sweep-mcp

MCP server that exposes the homelab model-testing sweep automation as a single tool: run_sweep. Composes Phase 3 of model-testing#36 — the operator says "Run qwen3.7:35b through the sweep test" in OWUI chat, the LLM dispatches run_sweep, this server forwards to the n8n model-sweep-kickoff webhook, and the existing Phase 1 + 2 chain takes over (GHA sweep → dual-summarize → tracking issue + Discord ping).

This repo holds the MIT-licensed source + Dockerfile. Deploy manifests live in dvystrcil/sweep-mcp.

What it does

Just one tool for v1:

Tool	Inputs	Returns
run_sweep	model (required), optional payload, runs, requester	{sweep_id, run_url, issue_url, issue_number, model, requester, eta} from the n8n kickoff

The MCP layer is intentionally thin — it just JSON-marshals the kickoff and surfaces the n8n response. The single-sweep mutex, GHA dispatch, and tracking issue creation all live in the n8n workflow that this server forwards to.

Follow-ups that are NOT in v1 but make sense to add later:

• sweep_status(sweep_id) — polls the GHA run, returns running/completed
  • percent-complete
• list_recent_sweeps(limit) — returns last N sweep_ids with model + status for the LLM to pick from when the operator says "rerun the last sweep"

Both require GitHub API auth (the kickoff doesn't); leaving them out keeps v1's blast radius tight.

Run locally

# Build + run against a real n8n
go run ./cmd/sweep-mcp --http=:8080 \
  --n8n-kickoff-url=http://localhost:5678/webhook/model-sweep-kickoff

# In another shell, invoke a tool over the streamable HTTP MCP endpoint:
curl -X POST -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  http://localhost:8080/mcp \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"smoke","version":"0.0.1"}}}'

Architecture

sequenceDiagram
    participant OWUI
    participant SweepMCP as sweep-mcp
    participant N8N as n8n model-sweep-kickoff
    participant GHA as model-sweep.yaml

    OWUI->>SweepMCP: tools/call run_sweep {model: ...}
    SweepMCP->>N8N: POST /webhook/model-sweep-kickoff
    N8N->>N8N: check in-flight (409 if busy)
    N8N->>GHA: workflow_dispatch
    N8N->>SweepMCP: 200 {sweep_id, issue_url, ...}
    SweepMCP-->>OWUI: tool result (JSON)
    Note over GHA: sweep runs ~5-90 min
    GHA->>N8N: notify complete (different workflow)
    N8N->>GH: post results comment + Discord ping

Memory hooks

• [[feedback_three_repo_split_for_new_services]] — build (this repo) + deploy (sweep-mcp) + ArgoCD app (argocd-projects)
• [[feedback_owui_skillids_vs_skills]] — binding the MCP into OWUI requires BOTH tool_server.connections AND the preset's meta.toolIds
• [[feedback_owui_mcp_dispatch_chain]] — tool_ids lives on form_data, not on meta.tools in the chat request
• [[feedback_n8n_expression_must_json_stringify]] — sibling memory; not relevant to this repo but mentioned for the related n8n workflow that receives our POSTs

Related

• Epic: https://github.com/dvystrcil/model-testing/issues/36 (Phase 3)
• Parent thesis: https://github.com/dvystrcil/homelab/issues/292 (A2 slot 4 — autonomous domain agents)
• Sibling MCP servers as conventions: tor-character-mcp-docker, rpg-dice-mcp-docker, wiki-mcp-docker
• n8n workflows we forward to:
  • https://github.com/dvystrcil/n8n-workflow/blob/main/workflows/model_sweep_kickoff.json
  • https://github.com/dvystrcil/n8n-workflow/blob/main/workflows/model_sweep_complete.json