External agents and MCP servers are optional workers. Eyra owns the route, sandbox, approvals, timeouts, output caps, logs, cancellation, and final status. For new worker integrations, prefer Connectors. The connector manifest is the newer universal contract for CLI, MCP, local HTTP, remote HTTP, browser-agent, coding-agent, and guarded Python-module workers. The older external-agent and MCP tool gates remain available for built-in adapter workflows.

​
Capability gates

AGENT_TOOLS_ENABLED=false
EXTERNAL_AGENT_TOOLS_ENABLED=false
EXTERNAL_AGENT_CONFIG_PATH=~/.config/eyra/agents.json
MCP_TOOLS_ENABLED=false
MCP_CONFIG_PATH=~/.config/eyra/mcp.json
Enable only the bridge you need.

​
External agent config

{
  "agents": [
    {
      "name": "openhands",
      "type": "cli",
      "command": ["openhands", "run"],
      "cwdPolicy": "filesystem_default_path",
      "network": false,
      "mutatesFiles": true,
      "requiresApproval": true,
      "timeoutSeconds": 600
    }
  ]
}
Use static argv only. Eyra does not accept model-selected dynamic commands.

​
Coding jobs

Start a coding job with the configured terminal agent to update the README.
What is the coding agent doing?
Cancel that.
Coding jobs run as background tasks after approval. They are bounded and logged.

​
MCP bridge

list_mcp_tools reads configured stdio server tools. call_mcp_tool requires action-specific approval and returns capped output. MCP calls are local stdio processes from MCP_CONFIG_PATH. Treat each server’s side effects according to that server’s own behavior.

​
Safety model

  • Bridges are disabled by default.
  • Delegation is routed as DELEGATED_AGENT risk.
  • Mutating delegation requires approval.
  • Output is redacted and capped.
  • Working directories stay under sandbox policy.
  • Realtime voice does not get agent delegation by default.