# Agent Trust Map

When AI agents interact through MCP tool calls, SolonGate tracks which agent made each call, computes a trust score per agent, and visualizes trust relationships in an interactive graph. Available in the dashboard under Trust Map.

Why Trust Map? When multiple AI agents (Claude Code, Cursor, custom bots) access the same MCP servers through SolonGate, you need visibility into who does what. Which agent triggers the most denials? Which agents share tools? Is a new agent behaving suspiciously?

How It Works

The trust map works in 4 steps, from identity capture to visualization:

STEP 1

Agent Identity Capture

CLI flag, HTTP headers, or MCP clientInfo

STEP 2

Audit Log with Agent ID

Every tool call tagged with agent identity

STEP 3

Trust Score Computation

Based on allow/deny ratios and PI detections

STEP 4

Dashboard Visualization

Interactive force-directed graph

Agent Identity Sources

SolonGate captures agent identity from 3 sources, in priority order. The first available source is used:

PRIORITY 1

CLI Flag

--agent-name "My Agent"

PRIORITY 2

HTTP Headers

X-Agent-Id, X-Agent-Name

PRIORITY 3

MCP ClientInfo

Auto-detected from initialize

Note: Most MCP clients (Claude Code, Cursor) automatically send their name via the MCP initialize handshake. For custom agents, use --agent-name or HTTP headers.

Setup

Trust map works automatically with zero configuration. Agent identity is captured from MCP client info on every connection. For explicit control:

Option 1: CLI Flag (stdio mode)

bash

1# Name the agent explicitly
2solongate-proxy --agent-name "Claude Code" --api-key sg_live_xxx -- npx @playwright/mcp@latest
3
4# The proxy will tag all audit logs with agent_id: "claude-code"
5# and agent_name: "Claude Code"

Option 2: HTTP Headers (HTTP mode)

bash

1# Start proxy in HTTP mode
2solongate-proxy --port 3100 --api-key sg_live_xxx -- npx @playwright/mcp@latest
3
4# Send requests with agent identity headers
5curl -X POST http://localhost:3100/mcp \
6  -H "X-Agent-Id: my-custom-bot" \
7  -H "X-Agent-Name: My Custom Bot" \
8  -H "Content-Type: application/json" \
9  -d '{"jsonrpc":"2.0","method":"tools/call","params":{"name":"file_read","arguments":{"path":"README.md"}}}'

Option 3: Automatic (MCP ClientInfo)

typescript

1// When an MCP client connects, SolonGate reads clientInfo from the
2// initialize handshake — no configuration needed.
3//
4// Example: Claude Code sends:
5//   { "name": "claude-code", "version": "1.0.32" }
6//
7// SolonGate captures this as:
8//   agent_id: "claude-code/1.0.32"
9//   agent_name: "claude-code"

Trust Score Algorithm

Each agent gets a trust score from 0 to 1 (shown as 0-100%) based on their behavior:

Trust Score Formula

Allow Rate

allowedCalls / totalCalls

60%

Clean Rate

1 - (piDetections / totalCalls)

30%

Recency

Active in last 24h = 1.0, 7d = 0.5

10%

trustScore = (allowRate * 0.6) + (cleanRate * 0.3) + (recency * 0.1)

Trust Levels

Trust scores map to visual indicators on the graph and table:

High Trust

80-100%

Agent behaves normally, few denials

Moderate

60-79%

Some denials or PI detections

Low Trust

40-59%

Significant denials or PI activity

Untrusted

0-39%

Mostly denied, potential threat

Agent Relationships (Edges)

The trust map doesn't just show individual agents — it shows how they relate to each other through shared tool usage.

How Edges Work

Shared Tool Usage

Two agents are connected if they both called the same tool(s) within the same project. For example, if Agent A and Agent B both use file_read and shell_exec, they share an edge.

Edge Thickness

Thicker edges mean more shared interactions. Edge trust score is the minimum of both agents' trust scores.

API Reference

The trust map data is available via the REST API:

bash

1# Get trust map data
2curl -H "Authorization: Bearer sg_live_xxx" \
3  "https://api.solongate.com/api/v1/trust-map?period=7d"
4
5# Response:
6{
7  "nodes": [
8    {
9      "id": "claude-code/1.0.32",
10      "name": "claude-code",
11      "trustScore": 0.92,
12      "totalCalls": 150,
13      "allowed": 145,
14      "denied": 5,
15      "piDetections": 1,
16      "lastSeen": "2026-03-14T10:00:00Z"
17    }
18  ],
19  "edges": [
20    {
21      "source": "claude-code/1.0.32",
22      "target": "cursor/0.45",
23      "sharedTools": ["file_read", "shell_exec"],
24      "interactionCount": 23,
25      "trustScore": 0.85
26    }
27  ],
28  "summary": {
29    "totalAgents": 3,
30    "totalInteractions": 45,
31    "avgTrustScore": 0.88
32  }
33}

Query Parameters

Parameter	Default	Description
`period`	30d	Time window: 24h, 7d, or 30d

Testing the Trust Map

To verify the trust map is working end-to-end:

Run proxy with an agent name

bash

1solongate-proxy --agent-name "Test Agent" --api-key sg_live_xxx -- npx @playwright/mcp@latest

Make some tool calls through the proxy

Use Claude Code, Cursor, or any MCP client connected through the proxy. Each tool call is logged with the agent identity.

