Developer Walkthrough & Skill Management
A complete, step-by-step developer guide for initializing a new Claude Code / Clembot build, recommending pre-vetted MCP tools from actual prompt history, and managing existing skill rosters so you never spend money or tokens on redundant capabilities.
Step 1: Fresh Init Step 2: Recommendations Step 3: Verification
Click canvas or button to inject a rogue tool and watch the scanner cage it
Step 1 · Fresh Init & Baseline Inspection
doorman doctor
When you run claude init in an empty repository, Claude Code sets up project scaffolding
(such as CLAUDE.md and .claude/settings.json). Before installing third-party tools,
you need to know exactly what an agent in that directory can currently reach.
Zero runtime dependencies. Local, read-only, and costs no tokens.
# Global CLI Install via npm
npm install -g clembot-doorman
# Or via Claude Code Marketplace:
claude plugin marketplace add clemenswan/clembot-doorman && claude plugin install clembot-doorman
# 1. Open Local Visual HTML Dashboard (zero servers, zero ports, zero cloud dependencies)
doorman dashboard
# 2. Instant Terminal Posture Doctor
doorman doctor
# 3. Reverse-Mine Prompt Needs from team session history
doorman needs
# 4. Unified Audit & Grading Sweep
doorman audit
# 5. Schedule Automated Fleet Audits (GitHub Actions, crontab, /schedule)
doorman scheduleVerbatim terminal output on a fresh Claude Code build:
# Your build
`C:\Users\username\my-new-project`
Read-only. Nothing here was executed, sent anywhere, or billed.
## Harness
- **Claude Code** (CLAUDE.md)
## MCP servers this project can reach
None declared. Nothing to grade yet, and nothing to gate.
## Agents
No .claude/agents/ directory.
## The gate
**not installed**
- hook present: no
- wired in settings.json: no
- registry present: no
---
_doorman doctor. Static, local, free. It reports what is here; it does not
say whether any of it works. That is doorman report and doorman eval._
Run doorman install to wire the deterministic offline security hook.
This copies mcp-gate.sh into your project, creates registry/allowlist.json,
and registers the hook in .claude/settings.json. Any unapproved tool call or injected payload
is blocked locally at exit 2 before the agent's context window can be compromised.
Step 2 · Prompt-Driven Recommender
Once you work in your project for a few sessions, don't guess which MCPs or skills to install.
doorman needs reads your local transcript history from ~/.claude/projects/,
extracts human prompts, maps them across 12 capability taxonomies, and cross-references them against
the free public scorecard feed.
Docs lookup, cloud deploy, live web search, browser automation, database queries, observability, payments, comms, and design.
Transcripts are parsed and matched locally on your machine. The only network call is an anonymous GET for public grades.
UNMET with vetted candidates, blocked with hard-fail reasons, or GAP when the ecosystem lacks coverage.
doorman needs — 42 prompts read from this build’s own history
history: ~/.claude/projects/C--my-new-project
feed: 24 graded rows
UNMET Current documentation for a library it does not know
14 prompts across 3 sessions · matched "latest docs", "deepwiki"
> …can you look up the latest docs for drizzle orm…
> …check the documentation for cloudflare workers assets…
worth-measuring DeepWiki MCP [A (85.7)] matched "docs"
https://mcp.deepwiki.com/mcp
worth-measuring Cloudflare Docs MCP [A (94.2)] matched "documentation"
https://docs.mcp.cloudflare.com/mcp
UNMET Driving a real browser
6 prompts across 2 sessions · matched "screenshot the page", "playwright"
> …take a screenshot of the landing page at 390px…
worth-measuring Peek Browser MCP [C (66.7)] matched "screenshot"
https://mcp.peek.com
UNMET Reading the live web
5 prompts across 1 sessions · matched "search online"
> …search online for the error code…
blocked WebZum Search [F (49.0)] matched "search"
hard fail: injection-shaped content in tool:host_site.description
https://webzum.com/api/mcp
GAP Production errors and logs
3 prompts across 1 sessions · matched "tail the logs"
> …tail the logs from production…
nothing graded covers this. The feed has the gap, not your build.
3 unmet, 1 of them with nothing graded to offer.
What this is: your own prompts, counted, against capability text those
candidates published about themselves. Nothing here was driven, so
nothing here is a claim that a server works. `worth-measuring` means
exactly that — run `doorman eval` with your key and find out.Step 3 · Roster Management
If you already have a rich collection of skills (e.g., in .claude/skills/ or .agents/skills/),
the most important job of a package manager is saying no to redundant tools.
coveredBy(inv) in Needs
When you run doorman needs, it checks your active MCP servers and tools held by subagents.
If a capability is already covered, it prints COVERED · already covered by: [tool]
and suppresses candidate recommendations so your terminal is never cluttered with tools you already have.
already-installed in Watch
doorman watch continuously monitors the public feed. Any server your project already runs
is tagged already-installed. If an installed server is downgraded or caught with hostile prompt injection
in a newer audit, watch alerts you to revoke it from your allowlist.
Doorman reads only the YAML frontmatter (name, description) of your skills,
never the full markdown bodies. In our production vault, reading full bodies was 642 KB; reading frontmatter
was 30 KB. Only frontmatter fits safely in a single model turn without blowing out context.
Before spending any tokens or paying a scorecard, node scripts/vet.mjs <target> --dry-run
runs the Fit Review. If an existing skill already covers the task, the verdict is REDUNDANT,
cites your exact skill name, and halts at Phase 1 ($0.00 spent).
candidate https://github.com/example/git-mcp
inventory C:\Users\username\my-project — 3 agents, 22 skills, 1 mcp servers
FIT REDUNDANT
Already covered by existing skill: git-pr covers reviewing and merging pull requests.
