AI SREClickHouse Workshops
Build workshopInstructor

Archived: 06b AI SRE with hosted LibreChat

Archived instructor reference for the former optional hosted LibreChat module.

Your computer
macOS terminal: Run workshop commands in Terminal using zsh or bash.

Archived module

Module 06b is no longer part of the run of show. Do not provision LibreChat, learner accounts, or its MCP connections for the current workshop. Continue directly from Module 06 to Module 07. The notes below remain only as a historical reference.

Facilitator companion to 06b AI SRE with hosted LibreChat.

Historical timing

This module previously took about 15 minutes immediately before Module 07. Module 06 already covers the remote ClickStack MCP workflow.

Historical cloud prework

Learners must not deploy LibreChat or MongoDB locally.

  1. Deploy one HTTPS LibreChat instance before the event using the official Railway one-click guide, or use an organization-approved hosted instance.
  2. Create learner accounts or configure your identity provider. Never share a default workshop password.
  3. Configure a tool-capable LLM provider and model, store its API key only in the hosted deployment, and apply a workshop spend cap.
  4. Add and share these remote MCP definitions:
    • clickstack → https://mcp.clickhouse.cloud/clickstack using per-user CLICKSTACK_SERVICE_ID and OAuth.
    • github → https://api.githubcopilot.com/mcp/ using a per-user fine-grained PAT, read-only headers, and the repository toolset.
  5. Mount or copy app/librechat/librechat.yaml if the hosted deployment is managed as code. It contains no shared learner token.
  6. Create the SRE agent from app/librechat/sre-agent-definition.md and share it with learners as Viewer, or let learners create it during the exercise.
  7. Test from a non-admin learner account: run one ClickStack MCP query and one GitHub source read with the configured model. Both callbacks must return to the hosted HTTPS domain, not localhost.

LibreChat documents remote MCP OAuth, per-user credentials, and agent sharing in its MCP guide and Agents guide.

Talk track

  • Module 06 is an agent in the editor; 06b was the same evidence chain in a hosted chat UI.
  • LibreChat is a client surface. ClickStack/HyperDX and the telemetry remain managed in ClickHouse Cloud; GitHub remains the source-of-truth service.
  • Each learner owns their ClickHouse OAuth session and GitHub PAT.

Common failures

  • Hosted URL is unavailable: skip this archived module and continue with Module 06. Do not improvise a local Docker deployment during the workshop.
  • OAuth returns to the wrong host: correct LibreChat's public DOMAIN_SERVER and DOMAIN_CLIENT to the HTTPS deployment URL, then reconnect.
  • Wrong ClickStack service: replace the learner's CLICKSTACK_SERVICE_ID with the UUID from their ClickStack connection, then reconnect.
  • GitHub tool errors: replace the learner's PAT with a fine-grained token restricted to ClickHouse/ClickHouse_Demos, Contents: read.
  • No browser fault telemetry: open the local app after injecting the fault so the browser SDK records it.

Reset steps

  • Revoke the learner's MCP connections or PAT in hosted LibreChat.
  • Remove temporary learner accounts after the event according to your access policy.
  • Reset the local workshop app with Module 07's stash-and-switch steps.
  • Keep the hosted instance patched or tear down the Railway project after the workshop.

On this page

EN