Trust Lattice

Evidence-weighted trust graph / reputation lattice for multi-agent systems. Nodes are agents, tools, and claim-sources. Edges carry time-decayed evidence (attestations, challenges, endorsements). High-risk actions are gated until trust thresholds and endorsement quorums are met.

Full MCP surface so orchestrators (Cursor, Claude Desktop, custom agents) can query trust before tool calls.

Features

• Beta / Bayesian-ish trust updates with exponential evidence decay
• Path-trust aggregation (noisy-OR of path products)
• Sybil-resistance basics via identity verification stubs + influence caps
• JSON policy DSL for risk tiers: read | write | irreversible
• Explainable gate decisions (“why was this gated?”)
• Export: JSON, Mermaid, Graphviz DOT
• SQLite persistence, CLI, Vitest suite, optional Vite dashboard

See docs/ARCHITECTURE.md for the math and model.

Quick start

npm install
npm run build
npm run seed -- --force
npm test

MCP server (stdio)

npm run serve
# or after build:
node dist/index.js

Cursor MCP config

Add to your Cursor MCP settings (path may vary by OS):

{
  "mcpServers": {
    "trust-lattice": {
      "command": "node",
      "args": ["/ABS/PATH/TO/trust-lattice/dist/index.js"],
      "env": {
        "TRUST_LATTICE_DB": "/ABS/PATH/TO/trust-lattice/data/trust-lattice.db",
        "TRUST_LATTICE_ADMIN_TOKEN": "replace-with-a-long-random-secret"
      }
    }
  }
}

Or during development:

{
  "mcpServers": {
    "trust-lattice": {
      "command": "npx",
      "args": ["tsx", "/ABS/PATH/TO/trust-lattice/src/index.ts"],
      "env": {
        "TRUST_LATTICE_DB": "/ABS/PATH/TO/trust-lattice/data/trust-lattice.db",
        "TRUST_LATTICE_ADMIN_TOKEN": "replace-with-a-long-random-secret"
      }
    }
  }
}

Mutating tools (tl_register_node, tl_attest, tl_challenge, tl_endorse, tl_promote_identity, pubkey challenge tools) require an adminToken argument that matches TRUST_LATTICE_ADMIN_TOKEN (min 16 chars). Writes fail closed when the env token is unset.

CLI

npx tsx src/cli.ts seed --force
npx tsx src/cli.ts query agent:planner tool:web-search
npx tsx src/cli.ts gate agent:planner tool:shell write "run sandbox command"
npx tsx src/cli.ts export mermaid
npx tsx src/cli.ts promote agent:researcher email --issuer [email protected]

Dashboard

npm run dev
# → http://127.0.0.1:5173

MCP tools

Tool	Description
tl_register_node	Register/update node (always unverified; needs adminToken)
tl_promote_identity	Operator promote verification level (adminToken)
tl_begin_pubkey_challenge / tl_complete_pubkey_challenge	Ed25519 proof → pubkey
tl_attest	Positive evidence on an edge (adminToken)
tl_challenge	Negative evidence on an edge (adminToken)
tl_endorse	Path-scaled endorsement (adminToken)
tl_query_trust	Decayed trust + paths (read)
tl_gate_action	Allow/deny with explanation (read; advisory)
tl_explain_path	Supporting paths (read)
tl_export_graph	json | mermaid | dot (read)

Policy

Default policy lives in policies/default.json. Override with TRUST_LATTICE_POLICY or --policy.

Security

• Set TRUST_LATTICE_ADMIN_TOKEN (≥16 chars) for MCP writes; mutating tools fail closed without it.
• Clients cannot set identity.verification via registerNode / tl_register_node — only operator promote / tl_promote_identity or a completed Ed25519 pubkey challenge.
• Email/org verification remain operator stubs (no external IdP); pubkey requires a signature challenge.
• Treat the SQLite DB as sensitive operational state.
• MCP logs only to stderr (stdio transport).
• Dashboard binds to 127.0.0.1 by default.
• Gate decisions are advisory — orchestrators must enforce deny.

License

MIT