Fast Hands

Local-first execution and research layer for AI agents on Windows, Linux and macOS.

Fast Hands gives an MCP-capable assistant a fast local execution path while keeping a human operator in control. Core execution uses persistent PowerShell 7 on Windows, Linux and macOS; Windows UI automation and desktop screenshot capture remain optional Windows-only features.

Fast Hands does not include an LLM and does not require a paid API for its core features. The calling MCP client supplies the reasoning model.

Why it is different

Most computer-use servers focus on clicking and typing. Fast Hands treats execution as a controlled transaction loop:

plan -> safe point -> action -> checkpoint -> operator gate -> next action

If the operator interrupts a run, completed work is preserved. The caller can replace only the not-yet-executed tail with fast_revise, then continue with fast_resume instead of replaying successful steps.

Related MCP server: MSG - MCP Swarm Gateway

Highlights

• fast_exec - low-latency command execution through a persistent hidden PowerShell process.

• fast_batch - 1-200 PowerShell commands with separate results and checkpoints.

• fast_run - mixed multi-step workflows with durable progress and operator control, including native windows_ui steps when Windows UI Direct is configured.

• fast_resume / fast_revise - resume from a checkpoint or replace only the unexecuted tail; resume verifies tracked file artifacts before continuing.

• fast_reconcile_external - resolve an external side effect with unknown outcome after remote read-back, either confirming it so replay is skipped or marking it failed so the same operation can be retried safely.

• Local operator monitor with RUN / PAUSE / EMERGENCY STOP and interrupting messages.

• Windows UI Direct adapter via a compatible windows-mcp-server executable.

• Optional legacy UIA batch adapter through FAST_HANDS_LEGACY_UIA_OPERATOR.

• fast_web_search, fast_web_read, fast_web_deep - local public-web research.

• youtube_research - YouTube search, metadata, captions, local Whisper fallback and frame extraction.

• Loopback-only HTTP endpoints by default (127.0.0.1).

• MCP stdio mode for normal MCP clients.

Downloadable builds

GitHub Releases also provide platform builds that bundle the Node.js runtime:

• FastHands-Windows-x64.exe - single-file Windows x64 launcher with embedded Fast Hands payload.

• FastHands-Linux-x64.tar.gz - Linux x64 native launcher + bundled Node runtime.

• FastHands-macOS-x64.tar.gz - macOS Intel x64 native launcher + bundled Node runtime.

PowerShell remains the execution engine (built into supported Windows as a compatibility fallback; PowerShell 7 is required on Linux/macOS). Optional research features still require their documented Python/FFmpeg dependencies.

Install from npm

npm install -g fast-hands-mcp
fast-hands --version

Start MCP over stdio with fast-hands, loopback HTTP with fast-hands server, or the operator monitor with fast-hands monitor. You can also use npx fast-hands-mcp --help without a global install.

Requirements

Core:

• Windows 10/11, Linux, or macOS

• Node.js 20+

• PowerShell 7 (pwsh) on Linux/macOS

• PowerShell 7 recommended on Windows; Windows PowerShell remains a compatibility fallback.

Optional research:

• Python 3.10+

• FFmpeg for YouTube frame/audio operations

• Playwright browser support for rendered-page fallback

Optional Windows UI Direct (Windows only):

• A compatible windows-mcp-server executable. Fast Hands does not redistribute third-party binaries.

Quick start

git clone https://github.com/tomaszteee/FastHands.git
cd FastHands
npm install
npm test

Run over stdio:

npm run stdio

Run loopback HTTP MCP:

npm start
# MCP:    http://127.0.0.1:8797/mcp
# health: http://127.0.0.1:8797/health

Run the operator console (it also supervises the local HTTP MCP server):

npm run monitor
# http://127.0.0.1:8796/

MCP client example

{
  "mcpServers": {
    "fast-hands": {
      "command": "node",
      "args": ["C:/path/to/FastHands/server.mjs"]
    }
  }
}

Platform support

Linux/macOS use the same PowerShell 7 execution engine and checkpoint semantics. UI-specific tools fail clearly as unsupported outside Windows rather than degrading core health.

