Tool Permissions & Approvals

Agent Bridge includes a powerful permission system that gives you control over what tools the AI agent can use.

Overview

There are two layers of permission control:

1. Tool Permissions - Enable or disable specific tools from an MCP server
2. Tool Approvals - Control whether the agent needs your approval before executing a tool

1. Tool Permissions (Enable/Disable)

You can selectively enable or disable individual tools from any connected MCP server.

How It Works

• Navigate to Settings > Connections
• Click on an MCP server to expand its tools
• Toggle each tool on or off

Use Cases

• Disable dangerous tools (e.g., delete_file) while keeping others active
• Limit agent capabilities for specific use cases
• Test specific tool integrations in isolation

API Endpoints

Endpoint	Method	Description
/api/mcp/settings/{server_id}/tools	GET	List all tools with enabled status
/api/mcp/settings/{server_id}/tools/{tool_name}	PATCH	Toggle a tool's enabled state

2. Tool Approvals (Human-in-the-Loop)

Before the agent executes certain tools, it can pause and ask for your approval. This is the "Human-in-the-Loop" feature.

Approval Types

Type	Behavior
always	Tool runs automatically without asking
once	Ask once, then auto-approve for 1 hour
never	Always ask before running this tool

How It Works

1. The agent decides to call a tool (e.g., create_github_issue)
2. The system checks your approval preferences for that tool
3. If approval is required, the agent pauses and sends you a prompt
4. You approve or deny the action
5. The agent resumes (or reports the denial)

Setting Approvals

In the chat interface, when the agent asks for permission, you can choose:

• Allow Once - Approve this specific call
• Always Allow - Never ask again for this tool
• Deny - Reject this action

API Endpoints

Endpoint	Method	Description
/api/tool-approvals	GET	List all your approval preferences
/api/tool-approvals	POST	Create or update an approval preference
/api/tool-approvals/{tool_name}	DELETE	Remove an approval preference

Architecture

The permission system is implemented in the LangGraph agent orchestrator:

Agent decides to call tool
        ↓
[route_tools] checks approval status
        ↓
   ┌────┴────┐
   ↓         ↓
Approved   Needs Review
   ↓         ↓
Execute   [human_review_node] → Pause Graph
   ↓         ↓
Resume ←──────┘

Key Files:

• server/app/routes/tool_permissions.py - API endpoints
• server/app/services/agent/agent_orchestrator.py - Graph logic
• server/app/services/security/permissions.py - Approval checking logic