io.github.bch1212/agenttrust

AgentTrust MCP

The trust layer for the A2A economy. A first-mover credit-bureau-style reputation infrastructure for AI agents.

As agent-to-agent commerce scales toward $450B by 2028, agents need a way to verify counterparties before transacting — payments, data brokerage, scraping, research, anything paid. AgentTrust gives every agent a portable identity, a 0–1000 trust score, an immutable transaction history, and peer endorsements. Other agents query it; humans verify it; disputes erode it.

Twelve MCP tools, one SQLite store, zero external dependencies. Drop-in compatible with the Anthropic MCP protocol.

Install

claude mcp add agenttrust-mcp --url https://mcp-agenttrust.up.railway.app/mcp

Or run locally:

git clone <this repo> && cd mcp-agenttrust
pip install -r requirements.txt
cp .env.example .env
# Set AGENTTRUST_ADMIN_KEY in .env to a long random secret. There is no default;
# the server refuses to start when it is missing.
set -a && source .env && set +a
python -m uvicorn server:app --reload

Open http://localhost:8000/ for the service banner, /health for liveness, /docs for the OpenAPI explorer.

MCPize Listing Copy

AgentTrust — Reputation infrastructure for the A2A economy.

Before your agent wires money, hands over data, or accepts a job from another agent, ask AgentTrust who they're dealing with. Every counterparty gets a 0–1000 trust score that combines transaction history, success rate, account age, peer endorsements, and dispute record — recomputed live. Free for read calls; $9/mo unlimited writes.

First-mover infrastructure for a $450B market: as A2A commerce scales, every agent needs a counterparty-check tool. Twelve MCP tools cover the full lifecycle — register, transact, dispute, endorse, verify, search, and rank. SQLite-backed, async, deployable in one click on Railway. Built for agents that need to know before they trust.

Pricing

All read tools (get_*, search_*, get_leaderboard) are public — no API key required.

Trust Tiers

Trust Score Algorithm (plain English)

Every score is recomputed live on read — no stale cache. Six independent signals combine into a number between 0 and 1000:

Volume (0–200, log scale). A first transaction is worth ~12 points; ten transactions ~40; a hundred ~80; a thousand caps at 200. We use a log curve so high-volume agents don't dwarf legitimate small ones — being active matters, but not exponentially.

Success rate (0–400, the heaviest signal). The percentage of transactions marked successful, multiplied by 400. A 99% success rate is worth 396 points; a 50% rate is only 200. This is the single most important factor — agents that ship reliably climb fastest.

Age (0–100, capped at one year). Each day since registration adds about a quarter of a point, maxing out at 365 days. Not a huge factor, but real: a one-day-old agent with five great transactions is still less trustworthy than a one-year-old agent with the same record.

Endorsements (0–200). Every peer endorsement (quality / reliability / speed / domain expertise) is worth 10 points, capped at 200. This is the social signal — other agents and operators vouching for this one.

Disputes (unbounded penalty). Each upheld dispute against the agent removes 30 points, with no floor below zero. A bad actor can score 1000 from the other categories and still bottom out at 0 from disputes alone.

Verification bonus (+100). A one-time bump when an operator verifies the agent (DNS proof, KYC, or platform-specific check). It's worth roughly the same as 10 endorsements — meaningful, not decisive.

The score then clamps to [0, 1000] and maps to a tier. Recompute happens on every transaction, endorsement, dispute resolution, and verification — so the score is never more than one event behind reality.

MCP Tools

Demo Data

On startup, five seed agents land in the DB at different tiers — demo-platinum-001 (AlphaAgent), demo-gold-001 (BetaAgent), demo-silver-001 (GammaAgent), demo-bronze-001 (DeltaAgent), demo-new-001 (EpsilonAgent). Use them for quickstart demos.

Quickstart Calls

# Public read — no key needed
curl -s http://localhost:8000/tools/get_leaderboard \
  -H 'content-type: application/json' \
  -d '{"limit": 5}' | jq .

# Register a new agent. AGENTTRUST_DEV_KEY is the free-tier key from your environment.
curl -s http://localhost:8000/tools/register_agent \
  -H 'content-type: application/json' \
  -H "x-api-key: ${AGENTTRUST_DEV_KEY}" \
  -d '{"agent_id":"my-agent","name":"My Agent","capabilities":["search","rag"]}' | jq .

# Record a successful transaction
curl -s http://localhost:8000/tools/record_transaction \
  -H 'content-type: application/json' \
  -H "x-api-key: ${AGENTTRUST_DEV_KEY}" \
  -d '{"from_agent":"my-agent","to_agent":"demo-platinum-001","amount_usd":42.0,"success":true,"description":"data lookup"}' | jq .

JSON-RPC MCP Transport

POST /mcp speaks the standard MCP JSON-RPC dialect:

{"jsonrpc":"2.0","id":1,"method":"tools/list"}
{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"get_trust_score","arguments":{"agent_id":"demo-gold-001"}}}

tools/call results return as {"content":[{"type":"text","text":"<JSON>"}],"isError":<bool>}.

Tests

pytest -v

42 tests covering: HTTP surface (16), trust-score edge cases (18), and fail-closed admin-key configuration (8). Specific edge cases included: zero-transaction agents, 100% dispute rate, max-endorsement cap, failure rates that erode score, score-bounded between 0 and 1000, and refusal to start or serve admin actions when AGENTTRUST_ADMIN_KEY is unset.

Deploying to Railway

./deploy.sh

Loads RAILWAY_TOKEN from the gitignored deploy-secrets file next to this repo, provisions the service via nixpacks, and sets env vars. AGENTTRUST_ADMIN_KEY must already be set; the script refuses to deploy without it and does not print key values. The public URL is printed when Railway has assigned one.

License

MIT.