelcontextprotocol.io/v0.1/servers/io.github.MM-sheng%2F404-directory/versions/latest |
|
Update: v0.9.0 now turns registry discovery into an actual Agent activation path.
The same Registry server remains packaged as a portable Agent Plugins 1.0 plugin plus a lightweight Claude marketplace package. The package uses a dependency-free local bridge to generate and preserve one random non-personal ID per plugin installation. The raw ID stays local; 404.directory stores only an HMAC digest after a successful tool execution. Prompts, arguments, results, probes, anonymous calls, self-tests, and duplicates are excluded from the public adoption target. Claude Code direct install: GitHub Copilot's external-plugin intake passed its automated gates and is awaiting maintainer review: github/awesome-copilot#2767 Release: https://github.com/MM-sheng/404-directory/releases/tag/v0.9.0 |
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
404.directory v0.8.0 is live in the Official MCP Registry as
io.github.MM-sheng/404-directory.The problem it targets is tool selection before execution: an Agent should be able to find a useful tool, inspect its trust evidence, and call only an approved read-only operation without wiring a different server for every task.
The public remote MCP endpoint currently exposes 12 tools with no account or API key. The lowest-friction entry point is
search_official_docs, which searches current first-party documentation from OpenAI, Microsoft Learn, AWS, and Cloudflare in one call and returns source URLs.A useful first prompt is:
For broader tool discovery, the Agent can use:
search_toolsandrecommend_toolsfor capability discoveryget_trust_scoreandcompare_toolsbefore selectioninspect_tool_serverto retrieve the current approved schemas for a curated remote MCP serverinvoke_registered_toolto execute one approved read-only operationSafety boundaries are deliberately narrow: registered remote calls are allowlisted and read-only, remote descriptions/results are treated as untrusted data, and callers are warned not to send credentials, private code, personal data, or secrets.
Install and verify:
https://404.directory/mcpio.github.MM-sheng/404-directoryEach generated install configuration includes a fresh, random, non-personal
X-404-Agent-ID. The raw value is never persisted; only an irreversible HMAC digest is stored after a successful tool execution. Prompts, tool arguments, results, and raw identifiers are excluded from product analytics. The de-duplicated public adoption metric is visible at https://404.directory/v1/metrics/agents.I would especially value feedback on two questions: whether
search_official_docsis specific enough for reliable automatic tool selection, and what additional evidence an Agent should require before invoking a third-party read-only tool through a trust gateway.All reactions