bladebro
面向 AI 智能体的隐蔽高效代理式浏览器驱动。工具少、控制全面、真正隐匿，并最大限度提高令牌效率。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add bladebro@4.6.2Bladebro 以 bladebro 发布，当前版本为 4.6.2。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
Stealthy and efficient agentic browser driver for AI agents. Few tools, full control, real stealth, max token efficiency.
npm install -g bladebro
Or use without installing:
npx bladebro mcp
Add Bladebro as a native plugin in the Web UI's Plugins page, or run:
dsh plugin --profile web add bladebro
Restart the profile and start a new chat. Desktop users can install bladebro
from the same Plugins page. For terminal setup, enable Desktop's Manage dsh
Command, fully quit the app, then use its bundled command:
dsh plugin --profile desktop add bladebro
Use Desktop's command for its profile; the npm DSH CLI cannot manage that app-owned profile. Open Desktop again after installation.
DSH discovers the five tools as mcp__bladebro__act, see, state, run and
vision (each with the same prefix). Schemas and descriptions come from the
binary; DSH handles permissions, image attachments, reconnection and cleanup
through its native MCP bridge. DSH may also expose its own generic MCP resource
helpers and choose PTC presentation. No API key, separate MCP configuration, global
Bladebro install or extra plugin package is needed. Chrome/Chromium must be
installed on the machine running the DSH host. A remotely hosted Web UI controls
that host's browser, rather than the device viewing the UI.
Disable/re-enable the Bladebro bundle in Plugins, or remove it with
dsh plugin --profile web remove bladebro (use desktop with Desktop's command).
Restart the profile and begin a new chat after package changes. Other tools and
settings remain intact. Do not also configure a second MCP server named
bladebro; DSH refuses the duplicate namespace.
Updates use DSH's package manager: dsh plugin --profile web update bladebro,
or Update in Plugins. The plugin and native binary travel in the same
package, so ordinary Bladebro releases update both. Existing global/Pi installs
are separate. No background downloader or installation scripts run.
Advanced configuration belongs in the profile's cordis.patch.yml:
- id: bladebro
config:
# Optional absolute source-build/portable executable; package binary is default.
binaryPath: /absolute/path/to/bladebro
env:
CHROME_PATH: /absolute/path/to/chrome
BLADE_HOME: /absolute/path/to/private/browser-data
toolCallTimeoutMs: 180000
Windows YAML paths can use forward slashes, such as C:/Tools/bladebro.exe.
The optional cwd and reconnect settings use DSH's native MCP configuration.
The default per-call budget is 180 seconds; raise it for longer run sequences.
Cancellation stops waiting, but an already-dispatched browser mutation may
finish: observe its destination before trying it again.
The npm CLI keeps its Node.js >=14 requirement; the DSH plugin uses the host's
Node.js requirement. Prebuilt binaries cover Linux x64/ARM64 (glibc >=2.28),
macOS Intel/Apple Silicon and Windows x64. Other targets need a source build and
binaryPath. See the official DSH plugin commands
and Desktop command runtime.
For AI agents that speak MCP (Model Context Protocol). Point your agent at it:
{
"mcpServers": {
"bladebro": {
"command": "bladebro",
"args": ["mcp"]
}
}
}
Five tools: act, see, state, run, vision. Full docs at github.com/dondai44423/bladebro.
For AI agents that run shell commands. One persistent Chrome instance across all commands (auto-daemon):
bladebro nav https://example.com # auto-starts daemon + Chrome
bladebro see content # uses same Chrome
bladebro act click e5 # uses same Chrome
bladebro stop # cleans up
# JSON output for agents:
bladebro nav https://example.com --json
bladebro see content --json
# Agent discovery (returns tool schemas + CLI mapping):
bladebro help --json
Prebuilt binaries available for:
linux-x64)linux-arm64)darwin-arm64)darwin-x64)windows-x64)Other platforms: build from source.
Apache-2.0