Repository navigation
Sentry's first stable release turns it from a system monitor into an AI-operated one: an in-app chat panel lets a language-model agent inspect and act on your machine through a full MCP tool surface, with a new usage audit trail showing exactly who called what.
sentry-core
- New chat storage: persistent chat spaces with full message history (user/assistant/error turns), so conversations with the agent survive restarts.
- New MCP client identification: every tool call is attributed to a specific caller (Claude Code, Claude Desktop, the bundled Python agent, ChatGPT, Grok, or unknown), resolved from both the MCP handshake and the OS process behind the connection.
- New loopback connection resolution: matches an inbound MCP request to the actual OS process that made it, by inspecting the machine's own TCP table.
sentry-tauri-ui / sentry-mcp (in-process MCP server)
- The desktop app now hosts a full MCP server (Streamable HTTP over loopback) exposing 22 tools: browsing/searching processes and their details, killing a process safely (name-verified, with protection for critical OS processes), reading live system/disk/network summaries, querying historical CPU/memory timelines, starting/stopping/archiving tracked monitoring sessions, and reading/writing the chat panel (including posting replies, thinking-step progress, and error reports).
- Every tool call is now logged with its caller, arguments, status, and duration, feeding a new MCP client usage audit page.
- The app now launches and supervises a bundled Python AI agent process automatically — no manual setup required — and tracking sessions can be ended and archived to disk so their history isn't lost to normal data retention.
sentry-tauri-ui frontend
- New Chat tab: a full conversational UI for talking to the agent, with live streamed progress ("Calling get_system_summary", etc.), markdown rendering, and error display.
- New MCP Clients sidebar and usage page: shows every connected AI client, its call history, and per-call status/timing — a transparency/audit view into what an agent has been doing to the machine.
- Tracking (Track tab) gained richer session detail: stat cards, metric toolbar, pagination, and an ended-session banner tied to the new archival feature.
python-agent (new)
- A standalone LangGraph-based agent package that connects to the app's MCP server, answers chat messages using a selectable LLM (GPT-OSS 120B/20B, Qwen, or Llama, served via Groq), and streams its tool-use progress back into the chat panel in real time.
- Supports hot-swapping models mid-conversation from the app's UI without losing chat history, and automatically reconnects and catches up on missed messages if the app restarts.