dsh-doyourweb

Reference and control your existing, logged-in Chrome and Edge tabs from DeepSeek Harness, with compact page context and batched browser actions.

Version 0.2.3 fixes default DOM reads on pages with additional body elements inside shadow roots, while retaining shadow and iframe content. Navigation batches reject missing or unsupported step.value URLs before dispatching any action. Installation and connection guidance now explains the browser entry points, unique read roots, snapshot replacement and cursor pagination. See browser workflow and diagnostics, optimization decisions and app control.

English | 简体中文

Local Demo · Quick Start · Usage

Current browser panel rendered with the installed DSH React and UI primitives, using sample tabs. This is a component preview, not a full DSH host screenshot.

---

01 · Introduction

Browser tasks often depend on pages you have already opened and accounts you have already signed into. dsh-doyourweb connects those tabs to DeepSeek Harness instead of launching a separate automation browser or exporting cookies.

The project has two parts: a DSH plugin and a companion Manifest V3 extension. The plugin adds browser references to the native @ and / menus; the extension supplies tab metadata and browser control. Once paired, the extension reconnects in the background without keeping its popup open.

Current version: 0.2.3 for both the package and browser extension. See the changelog.

02 · Preview

1. Use /browser to open the connection page and locate the extension folder.
2. Load the extension in Chrome or Edge, then approve its matching four-digit pairing code.
3. Open Browser from the left navigation, or use @, to add an existing page to the current chat draft. Send the draft when ready.

The real annotation toolbar rendered on a local test page. Collapse keeps the toolbar visible; dock leaves a small edge button and preserves the draft and position.

The real local connection page rendered with a sample pairing request. These previews are not an automation recording.

Local Demo

There is no public hosted demo. After the DSH plugin starts, its default local connection page is http://127.0.0.1:43187/setup. Browser control runs through this local bridge and the installed extension.

03 · Core Features

Feature	Current behavior
Existing tabs	Reuses regular HTTP/HTTPS and authorized local file:// tabs and their login state in Chrome and Edge.
Native references	Lists page titles, sites and icons through DSH's @ and / menus.
Browser sidebar	Native left navigation opens a session-owned right panel for pages, running apps and local diagnostics. Picks append references to the current draft without sending.
Compact context	Page/app context, annotations and manually referenced logs appear as expandable chips; their serialized model context is preserved.
Compact reading	Accessibility snapshots support selector/query narrowing and cursor pagination. Initial reference context shares a default 2,400-character budget.
Batched actions	Runs up to 12 typed actions per call, including uploads; stops on failure/document navigation. Query/hash routes only continue observations or target-free steps.
Result extraction	Returns selected text, value and href fields, with up to 20 rows per extraction.
Scoped receipts	summary skips the next snapshot; handle returns a fresh ID/cursor without text. observation explicitly narrows the post-action snapshot. Read the handle cursor before using element refs.
Final checks	Optional expect checks visibility, contained text, an exact input value or exact URL once after the batch. The result stays separate from completed actions and does not replay effects.
File uploads	Hidden inputs, custom choosers and iframe selectors; requires a selected tab and exact user-supplied absolute file paths.
Drag and drop	Drag between elements in the same frame, or drop authorized local files and MIME data onto an element.
Network metadata	Explicit start/stop capture, 200-entry capacity, filtered pages of 10 (up to 20). No query strings, credentials, headers or bodies.
Local diagnostics	Bounded in-memory command stages, duration and error codes; disable, clear, export or explicitly add logs to the draft.
Browser stop	Input displays private stop controls and a brief target highlight; cancellation stops later steps and reports uncertain effects.
Screenshots	Saves PNG files locally and returns their paths instead of embedding base64.
Background connection	Saves pairing settings, reconnects automatically, and releases idle debugger connections.
Element annotations	Stage multiple comments with selectors, styles, bounds and optional crops; edit or delete staged notes. The session-bound queue survives collapse, docking and reattachment. Confirm notes into the chat draft, then send from chat.
Running apps	Search visible Windows application windows with @, with cached program icons. References include process lifetime and window identity.
MCP / CLI	Known official connectors, MCP Registry and npm CLI discovery, stdio / Streamable HTTP, and reviewed custom configurations.
Installation approval	Native DSH questions or a local /apps page; exact package versions and npm lifecycle scripts disabled. Optional repeat-install bypass is restricted to reviewed, trusted sources.
Desktop fallback	Cua Driver MCP, bounded text-only reads by default, on-demand images, cached checks and automatic launch for selected apps and explicit stops for repeated polling.
Mouse input	app_mouse batches up to 8 clicks, double/right clicks, scrolls and drags from one visual snapshot; separate foreground approval.

