the tool
Point it at your
own build
Everything else on this site is a record of one harness. Doorman is the part of it built to be installed by somebody who does not have this vault, and it works on a build that looks nothing like this one.
What it does today
Built and working
- Reads the prompts you have already typed and names what this build keeps reaching for and does not have. Your own sentences are the evidence, counted, rather than a guess about your intentions.
- Grades an MCP server before you add it, and puts the scanned text on a public transcript so the evidence for an accusation does not sit behind the accuser's token.
- Refuses a server that has not been graded, at the point of use. The hook is offline and dependency-free: no network call, no jq, no node.
- Drafts usage rules for a mediocre server, so a C grade becomes workable instead of only labelled.
- Leaves your trust list alone. The installer reports that it kept your registry rather than replacing it, and a plugin update cannot reach it.
Not yet
- Ranking what it finds by what it has already cost you. Specified and queued, not built. It currently reports needs without pricing them.
- A grade for a skill or a setup. The scorecard grades servers, because a skill has no tools to drive, so those are pointers and never scores.
- Anything across more than one machine. There is no account, no telemetry and no fleet view, and nothing about your build leaves it.
- A verdict you should trust on faith. A match reads
worth-measuring, neverfits, and a need with no candidate is reported as a gap rather than quietly dropped.
The arguments for any of this are checkable rather than claimed:138 commits and 169 evidence records areon the proof page, 6 shipped projects are linked from the overview, the rules arewritten down with the failures that caused them, and the approval gate is something you can click.
Install it
A Claude Code plugin. The static grade costs nothing and needs no key; the behavioural layer needs your own model key, and an unmeasured layer is reported as unmeasured rather than as a zero.
Looking to work together rather than to install something? That is a studio conversation, and it happens at wanessalabs.com ↗. Clembot itself is not for sale: it is how the work gets done, published so you can audit the method before trusting the output.