WinKit
Local-first MCP server · Read-only · No telemetry

Your machine,
visible to your agent.

WinKit gives coding agents a structured, permissioned view of Windows — answer real questions about crashes, processes, and health without guessing.

Get started View source
opencode — WinKit liveread-only
$ opencode --mcp-config examples/mcp/opencode.json
User: Why did my PC restart overnight?
Agent › shutdown_analysis(since_minutes=720)
{ boots: 1, unexpected_shutdowns: 1, last_shutdown_kind: "power_loss" }
"You had a power loss at 8:13 PM — event 6008 + 41. Not an update."
14
read capabilities
~17ms
median tool latency
18
failure eval scenarios
0
writes · cloud · telemetry

01 — CAPABILITIES

Evidence first.
Never a guess.

Diagnostics that read like a report

Measurements become signals, signals become ranked findings. Every report carries evidence-completeness so agents never over-read a partial view.

system_healthsystem_diagnoseshutdown_analysissnapshot

Full Windows surface

Processes, services, windows, drives, ports, event logs — one permissioned surface.

Fail-closed permissions

Four modes gate 14 capabilities. Denials carry a precise reason.

See the model

Sub-100ms reads

Every tool except snapshots stays sub-100ms regardless of system size.

Limits as output

Denied reads list as null with reasons — never silently dropped.

02 — PERMISSION MODEL

Fail closed.
Explain every no.

Safe and read-only grant every read capability. Anything else requires approval — and the agent always learns exactly why.

safeall reads · allow
read_onlyall reads · allow
approvalactions · ask first
unrestrictedpolicy off · audit

03 — INSTALL

Running in three commands.

01

Build

cargo build --release
02

Check health

winkit doctor
03

Connect agent

--mcp-config examples/mcp/opencode.json

04 — PERFORMANCE

Fast enough to ask twice.

17ms
system_info
28ms
list_services
71ms
list_processes
1.4s
full diagnose

Release build, fresh process per call incl. startup + MCP handshake. Windowed tools scale with window, not system size.

Stop guessing.
Start observing.

MIT · local-first · cargo test --features mocks