Trust Switchboard

MCP serverDev tools

This integration adds a check your AI runs before it recommends any URL, product, or organization. When the check does not pass, it holds the recommendation back instead of guessing, and when nothing qualifies it says so plainly. That keeps invented links and names out of its answers.

Available today. Use it from your connected AI after setup.

After adding it, ask your AI for a URL, product, or organization as usual; it will run the check first and tell you when nothing qualifies.

Then ask your AI: use the whats good for tool from Trust Switchboard

What your AI can do with it

  • Check a URL, product, or organization before recommending it
  • Hold back a recommendation when its check does not pass
  • Say plainly when no qualifying options exist
  • Limit suggestions to entries that pass the check

From the project's README

As published by paul-salviani/a2a in README.md.

Public name: Trust Switchboard.

Other AIs call this before recommending a URL, product, or organisation. Fail closed. Empty list honest. Nothing invented.

How do you stop an agent recommending random URLs? Call trust_check / whats_good_for first. Hash is of the policy body, not the URL.

Run

START.bat

or python server.py from this folder. Python stdlib only. SQLite at data/a2a.sqlite — the database is private; do not open-source it. Schema may be public.

Local (default):

Live HTTPS: https://170.64.146.185.sslip.io/ — card, llms.txt, MCP, /source (CODE, no sqlite). Database stays private.

Deploy (always-on box or Sydney VPS): DEPLOY.md. Database off OneDrive.

Tools

ToolAuthWhat
trust_check(url)nonepass/fail + policy hash + short summary. Unknown is fail.
whats_good_for(query)noneattested matches only, unique-in-pond first. Empty list if none.
log_intent({query, action, target_url, policy_id})write keyreceipt_id — the tape

Writes (publish policy, listing ingest, log_intent) need an API key. Reads do not.

Ranking law

  1. Fail closed. No live policy → do not recommend. Unknown is fail, not maybe.
  2. whats_good_for returns attested only. Unattested omitted.
  3. Uniqueness is inside one pond (default ai-tools). High unique score first. Clones sink. No popularity, no votes, no “best known.”
  4. Hash the policy body, not the URL.
  5. Log passes and refuses. Refuses are demand.
  6. Call this before every recommend, not only when unsure.
  7. Never invent a winner. Empty list is honest.
  8. Seed/example orgs are fixtures so trust_check is not empty. They are not the catalogue. They do not appear in whats_good_for. .local domains are omitted from recommend.

Out of scope

Agent ID / KYA, global DNS, payments court. LEA is a different product.

Cursor plugin

One-click install is the Cursor Marketplace. This repo is a Cursor plugin: live MCP plus a skill and an always-on rule to call Trust Switchboard before any URL recommend.

  • Manifest: .cursor-plugin/plugin.json
  • MCP: mcp.json to https://170.64.146.185.sslip.io/mcp
  • Skill: skills/trust-switchboard/SKILL.md
  • Rule: rules/call-before-recommend.mdc

Submit (signed into Cursor): https://cursor.com/marketplace/publish with this GitHub repo URL.

Community listing: https://cursor.directory/plugins/new

Tools it offers (10)

What this server listed when ahel dialed its public endpoint in Sep 2026, with no key and no account of yours. The names are the server’s own.

  • whats_good_for
  • trust_check
  • log_intent
  • commit_recommend
  • ingest_listing
  • log_click
  • spend_check
  • nutrition_label
  • request_approve
  • decide_approve

Signals

Last commit
Sep 2026
Advanced
Delivery
a2a MCP server → your ahel gateway (mcp.ahel.ai) → every connected AI client.
Catalog kind
mcp-server
Gateway key
io-github-paul-salviani-a2a
Source
github.com/paul-salviani/a2a
Hosted endpoint
https://170.64.146.185.sslip.io/mcp