Skip to content

Dashboard

Open Brain ships with a dark-themed monitoring dashboard built on customtkinter. It shows memory stats, service health, recent captures, most-accessed memories, and OpenTelemetry observability metrics.

Features

Panel What it shows
Stat Cards Total memories, +1 hour, +24 hours, pinned, upvoted, accessed
Service Pills PostgreSQL, Ollama, MCP Server status (green/red/yellow). One-click Start button when a service is down.
Titlebar Dots Per-service colored dots at a glance
GPU Selector (v0.10.0+) Dropdown auto-populated from nvidia-smi -L showing the available CUDA devices ("RTX 5090", "RTX 3080 Ti", "Auto"). Hidden if nvidia-smi reports no GPUs. Apply button does a graceful Ollama restart so the new selection takes effect.
By Type / By Project Bar charts showing memory distribution
Recent Memories Last 10 memories with clickable rows to read full content
Most Accessed Top 10 most-recalled memories (quality signal)
Observability Strip Tool call counts, error rate, top tools, last restart. Reads from OpenTelemetry traces.
Server Log Live tail of OTel trace log + Ollama log

Event-Driven Refresh

The dashboard does not poll on a timer. A PostgreSQL trigger fires pg_notify('memories_changed') on every INSERT, UPDATE, or DELETE to the memories table. The dashboard listens on a dedicated connection and refreshes data widgets instantly when something changes.

Service health checks (Ollama, MCP) run on a separate 60-second interval since they don't need to be instant.

This eliminates the "strobe" effect that timer-based polling causes.

Launching

The dashboard detects whether Open Brain is already running. If services are up, it opens directly. If not, it shows a startup splash that drives scripts/infrastructure.py's bring_up() in a worker thread — bringing up Docker, the open-brain-db container, and ollama serve in sequence while streaming structured progress to the splash and to logs/startup.log. Since v0.8.0 this is a pure-Python flow (no .cmd chain, no pipe inheritance); since v0.9.0 shutdown uses a real Win32 Ctrl+Break graceful signal for ollama instead of taskkill /F; since v0.10.0 a single-instance guard inside dashboard.py itself prevents a duplicate window when the desktop shortcut is double-clicked twice — it focuses the existing window instead of spawning a second one.

Double-click Open Brain Dashboard on your Desktop. To create shortcuts:

powershell -ExecutionPolicy Bypass -File scripts\windows\create-desktop-shortcuts.ps1

# Windows (no console window)
.venv\Scripts\pythonw.exe dashboard.py

# Windows (with console for debugging)
.venv\Scripts\python.exe dashboard.py

Restart MCP

The titlebar has a Restart MCP button that:

  1. Kills the existing server.py process (via WSL pkill)
  2. Polls until the process is confirmed dead
  3. Launches a new server in a tmux session
  4. Polls until the process is confirmed alive
  5. Shows step-by-step progress in the status bar

Memory Popup

Click any row in the Recent Memories or Most Accessed tables to open a popup showing the full memory content, type, project, and timestamp.

Observability

The dashboard reads OpenTelemetry spans from logs/otel-traces.jsonl. Every MCP tool call produces a span with:

  • mcp.tool attribute (tool name)
  • status (OK or ERROR)
  • duration_ms (call latency)
  • trace_id and span_id for correlation

The observability strip aggregates these into: total calls, total errors, error rate %, top 3 tools by call count, and last server restart timestamp.

The dashboard itself writes its own spans to the same file (service name open-brain-dashboard) so refresh cycles are also traceable.

Requirements

The dashboard requires two additional Python packages (both included in requirements.txt):

  • customtkinter -- dark-themed Tk widgets
  • pillow -- image handling for the app icon

Troubleshooting

Symptom Fix
Dashboard hangs on launch Check that PostgreSQL is reachable. The dashboard uses 127.0.0.1 (not localhost) to avoid DNS resolution issues on Windows.
Taskbar shows generic Python icon Reboot Windows to flush the icon cache. The dashboard sets SetCurrentProcessExplicitAppUserModelID and DPI awareness.
Observability strip shows zeros Make some MCP tool calls (search, remember, etc.) to generate OTel spans in logs/otel-traces.jsonl.
MCP status shows offline If the server was active within the last 5 minutes, it should show online. Check that logs/otel-traces.jsonl has recent spans with "service": "open-brain".

v2 Monitoring

The v1 dashboard does not currently show brain_v2 stats. During the v2 soak period, use these MCP tools from any agent session to monitor v2 health:

  • health_v2 — DB connectivity, Ollama reachability, table row counts, server uptime, tool list
  • metrics_v2 — per-tool call counts, error rates, avg/p99 latency
  • recent_errors_v2 — last N errors from the in-memory ring buffer

v2 also writes persistent telemetry to the tool_events table (every tool call with timing) and structured JSONL to logs/brain_v2.jsonl (rotating, 5MB x 5 files).

A unified dashboard covering both v1 and v2 is planned for post-soak.