Integration

OpenAI Agents SDK

Know when your agent silently stops.

from nedwatch.integrations.openai_agents import NedRunHooks          # pip install ned-watch; set NED_CALLBACK_URL
result = await Runner.run(agent, "Triage today's tickets", hooks=NedRunHooks(every="1h", max="15m"))

event	what Ned gets
the first agent starts	the deadman is armed; an overrun run starts (with `max=`). Handoffs don't start a new run.
the run produces its final output	the overrun run finishes; the deadman checks in
an LLM reply / tool result (with `heartbeat=True`)	a check-in, at most one per 30 s
the run raises	the SDK has no error hook, so nothing is sent: no check-in (the deadman fires), and the open overrun fires at its deadline

The watch name defaults to the first agent's name.

To report a raised run as failed (Ned fires at once, with the error), use the wrapper:

from nedwatch.integrations.openai_agents import run
result = await run(agent, "Triage today's tickets", every="1h", max="15m")

Per agent

AgentHooks watch one agent wherever it runs:

from nedwatch.integrations.openai_agents import NedAgentHooks
billing_agent.hooks = NedAgentHooks("billing-agent", every="1h")

Give the agent Ned's own tools (MCP)

Ned has an MCP server, so an agent can register its own watches and check in by itself.

from nedwatch.integrations.openai_agents import mcp_server, hosted_mcp_tool

async with mcp_server() as ned_mcp:                       # your process connects to https://api.ned.watch/mcp
    agent = Agent(name="ops", instructions="...", mcp_servers=[ned_mcp])

agent = Agent(name="ops", instructions="...", tools=[hosted_mcp_tool()])   # or OpenAI's servers connect for you

Both send X-Ned-Ref: openai-agents (Ned's MCP passes it on) and Authorization: Bearer $NED_AGENT_KEY when it's set. mcp_server_stdio() runs uvx ned-watch-mcp locally with NED_REF=openai-agents. With no key, the agent's first quickstart or watch_register call creates one. Tools: quickstart, deadman_checkin, overrun_watch, overrun_start, overrun_finish, content_watch, watch_register, watch_get, watch_cancel, balance, pricing.

The same config by hand:

MCPServerStreamableHttp(params={"url": "https://api.ned.watch/mcp", "headers": {"X-Ned-Ref": "openai-agents", "Authorization": f"Bearer {key}"}})
HostedMCPTool(tool_config={"type": "mcp", "server_label": "ned_watch", "server_url": "https://api.ned.watch/mcp",
                           "require_approval": "never", "headers": {"X-Ned-Ref": "openai-agents", "Authorization": f"Bearer {key}"}})

Options

name, every (default "1h" when max isn't set), max, on_error, heartbeat, ned. New agents are tagged openai-agents.

Example

examples/openai-agents/ runs an agent with a scripted stand-in model (no API key), including a handoff and an agent that calls Ned's MCP quickstart tool, against a local Ned Watch and a local copy of the MCP server.