Check the API

bash

1curl -H "Authorization: Bearer sg_live_xxx" \
2  "https://api.solongate.com/api/v1/trust-map"

You should see your agent in the nodes array.

Open the Dashboard

Go to Trust Map in the sidebar. Your agent should appear as a node with its trust score. If multiple agents connect, edges will form between agents that share the same tools.

Sub-Agent Tracking

When an AI client spawns sub-agents (e.g. Claude Code's Task tool), they share the same MCP connection. SolonGate can still track each sub-agent individually using the MCP _meta field or HTTP headers.

Option 1: MCP _meta field

Include sub-agent identity in the _meta field of each tools/call request:

json

1{
2  "jsonrpc": "2.0",
3  "method": "tools/call",
4  "params": {
5    "name": "file_read",
6    "arguments": { "path": "README.md" },
7    "_meta": {
8      "io.solongate/agent": {
9        "id": "researcher",
10        "name": "Research Agent"
11      }
12    }
13  }
14}

The proxy reads io.solongate/agent from _meta and creates a composite agent ID: parent-agent::researcher. This appears as a separate node in the trust map, linked to its parent.

Option 2: HTTP Headers (HTTP mode)

bash

1# Start proxy in HTTP mode
2solongate-proxy --port 3100 -- npx @playwright/mcp@latest
3
4# Send requests with sub-agent identity
5curl -X POST http://localhost:3100/mcp \
6  -H "X-Agent-Id: claude-code" \
7  -H "X-Sub-Agent-Id: researcher" \
8  -H "X-Sub-Agent-Name: Research Agent" \
9  -H "Content-Type: application/json" \
10  -d '{"jsonrpc":"2.0","method":"tools/call","params":{"name":"file_read","arguments":{"path":"README.md"}}}'

Visual cue: Sub-agents appear as smaller nodes with dashed borders in the trust map. Parent→sub-agent edges are shown as dotted lines.

Tips

Click a node in the graph to highlight its connections and see detailed stats.

Period selector lets you view 24h, 7d, or 30d of data. Use shorter periods for recent activity.

No agents showing? Make sure you're using a sg_live_ key. Test keys (sg_test_) skip audit log forwarding.

AI Judge Dashboard Guide

# Agent Trust Map

How It Works

The trust map works in 4 steps, from identity capture to visualization:

STEP 1

Agent Identity Capture

CLI flag, HTTP headers, or MCP clientInfo

STEP 2

Audit Log with Agent ID

Every tool call tagged with agent identity

STEP 3

Trust Score Computation

Based on allow/deny ratios and PI detections

STEP 4

Dashboard Visualization

Interactive force-directed graph

Agent Identity Sources

SolonGate captures agent identity from 3 sources, in priority order. The first available source is used:

PRIORITY 1

CLI Flag

--agent-name "My Agent"

PRIORITY 2

HTTP Headers

X-Agent-Id, X-Agent-Name

PRIORITY 3

MCP ClientInfo

Auto-detected from initialize

Note: Most MCP clients (Claude Code, Cursor) automatically send their name via the MCP initialize handshake. For custom agents, use --agent-name or HTTP headers.

Setup

Trust map works automatically with zero configuration. Agent identity is captured from MCP client info on every connection. For explicit control:

Option 1: CLI Flag (stdio mode)

bash

1# Name the agent explicitly
2solongate-proxy --agent-name "Claude Code" --api-key sg_live_xxx -- npx @playwright/mcp@latest
3
4# The proxy will tag all audit logs with agent_id: "claude-code"
5# and agent_name: "Claude Code"

Option 2: HTTP Headers (HTTP mode)

bash

1# Start proxy in HTTP mode
2solongate-proxy --port 3100 --api-key sg_live_xxx -- npx @playwright/mcp@latest
3
4# Send requests with agent identity headers
5curl -X POST http://localhost:3100/mcp \
6  -H "X-Agent-Id: my-custom-bot" \
7  -H "X-Agent-Name: My Custom Bot" \
8  -H "Content-Type: application/json" \
9  -d '{"jsonrpc":"2.0","method":"tools/call","params":{"name":"file_read","arguments":{"path":"README.md"}}}'

