dsh-browser-unified

Browser automation for DeepSeek Harness (dsh) on your own browser: a local WebSocket bridge plus a Chrome/Edge MV3 extension expose browser_* tools to the model, with a configurable URL policy in front. AGPL-3.0-only.

中文 · English

package	version	what it is
dsh-browser-unified	0.11.34	the plugin: 31 browser_* tools (27 driving a browser + 4 registry/tooling), GUI settings section
browser-unified-core	0.1.10	URL policy, approval layer, bridge (vendored from upstream), guarded bridge
dsh-browser-unified-mit	0.1.1	MIT edition without the bridge and without browser tools
extension DSH Browser Control	1.0.37	MV3 extension; use this repo's build, not upstream's

Install

Requirements: dsh web (0.1.7+ renders the settings card), Edge or Chrome 116+.

1. Extension — download DSH-Browser-Control-v1.0.37.zip from Releases, unzip anywhere, then edge://extensions (Chrome: chrome://extensions) → enable developer mode → Load unpacked. Source: extension/.

2. Plugin — download browser-unified-core-0.1.10.tgz and dsh-browser-unified-0.11.34.tgz from the same release, then:

   dsh plugin --profile web add \
     file:/abs/path/browser-unified-core-0.1.10.tgz \
     file:/abs/path/dsh-browser-unified-0.11.34.tgz
   

3. Restart dsh web. The bundled patch sets enabled: true, so the bridge listens on 127.0.0.1:9777 right away.

dsh plugin add github:Short-Arm-Ape/dsh-browser-unified does not work. pnpm resolves the repo root, which is not a package — it installs an @short-arm-ape/dsh-browser-unified@0.0.0 stub with no lib/ and no dsh.bundle.patch, so dsh never mounts it. GitHub installs only fit repos whose root is the plugin package. Use the release tarballs above, or file: a checkout's packages/ entry.

The MIT edition (dsh-browser-unified-mit, same release) carries the URL policy, approval and self-update modules but no bridge and no browser tools.

Configure

GUI: Settings → Browser (0.1.7+). Equivalent YAML in the profile's cordis.patch.yml:

- id: browser-bridge
  config:
    enabled: true
    urlMode: public      # public | intranet
    blockMetadata: true
    readOnly: false      # true = refuse every state-changing command before it is sent
    allowScript: true    # false = no browser_evaluate

Every field, its semantics and the validation rules: packages/browser-unified-plugin/README.md.

Tools

Navigation browser_navigate / browser_wait / browser_read / browser_find / browser_snapshot; interaction browser_click / browser_type / browser_press / browser_scroll; tabs and frames browser_tabs / browser_frames / browser_frame_eval; inspection browser_console_log / browser_network_log / browser_network_clear / browser_screenshot / browser_pdf / browser_emulate / browser_permissions / browser_dialog / browser_evaluate; session browser_info / browser_launch / browser_hold; housekeeping browser_cleanup / browser_tools; project registry browser_check_update / browser_design_show / browser_design_edit / browser_issues.

Security model

• The real boundary is the loopback address plus the token in the profile config. The agent can read and rewrite that config, so the token separates you from other local processes — not from the agent.
• URL policy — per-realm allow / ask / deny (internet, LAN, loopback), allow and deny host lists, cloud-metadata blocking, and an approval prompt for unknown hosts.
• readOnly / allowScript — refused commands never reach the browser, so they leave no side effects.
• Extension-side hard lock — the popup's 「锁定」 freezes every browser command; only the popup can undo it.
• Take-over overlay — browser_hold dims the page and shows what the model is doing; the user double-clicks to take the page back, after which the model's click / input / press / nav on that tab are refused until it turns the overlay off.

Verified

• node tests/run-all.mjs — 10 suites, 325 assertions, exit 0.
• CI on every push and PR: all three packages build from a clean checkout, those suites run, and lib/client.js is byte-identical to client/client.js.
• Real-browser checks (headless Edge plus the extension under test) are manual; CI never runs a browser.

Layout

packages/browser-unified-core/    AGPL core: url-policy / approval / bridge (vendored) / guarded-bridge
packages/browser-unified-plugin/  dsh plugin: 31 tools, client half, GUI settings section
packages/browser-unified-mit/     MIT edition (no bridge, no browser tools)
extension/                        MV3 extension source
design/registry.json              design entries, impact map, authorization policy
upstream/                         upstream license texts only; sources are pinned by SHA in upstream-baseline.json
tests/                            10 regression suites + run-all.mjs

Related projects

• caob23/dsh-browser-control — upstream bridge and MV3 extension this project vendors and extends (AGPL-3.0).
• dsh-intranet-browser — this plugin's predecessor (archived); the URL policy and approval layer started there.
• @deepseek-ai/dsh — the host.
• dsh-plugin.org — plugin hub; its install-failure report is where several diagnostics in this README come from.

License

Plugin and core: AGPL-3.0-only, which includes code derived from caob23/dsh-browser-control. Per-upstream attribution, versions and license matrix: NOTICE.md. The MIT edition is MIT-licensed and carries no upstream-derived code.

Upstream sources are not vendored: upstream/ keeps the license texts and upstream-baseline.json keeps one public repository plus commit SHA per upstream.