Windows UI Direct

Set the backend executable or command before starting Fast Hands:

$env:FAST_HANDS_WINDOWS_MCP = 'C:\tools\windows-mcp-server.exe'
node server.mjs

On Windows, Fast Hands opens the backend over MCP stdio and exposes convenience tools such as fast_ui_snapshot, fast_ui_click, fast_ui_type, and fast_ui_invoke. fast_ui_call provides access to the complete backend tool surface.

Inside fast_run or a saved macro, Windows UI Direct can be used as a normal checkpointed step:

{ "kind": "windows_ui", "tool": "Snapshot", "arguments": { "all_windows": true } }

If the backend is unavailable, shell/checkpoint functionality remains usable; UI-specific calls fail clearly.

External side-effect reconciliation

Any fast_run step can declare an external_effect when it may mutate state outside the local workspace, such as an API write or a browser/UI publish action:

{
  "kind": "windows_ui",
  "tool": "Invoke",
  "arguments": { "name": "Publish" },
  "external_effect": {
    "operation_id": "publish-post-2026-08-29-001",
    "target": "https://example.com/posts/123",
    "payload": { "action": "publish" },
    "confirmation": "reconcile"
  }
}

Before the action starts, Fast Hands durably records the operation identity, target and SHA-256 payload fingerprint with outcome unknown. This closes the crash window where a remote write may commit before the normal step checkpoint is advanced.

• confirmation: "execution" marks the effect confirmed when the underlying action returns successfully. A crash/timeout before that confirmation still leaves unknown and blocks replay.

• confirmation: "reconcile" is stricter and is the default for UI actions: even a successful click/call remains unknown until the caller reads the remote system back.

• Any external action failure or timeout is treated as unknown, even when fail_fast is false.

• fast_resume returns EXTERNAL_RECONCILIATION_REQUIRED and will not rerun an unknown operation.

• After read-back, call fast_reconcile_external with outcome: "confirmed" to advance past the step without replay, or outcome: "failed" to allow a controlled retry using the same operation identity and payload fingerprint.

• Reusing an operation_id with a different step, target or payload fingerprint is rejected.

Callers may provide payload_hash directly instead of payload when they already have a SHA-256 fingerprint. Optional receipt/read-back metadata is bounded before being persisted in runtime state.

Web and YouTube research

Install optional local research dependencies:

powershell -ExecutionPolicy Bypass -File .\scripts\setup-research.ps1

Linux/macOS research setup:

./scripts/setup-research.sh
# or: ./scripts/setup-research.sh --install-browser

To install Playwright Chromium too:

powershell -ExecutionPolicy Bypass -File .\scripts\setup-research.ps1 -InstallBrowser

FastWeb uses public search/read backends and local extraction. YouTubeResearch uses yt-dlp; captions are preferred, and faster-whisper can run locally when captions are missing. These integrations do not require a paid API.

Configuration

Safety model

Fast Hands is powerful software. A connected agent can run commands and, when UI backends are enabled, operate desktop applications.

The execution engine records durable run state and checks local operator control before and after safe units. File artifacts referenced by fs steps (and optional per-step artifacts paths) are fingerprinted with SHA-256 + size at checkpoints. fast_resume compares the current files with the last checkpoint and returns WORKSPACE_DRIFT_DETECTED instead of blindly continuing when tracked content changed. Steps marked with external_effect get a durable pre-action operation record; an unknown external outcome blocks resume/revise until fast_reconcile_external records remote read-back. Emergency Stop terminates active child processes and marks the affected step as uncertain when partial execution may have occurred.

Do not expose the monitor or HTTP MCP endpoint to an untrusted network without adding authenticated transport and least-privilege controls. See SECURITY.md.

Project boundaries

Included here: the Fast Hands server, operator monitor, research integrations and adapter code written for Fast Hands.

Not bundled: LLMs, paid APIs, third-party desktop binaries, personal logs/screenshots, local runtime state, credentials, machine-specific paths, or project handoff files.

Release integrity

The repository includes a tag-triggered release workflow prepared to verify the tag/version, rerun the full quality gate, build an npm-compatible source archive, export an SPDX SBOM, generate SHA-256 checksums, and create GitHub build-provenance attestations. No release or npm publication occurs merely by pushing main; a version tag is required.

Project policies: Architecture | Tools | Threat model | Security | Privacy | Scope | Support | Contributing.

License

MIT. Third-party dependencies and optional backends keep their own licenses; see THIRD_PARTY.md.

About

IMI Studio https://www.imistudio.pl