Skip to main content
Glama

Capability SDK

Give your existing agent browser, Office, terminal, code, charts and knowledge tools through one SDK or MCP server.

简体中文 · Tool reference · Setup · SDK development

Capability SDK packages tool implementations, input schemas, operating instructions, structured results and local runtime preparation together. Connect a supported MCP client, or mount selected capabilities in your own agent process. Your agent keeps its model and orchestration.

The current standalone release is 0.3.3, distributed through GitHub Releases. The installer supports GitHub packages directly; npm publication is optional. npm 0.3.0 does not contain the unified setup command.

What your agent can do

Capability

Included work

Setup

Browser

Playwright sessions, DOM reads, screenshots and interactions

Managed Chromium

Files and Office

Read, generate, edit, render and convert documents, spreadsheets, presentations and PDFs

LibreOffice/UNO and format libraries

Terminal and code

Run commands, manage terminal sessions and execute local code

Local processes and managed Python

Charts

Create ECharts and 3D visualizations, preview and export

Bundled visualization UI

Knowledge

Ingest content and retrieve relevant passages

Local storage; configure optional services

Media

Inspect media and extract frames

FFmpeg; OCR, transcription and generation need providers

Computer

Observe and control a Windows desktop

Windows driver or a configured endpoint

Maps, data and integrations

Map tools, database adapters, connectors and enterprise communication

Opt-in configuration and service credentials

Use tool configuration to select what you expose. Sensitive-data detection is a separate model middleware; adding an MCP connection does not automatically intercept model requests.

Related MCP server: UBA MCP Server

Install into an existing agent

Requires Node.js >=22.16 with npm. Automatic runtime preparation supports Windows x64 and Ubuntu 22.04/24.04 or Debian 12/13 on x64/arm64. macOS automatic runtime preparation is not currently supported.

Windows PowerShell:

irm https://raw.githubusercontent.com/cyxysky/capability-sdk/main/install.ps1 | iex

Supported Linux systems:

curl -fsSL https://raw.githubusercontent.com/cyxysky/capability-sdk/main/install.sh | sh

These scripts install release 0.3.3 into ~/.capability-tools, prepare runtimes and merge supported user-level MCP configurations with backups. When the npm version is unavailable, they download the matching GitHub Release package and verify its SHA-256 checksum. Reload your client and enable the capability-sdk connection. The first full runtime preparation can download several GB; Linux system dependencies may need root or passwordless sudo. See setup details.

For a reviewed local clone, preview or select clients:

node install.mjs --dry-run --clients cursor,codex,claude-code
node install.mjs --clients cursor,codex,claude-code --project ./agent-tools

Supported configuration targets: Cursor, Codex, Claude Code, Claude Desktop, VS Code, Windsurf and Devin. This installs tools and prepares connections; client applications are installed separately.

Use the tools in your own agent

Install the prepared release in your agent project:

npm install https://github.com/cyxysky/capability-sdk/releases/download/v0.3.3/cjfclonedeep-capability-sdk-0.3.3.tgz

Once the same version is available on npm, npm install @cjfclonedeep/capability-sdk@0.3.3 is equivalent. Mount only the capabilities you need, then give their schemas, instructions and execution functions to your model adapter:

import { createLocalCapabilities } from '@cjfclonedeep/capability-sdk/local';

const runtime = await createLocalCapabilities({ tools: ['browser', 'file', 'chart'] });
try {
  for (const { publicName, tool } of Object.values(runtime.snapshot.tools)) {
    console.log(publicName, tool.description, tool.input.jsonSchema);
  }
  // Pass runtime.snapshot and runtime.instructions to your agent integration.
} finally {
  await runtime.dispose();
}

Start the process with the tool project as its working directory. Native adapters are available at /ai-sdk, /openai and /mcp. See agent integration examples, framework integration and the runnable browser example.

Browser code uses the provided Node-side page binding. Read DOM inside the page callback:

var info = await page.evaluate(() => ({ title: document.title }));
nodeRepl.write(info);

Develop and release

The source package has its own TypeScript configuration and dependencies. It does not require the original agent application. Maintainers can typecheck without building or downloading the tool runtimes:

CAPABILITY_SKIP_RUNTIME_INSTALL=1 npm install
npm run typecheck

For the independent repository export and manual release workflow, see maintainer notes. Source code is licensed under MIT. Dependencies and downloaded third-party runtimes retain their own licenses.

Related MCP Connectors

Related MCP Servers

  • A
    license
    B
    quality
    A
    maintenance
    Provides AI agents with a real browser environment for web automation, memory, and secure credential management through 15 MCP tools.
    15
    MIT
  • A
    license
    Not graded
    quality
    B
    maintenance
    Enables AI agents to securely access local development capabilities such as files, Git, processes, browser control, Windows automation, WSL, and observability through MCP on Windows and macOS.
    MIT
  • A
    license
    Not graded
    quality
    B
    maintenance
    Enables AI agents to automate both web browsers and native desktop applications through a unified interface, supporting tasks like browsing, clicking, typing, and file clipboard operations via exposed MCP tools.
    1
    Apache 2.0