Skip to content

Releases: bujosa/computer-mcp

Release list

v0.1.0 — Sendtry Computer (@sendtry/computer-mcp)

v0.1.0 — Sendtry Computer (@sendtry/computer-mcp) Latest
Latest

Choose a tag to compare

View all tags
@bujosa bujosa released this 04 Nov 01:09
v0.1.0
63c2601
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Sendtry Computer (@sendtry/computer-mcp v0.1.0)

Give your local stdio MCP agent its own dedicated Linux desktop in Docker, share its screen when you want to help, and teach reusable skills by demonstration.

Install & Quickstart

1. Install the CLI globally from npm (or clone from source):

npm install -g @sendtry/computer-mcp
computer-mcp --version

2. Provision and start a dedicated computer:

# macOS / Windows (Docker Desktop)
computer-mcp doctor --context desktop-linux
computer-mcp init --name my-computer --context desktop-linux
computer-mcp start --name my-computer --open

# Linux (local Docker Engine)
computer-mcp doctor --context default
computer-mcp init --name my-computer --context default
computer-mcp start --name my-computer --open

3. Connect your local stdio MCP harness:

computer-mcp connection create claude-work --name my-computer --access work
computer-mcp connection create codex-work --name my-computer --access work --client codex
computer-mcp connection create observer --name my-computer --access read --client vscode
computer-mcp connection create custom-agent --name my-computer --access work --client stdio

4. Install the operator skill (sendtry-computer):

DISABLE_TELEMETRY=1 npx --yes skills@1.7.4 add bujosa/computer-mcp \
  --skill sendtry-computer --agent codex claude-code cursor --copy

Highlights

  • Dedicated local Linux desktop in Docker (isolated offline network by default, optional internet mode) with Chromium, workspace file explorer, bounded command runner, screenshots, and bounded guest X11 mouse/keyboard input.
  • Scoped identity & named stdio connections: Per-computer and per-workspace UUID pins, separately revocable work and read grants, exclusive control leases, and instant human takeover via noVNC.
  • Teach by demonstration: Record an action trace and optional silent guest WebM video, consent to sampled frame analysis (workflow_analyze), review the agent's proposed procedure and input parameters (workflow_propose), and export a pinned SKILL.md.
  • Owner-verified task assignments: Assign scoped tasks to named work connections (task_start / task_finish) with factual agent reporting separated from independent owner verification and private Activity metadata.
  • Verified across real clients & hosts: 160 automated contract tests, 17-phase E2E lifecycle, and live acceptance runs across Claude Code, Codex CLI, and Gemini via Antigravity CLI.

Documentation

Assets 4
Loading