MCP 服务器

Agent Trust Stack MCP Server

Provides tools for Chain of Consciousness provenance logging and Agent Rating Protocol reputation scoring to establish trust and accountability for AI agents. It enables tamper-evident activity tracking, integrity verification, and bilateral reputation management.

README

Agent Trust Stack MCP Server

MCP (Model Context Protocol) server exposing the Agent Trust Stack tools so any MCP-compatible AI agent can use them natively.

Provides Chain of Consciousness (CoC) provenance logging and Agent Rating Protocol (ARP) reputation scoring — the two operational protocols from the 7-protocol Agent Trust Stack.

Tools

Tool	Description
`coc_init`	Initialize a new cryptographic hash chain (genesis block)
`coc_add`	Append an entry — learn, decide, create, error, note, milestone, session_start/end
`coc_verify`	Verify chain integrity (hash linkage, sequence, completeness)
`coc_status`	Get chain stats (length, agents, time span, event types)
`coc_tail`	Get the last N entries
`coc_anchor`	Submit chain hash for external timestamping (OTS + RFC 3161 TSA)
`arp_rate`	Submit a bilateral blind rating for another agent
`arp_check`	Check an agent's reputation score
`trust_stack_info`	Get info about all 7 protocols with whitepaper links

Resources

URI	Description
`trust-stack://protocols`	Overview of all 7 protocols with links
`trust-stack://installation`	Installation instructions

Installation

pip install agent-trust-stack-mcp

For proper OpenTimestamps .ots file format (optional):

pip install agent-trust-stack-mcp[ots]

Configuration

Add to your MCP client config (Claude Code, Cursor, etc.):

{
  "mcpServers": {
    "agent-trust-stack": {
      "command": "agent-trust-stack-mcp",
      "args": []
    }
  }
}

Environment Variables

Variable	Default	Description
`COC_CHAIN_DIR`	`./chain`	Directory for chain files
`ARP_RATINGS_DIR`	`./ratings`	Directory for rating files

Custom data directories

{
  "mcpServers": {
    "agent-trust-stack": {
      "command": "agent-trust-stack-mcp",
      "args": [],
      "env": {
        "COC_CHAIN_DIR": "/path/to/my/chain",
        "ARP_RATINGS_DIR": "/path/to/my/ratings"
      }
    }
  }
}

Usage Examples

Once connected, any MCP-compatible agent can call these tools directly:

Start a provenance chain

→ coc_init(agent="my-agent")
← { "status": "chain_initialized", "sequence": 0, "entry_hash": "a1b2c3..." }

Log activity

→ coc_add(event_type="learn", data="Processed 500 documents from dataset X", agent="my-agent")
← { "status": "entry_added", "sequence": 1, "entry_hash": "d4e5f6..." }

Verify chain integrity

→ coc_verify()
← { "is_valid": true, "entry_count": 42, "agents": {"my-agent": 42} }

Rate another agent

→ arp_rate(rater="agent-a", ratee="agent-b", score=0.8, context="Delivered accurate research")
← { "status": "rating_recorded", "rater_hash": "7f8a9b...", "score": 0.8 }

Check reputation

→ arp_check(agent_id="agent-b")
← { "rating_count": 5, "average_score": 0.72, "unique_raters": 3 }

Running Directly

# stdio mode (default — for MCP client connections)
agent-trust-stack-mcp

# Or via Python module
python -m agent_trust_stack_mcp

How It Works

Chain of Consciousness (CoC): An append-only JSONL file where each entry contains a SHA-256 hash linking it to the previous entry, creating a tamper-evident log. Any modification to earlier entries breaks the hash chain, making tampering detectable. Optional external anchoring via OpenTimestamps (Bitcoin) and RFC 3161 TSA provides independent timestamp proof.

Agent Rating Protocol (ARP): Agents rate each other on a -1.0 to 1.0 scale after interactions. Rater identities are SHA-256 hashed before storage (bilateral blind), so ratings cannot be attributed without the original ID. Reputation is the aggregate of all received ratings.

Registry Listings

Smithery — agent-trust-stack
Glama — agent-trust-stack
mcp.so — agent-trust-stack
MCP Servers — agent-trust-stack

Part of the Agent Trust Stack

This MCP server exposes tools from the Agent Trust Stack — seven interlocking protocols for autonomous AI agent trust infrastructure:

Chain of Consciousness — Provenance logging (this server)
Agent Rating Protocol — Reputation scoring (this server)
Agent Service Agreements — Machine-readable contracts
Agent Justice Protocol — Dispute resolution
Agent Lifecycle Protocol — Birth, migration, retirement
Agent Matchmaking — Capability discovery
Context Window Economics — Token resource management

Full stack: pip install agent-trust-stack

Security

VAM-SEC v1.0 — All CoC and ARP operations are local file I/O. No credentials are required or stored. No network calls are made except during optional coc_anchor (OTS calendar servers + freeTSA.org). No API keys needed.