The Gate Log
Every verdict, flag, and refusal, dated and never quietly edited.
-
Gatehouse now ships as an MCP server
The gate is now a tool an agent can install: npx -y gatehouse-mcp. Three read-only tools put it in the agent's own tool list. check_tool answers 'is this safe to install' with the vetting verdict, get_tool returns the full audit write-up, and search_tools finds tools ranked vetted first. It reads the same live registry the site does, so verdicts and liveness stay current. We ran our own audit on it, since a registry that flags others for phoning home cannot phone home itself. Three tools, all read-only, no writes and no deletes. Its only network call is one fetch to agentgatehouse.com for the registry, so it sends no telemetry about you or your queries. Always-on cost is 525 tokens per turn, against roughly 16k for a 44-tool server like posthog-mcp. One runtime dependency. It clears its own gate.
-
cc-skills vetted, clean and worth cherry-picking
Samuel Berthe's 20 skills: about 4,000 always-on tokens, no scripts, no telemetry, MIT. The catch is the default install loads all 20, and a few triggers (deep-research, frontend-design-deslop) are greedy. Install by name with --skill.
-
claude-code-subagents flagged for over-broad tool grants
Seven subagents with useful prompts, flagged on packaging. Three inherit all tools (Bash and web) with no tools: line; security-auditor gets Write but no Read; project-task-planner is handed Bash and Write against its own 'never action tasks' rule. One commit from 2025, deprecated tool names. Copy the ones you review and pin each to least privilege.
-
kostja marketing-skills flagged, heavy and self-linking
172 well-researched marketing skills, flagged on two counts: about 17,000 always-on tokens at full install (ten times Anthropic's pack), and ten skill bodies embed the author's own alignify.co as a reference so it surfaces in your output. Clean source, MIT. Install the few you need with --skill.
-
marketing-skills vetted, clean but heavy
Corey Haines's 47 marketing skills are well-built and clean of injected self-promo, but the full pack is about 8,000 always-on tokens (five times Anthropic's pack), and it bundles 64 opt-in CLI scripts that read vendor API keys. Install the few you need with --skill.
-
posthog-mcp vetted, passed
The official PostHog MCP works and is honest about what it does, but it is not read-only: 18 of its 44 tools write and 5 delete production objects. The safety is real and server-enforced through the key's scopes, so mint a read-scoped personal API key unless you want an agent editing live feature flags.
-
vercel-agent-skills vetted, with one deploy caveat
Vercel's nine first-party skills, carefully built: about 850 always-on tokens, secrets redacted, execFile not shell, no telemetry. The one thing to know: deploy-to-vercel tars your project and uploads it to an unauthenticated Vercel endpoint for a publicly claimable preview. It strips .env first and asks before pushing.
-
writing-humanizer vetted, but install it by name
One clean skill that strips AI tells from writing, about 145 always-on tokens. It ships inside a 68-skill collection that costs 32,000 tokens at --all -g, most of it auto-generated and untested. Vet the skill, not the dump: --skill writing-humanizer.
-
context7 vetted, with a disclosure
Works as advertised for fresh docs, but it's a thin client for a hosted service, so your query text goes to context7.com. Expected, disclosed, and now stated on the listing instead of buried in the source.
-
playwright-mcp vetted, safe by default
A clean pass, and a rare one. Filesystem access is locked to your workspace and file:// is blocked unless you opt out. The caveat is what it is, not a flaw: it drives a real browser, so treat it accordingly.
-
sentry-mcp vetted, mind the data
Official and well-built, with a sensible device-code login. The thing to be deliberate about: it pulls production error data, stack traces and PII included, into the model's context, and it ships its own telemetry DSN you can override.
-
serena vetted, capable and it edits
Language-server-backed edits that are precise and token-cheap on big repos. Remember it edits your code and writes memory files into the project; analytics stay local, and the one outbound call is a benign news fetch.
-
Two skill packs vetted, and a way to measure them
A new method: skills are judged by what they cost you on every turn. superpowers (14 skills, ~577 always-on tokens) and anthropic-skills (17 skills, ~1,629) both passed, both lean and well-scoped. The measurement script now lives in the repo.
-
wshobson/agents, measured but not yet vetted
The catalog is big: 88 plugins, 199 agents, 106 commands, and about 25.6k always-on tokens if you install all of it. Useful per agent, but a heavy context tax at full install, so pick the plugins you need. A behavioral audit of the agents is still pending.
-
postgres-mcp vetted, passed
Restricted mode is genuinely read-only, held by two independent enforcement layers (pglast AST checks plus a real read-only transaction), and 66 of its own tests pass. One catch worth knowing: it isn't safe by default, so pin --access-mode restricted.
-
The registry is open
First ten tools in intake: seven MCP servers, a plugin, a skill pack, and a subagent catalog. The vetting method is public. Install it, run it, audit what it injects.