Already covered by:
- skill git-pr: handles GitHub pull requests and diff review locally
STOPPED before the paid grade. $0.00 spent.
Nothing was sent to the scorecard, and no client was built.Step 4 · Agent Integration
Yes. Doorman is engineered to run seamlessly inside your agent loop rather than requiring you to juggle a separate terminal window.
/vet Slash Command
In Claude Code, drop a command definition into .claude/commands/vet.md that calls
node scripts/vet.mjs $ARG --dry-run. When you encounter a tool link in chat,
type /vet <url>: your agent tests fit against your existing skills and reports
the verdict in-line.
Isolate tool adoption from your coding session. The doorman subagent is provisioned with
read-only tools and access to the Bazantic Scorecard gateway. It audits unknown tools in a sandbox
and never exposes your primary codebase to unallowlisted code.
Equip any agent harness with the Doorman skill. When the agent is tasked with adding a capability
(like searching the web or inspecting database schemas), it triggers doorman needs or
doorman report to evaluate options before asking you to install them.
Configure continuous tool governance without manual overhead. Run doorman schedule --github
to install a weekly GitHub Action that audits project posture and tool recommendations, or trigger
periodic scans via your agent harness or crontab.
Step 5 · What the Doorman Checks
Two do not always run. A probe that did not apply is recorded with its reason and left out of the average, rather than counted as a failure the server never earned. A seventh step is not a probe at all: it re-runs the cold open with the recipe, to find out whether the recipe works.
Connect, list tools, lint the schemas. A server with no tools exits here without spending anything on probes.
Is there anything here at all?
A fresh agent, the server's own descriptions, and a task built from what the subagent actually needed. Scored on first-try tool choice, first-try schema validity, completion, and how much wandering it took.
Can an agent that has never seen this succeed on the first try?
Fires only when two tool descriptions genuinely overlap. The task is worded in the wrong tool's vocabulary, so only a description that draws a real line will save it.
Do these two descriptions actually distinguish themselves?
Omit a required parameter on purpose, then measure the error. What is graded is the message, not the agent. "Invalid input" and "missing required repoName (string, e.g. facebook/react)" are the same failure and completely different products.
Can an agent fix itself from this error in two turns?
Two steps, where the second must consume a real value the first returned. Servers that look fine tool by tool fall over here. Skipped under three tools.
Do these outputs compose, or only look like they should?
Scan-only. Reads every description an agent sees before choosing a tool, looking for content shaped like an instruction rather than documentation. It never calls a tool: we do not execute a server to find out whether it is hostile.
Is this documentation, or is it giving my agent orders?
The same cold task again, with the drafted recipe in the agent's system prompt. Scored as recovered headroom, so a server that was already strong is not punished for having little room to improve. It is kept out of the behavioural average: folding it in would pay a server twice for one recovery.
We told it exactly what went wrong. Did that fix it?
What triggers an automatic F? Two findings cap a grade at F outright: injection-shaped content in advertised strings, and a non-TLS transport. A hard fail immediately overrides high static scores.
Are unmeasured layers counted as zero? No. A layer that did not run has its weight removed, renormalising remaining layers so honest partial audits are not penalized.
What does the guidance delta prove? It measures recovered headroom when the agent receives the drafted recipe. A server that recovers can be safely adopted behind a recipe; a server that still fails cannot be saved by documentation.
Step 6 · Production Status
Here is the honest status of what is tested and running vs. what requires configuration:
| Component | Status | Test Suite & Live Proof |
|---|---|---|
Static Scannerdoorman report |
Production Ready | Live SSE handshake, tool schema validation, 6-pattern injection detection. Caught WebZum 6,290-char injection in real-time. |
Needs Recommenderdoorman needs |
Production Ready | Parses real JSONL transcripts, deduplicates resumed sessions, matches 12 capability taxonomies against live feed. |
Harness Doctordoorman doctor |
Production Ready | Zero-dependency scanner for Claude Code, Cursor, Windsurf, Copilot, Gemini. Audits agent-to-tool exposure ratios. |
Security Gatemcp-gate.sh |
Production Ready | 180-line offline Bash hook. Passed 29/29 tests in test-gate.sh. Blocks rogue tools and enforces 5s timeout. |
| Fit Reviewfit-review.mjs | Production Ready | Compares candidates against skill frontmatter. Enforces temperature 0, strict JSON schema, and hallucination rejection. |
| Budget Ledgerbudget.mjs | Production Ready | Enforces per-run ($1) and per-day ($5) caps in USDC on Base; auto-releases unspent reserves on error. |
| Bazantic Gatewayx402 & MCP | Live in Production | clembot-doorman.bazgateway.com returns live x402 challenge ($0.01 USDC on Base) + MCP SSE tool stream. |
| Multi-Directory Ingest.agents/skills | Configuration Detail | Reads .claude/skills/*/SKILL.md by default. Configure DOORMAN_INVENTORY_ROOT for alternate directory layouts. |
Developer Q&A
Yes. You can install it directly as a Claude Code Marketplace plugin (claude plugin install clembot-doorman) or invoke it on demand using npx clembot-doorman audit. In both cases, zero global permissions or dependencies are needed.
doorman audit aggregates L0 doctor (wiring & inventory inspection), L0.5 prompt needs (capability gap extraction), and L2 feed watch into a single executive report. Pass --out .doorman/audit.md to save a markdown summary or --json for machine-readable output.
Doorman's local budget.mjs checks and reserves funds against your configured caps ($1.00 per run, $5.00 per day) in an offline append-only ledger (.claude/registry/ledger.jsonl) before any HTTP request is dispatched. If the network call fails, the reservation is immediately released.