Skip to main content

Overview

OpenSRE integrates with OpenClaw in a native, one-way support loop:
  • OpenSRE can search recent OpenClaw conversations during an investigation.
  • OpenSRE can read full OpenClaw threads for extra engineering context.
  • OpenSRE writes completed RCA findings back into OpenClaw through the OpenClaw bridge.
OpenSRE no longer ships a separate opensre-mcp command. You do not need to register OpenSRE itself as an MCP server inside OpenClaw.

Prerequisites

  • Node.js v22.12+ in the same shell that runs OpenSRE
  • OpenClaw CLI installed and healthy (openclaw --help)
  • OpenClaw Gateway running for conversation tools / write-back
Preflight:
If you use nvm:
nvm alias default 22 only affects future shells — it does not switch the current shell that launches uv run opensre.

Setup

Option 1: Interactive CLI / onboard

Recommended bridge settings (stdio — what setup configures):

Option 2: Environment variables

Setup locks stdio mode and clears URL/token fields so you do not accidentally point at the Control UI.

Credentials

For local stdio, the OpenClaw CLI on PATH is enough. For remote MCP (streamable-http / sse), set OPENCLAW_MCP_URL and optionally OPENCLAW_MCP_AUTH_TOKEN. Do not use http://127.0.0.1:18789/ as the MCP URL — that is the OpenClaw Control UI / Gateway, not the MCP bridge.

Investigation tools

Preferred investigation path: search → get conversation → send message. Raw bridge tools remain available as fallback.

Write-back behavior

After every completed investigation, OpenSRE attempts to write the full RCA report back into OpenClaw.
  • If state contains openclaw_context.conversation_id, OpenSRE appends to that conversation first.
  • Otherwise it creates a new conversation with the RCA report, root cause, remediation steps, and confidence score.

Accurate RCA

OpenClaw improves RCA with engineering context, but is not enough alone. Also configure at least one of:
  • Grafana or Datadog for metrics/logs/traces
  • AWS for infra topology and runtime state
  • GitHub for deploy/code-change context
Smoke test:
Good local validation sequence:
Contributor regression coverage:
  • tests/integrations/openclaw/test_integration.py
  • tests/tools/test_openclaw_mcp_tool.py
  • tests/utils/test_openclaw_delivery.py
  • tests/e2e/openclaw/ — see its README

Verify

Verify exercises a live MCP list_tools call (plus stdio preflight). If it fails:

Troubleshooting

Security

  • Prefer local stdio over exposing a remote MCP endpoint when possible.
  • Treat OPENCLAW_MCP_AUTH_TOKEN as a secret — store it in the keyring / .env, not source control.
  • Write-back posts RCA content into OpenClaw; keep conversation access appropriately scoped.