Trust Switchboard
Are you on Trust Switch?
One free call before your agent deals with another agent, vendor, or URL. It asks the question, records the answer, and lets business proceed. Nobody is blocked. The answer is the data.
| You are about to… | Call | What happens |
|---|
| deal with, wire, or talk to another agent or vendor | handshake(url) | Asks "Are you on Trust Switch?" Records yes/no. Business proceeds (activation is off). Never blocks. No key. |
| recommend a URL, product, or organisation to a human | trust_check(url) / whats_good_for(query) | Fails closed. No live policy → do not recommend. Empty list is honest. Nothing invented. No key. |
Most integrators only ever need the first row.
Handshake in one call
POST https://170.64.146.185.sslip.io/tools/handshake
{"url": "https://api.example.com", "asker": "my-agent", "context": "about to call their API"}
{
"question": "Are you on Trust Switch?",
"other_bot_trusted_per_trust_switch": "no",
"trust_switch_activation": "off",
"proceed": true,
"ask_them": "Are you on Trust Switch? If not: publish a policy at https://170.64.146.185.sslip.io and other AIs can verify you before they recommend you. Free. No account to read.",
"handshake_id": "hs_…"
}
Put ask_them in your message to the other agent. Send its reply back as their_answer. GET /handshakes shows who is being asked about most — that list is demand.
activation is off by default: proceed is always true. Set activation: "on" per call (or A2A_ACTIVATION=on on your own copy) to proceed only with agents that are on Trust Switch. Do not turn it on until enough people are on.
Drop-in for products
client/trust_switch_client.py (Python, stdlib) and client/trust_switch_client.mjs (Node ≥ 18). One file each. Never raise. Never block.
from trust_switch_client import note
note("https://api.retellai.com", asker="my-product", context="retell client")
note() runs in the background, 3 s timeout, one ask per vendor per hour, result appended to ~/.trust_switch/<asker>.jsonl. Off under pytest / unittest / node --test and when TRUST_SWITCH=off. Nothing in your product branches on it.
hs = handshake("https://drive.google.com", asker="my-product")
hs["proceed"]
MCP
Tools: handshake, trust_check, whats_good_for, nutrition_label (free); log_intent, commit_recommend, spend_check, request_approve, ingest_listing (write key). Streamable HTTP at /mcp. mcp.json in this repo is the client stanza.
For the org being asked about
Publish a policy (name, site, what you store, region, dispute email) → policy_id + hash of the policy body. Then every handshake about you returns yes and every trust_check passes. POST /orgs to sign up, POST /orgs/policy with your key. Free.
Ranking law (do not "optimise" this into votes)
- Handshake never blocks while activation is off. The no is the data.
- Recommending to a human fails closed. Unknown is fail, not maybe.
whats_good_for returns attested orgs only, unique-in-pond first. Clones sink. No popularity, no votes.- Hash the policy body, not the URL.
- Log passes and refuses. Refuses are demand.
- Seed/example orgs are fixtures. They never appear in
whats_good_for and handshake says no for them.
- Never invent a winner. Empty list is honest.
Run your own
Python stdlib only. SQLite at data/a2a.sqlite — the database is private; do not open-source it. Schema may be public. deploy/ has the systemd unit, Caddyfile, and push script. Database off OneDrive.
First run: copy seed.example.json seed.json and put a key from python deploy/new_write_key.py in it (seed.json is gitignored).
Tests: python tests/test_switchboard.py, python tests/test_46.py, python tests/test_handshake.py.
What this is not
Not a directory. Not an agent ID system. Not a payments court. Not a marketing site. The code is copyable — this repo is it. The log is the company.