Option 3: Automatic (MCP ClientInfo)

typescript

1// When an MCP client connects, SolonGate reads clientInfo from the
2// initialize handshake — no configuration needed.
3//
4// Example: Claude Code sends:
5//   { "name": "claude-code", "version": "1.0.32" }
6//
7// SolonGate captures this as:
8//   agent_id: "claude-code/1.0.32"
9//   agent_name: "claude-code"

Trust Score Algorithm

Each agent gets a trust score from 0 to 1 (shown as 0-100%) based on their behavior:

Trust Score Formula

Allow Rate

allowedCalls / totalCalls

60%

Clean Rate

1 - (piDetections / totalCalls)

30%

Recency

Active in last 24h = 1.0, 7d = 0.5

10%

trustScore = (allowRate * 0.6) + (cleanRate * 0.3) + (recency * 0.1)

Trust Levels

Trust scores map to visual indicators on the graph and table:

High Trust

80-100%

Agent behaves normally, few denials

Moderate

60-79%

Some denials or PI detections

Low Trust

40-59%

Significant denials or PI activity

Untrusted

0-39%

Mostly denied, potential threat

Agent Relationships (Edges)

The trust map doesn't just show individual agents — it shows how they relate to each other through shared tool usage.

How Edges Work

Shared Tool Usage

Two agents are connected if they both called the same tool(s) within the same project. For example, if Agent A and Agent B both use file_read and shell_exec, they share an edge.

Edge Thickness

Thicker edges mean more shared interactions. Edge trust score is the minimum of both agents' trust scores.

API Reference

The trust map data is available via the REST API:

bash

1# Get trust map data
2curl -H "Authorization: Bearer sg_live_xxx" \
3  "https://api.solongate.com/api/v1/trust-map?period=7d"
4
5# Response:
6{
7  "nodes": [
8    {
9      "id": "claude-code/1.0.32",
10      "name": "claude-code",
11      "trustScore": 0.92,
12      "totalCalls": 150,
13      "allowed": 145,
14      "denied": 5,
15      "piDetections": 1,
16      "lastSeen": "2026-03-14T10:00:00Z"
17    }
18  ],
19  "edges": [
20    {
21      "source": "claude-code/1.0.32",
22      "target": "cursor/0.45",
23      "sharedTools": ["file_read", "shell_exec"],
24      "interactionCount": 23,
25      "trustScore": 0.85
26    }
27  ],
28  "summary": {
29    "totalAgents": 3,
30    "totalInteractions": 45,
31    "avgTrustScore": 0.88
32  }
33}

Query Parameters

Parameter	Default	Description
`period`	30d	Time window: 24h, 7d, or 30d

Testing the Trust Map

To verify the trust map is working end-to-end:

Run proxy with an agent name

bash

1solongate-proxy --agent-name "Test Agent" --api-key sg_live_xxx -- npx @playwright/mcp@latest

Make some tool calls through the proxy

Use Claude Code, Cursor, or any MCP client connected through the proxy. Each tool call is logged with the agent identity.

Check the API

bash

1curl -H "Authorization: Bearer sg_live_xxx" \
2  "https://api.solongate.com/api/v1/trust-map"

You should see your agent in the nodes array.

Open the Dashboard

Go to Trust Map in the sidebar. Your agent should appear as a node with its trust score. If multiple agents connect, edges will form between agents that share the same tools.

Sub-Agent Tracking

Option 1: MCP _meta field

Include sub-agent identity in the _meta field of each tools/call request:

json

1{
2  "jsonrpc": "2.0",
3  "method": "tools/call",
4  "params": {
5    "name": "file_read",
6    "arguments": { "path": "README.md" },
7    "_meta": {
8      "io.solongate/agent": {
9        "id": "researcher",
10        "name": "Research Agent"
11      }
12    }
13  }
14}

The proxy reads io.solongate/agent from _meta and creates a composite agent ID: parent-agent::researcher. This appears as a separate node in the trust map, linked to its parent.

Option 2: HTTP Headers (HTTP mode)

bash

1# Start proxy in HTTP mode
2solongate-proxy --port 3100 -- npx @playwright/mcp@latest
3
4# Send requests with sub-agent identity
5curl -X POST http://localhost:3100/mcp \
6  -H "X-Agent-Id: claude-code" \
7  -H "X-Sub-Agent-Id: researcher" \
8  -H "X-Sub-Agent-Name: Research Agent" \
9  -H "Content-Type: application/json" \
10  -d '{"jsonrpc":"2.0","method":"tools/call","params":{"name":"file_read","arguments":{"path":"README.md"}}}'

Visual cue: Sub-agents appear as smaller nodes with dashed borders in the trust map. Parent→sub-agent edges are shown as dotted lines.

Tips

Click a node in the graph to highlight its connections and see detailed stats.

Period selector lets you view 24h, 7d, or 30d of data. Use shorter periods for recent activity.

No agents showing? Make sure you're using a sg_live_ key. Test keys (sg_test_) skip audit log forwarding.

AI Judge Dashboard Guide