Browser budgets count JavaScript characters, including JSON escaping and metadata. App tool limit values bound body text; status, steps and image-path metadata are additional. MCP results omit duplicate JSON, and CLI stdout/stderr share one body budget. Character counts are not model tokens, and output budgets do not limit DOM traversal time.

WebMCP-First Control (Experimental)

Version 0.1.1 checks for page-provided WebMCP tools before collecting a DOM snapshot. It prefers tools that cover the task, falls back to DOM when needed, and reuses unchanged catalogs and validators. Successful calls return a fresh token for the next call without relisting schemas. WebMCP requires a compatible browser API and site-provided tools; ordinary pages keep using DOM automation. See the experiment guide for requirements, tools and local browser tests.

04 · Architecture

Layer	Technology
DSH host plugin	Node.js 22+, JavaScript ESM, Cordis integration
Native client UI	React and DSH UI primitives
Browser engine	Playwright Core 1.64.0, Chrome DevTools Protocol
Site tools	document.modelContext, Ajv 8.20.0 argument validation
Local bridge	WebSocket (ws), Zod validation
Browser extension	Manifest V3, chrome.tabs, chrome.debugger
App connectors	Official MCP SDK 1.32.1, stdio / Streamable HTTP, literal CLI argument arrays
Build and checks	esbuild, Node.js test runner, GitHub Actions

flowchart LR
    subgraph DSH["DeepSeek Harness"]
        UI["Native @ / menus"]
        Host["Host plugin"]
        Grants["Conversation grants and context budget"]
        Engine["Playwright engine"]
        Bridge["Loopback WebSocket / CDP relay"]
    end
    subgraph Browser["Existing Chrome / Edge"]
        Extension["Manifest V3 extension"]
        Tabs["Logged-in HTTP/HTTPS tabs"]
    end
    UI <-->|"Strict Remote RPC"| Host
    Host --> Grants --> Engine
    Engine <-->|"CDP"| Bridge
    Bridge <-->|"Origin check and pairing token"| Extension
    Extension <-->|"Tab inventory and debugger"| Tabs
    Engine -->|"Snapshots, receipts, PNG paths"| Host
    Host --> Apps["App discovery and human approval"]
    Apps --> Structured["Official / community MCP and CLI"]
    Apps --> Desktop["Cua Driver desktop fallback"]

Only tabs referenced by the user in the current conversation are available to the model's browser tools. Page content is treated as untrusted data, not instructions or permission.

05 · Quick Start

Requirements

• Node.js 22+ and npm.
• DeepSeek Harness; plugin API compatibility has been checked against 0.2.0-rc.2.
• Chromium-based Chrome or Edge 125+, with developer mode available for loading an unpacked extension.

Install the DSH Plugin

For npm installation, open Plugins > Add Plugin in DSH, select the official npm registry, and enter:

dsh-doyourweb

This requires a published package. If the registry returns 404, use the local installation below. Mirrors may take time to sync a newly published version.

Build and Install Locally

git clone --branch ver0.2.3 https://github.com/SheepofSheep/dsh-doyourweb.git
cd dsh-doyourweb
npm ci
npm run build

In DSH, open Plugins > Add Plugin and enter the absolute path to this built project directory. The bridge starts when the plugin activates. No separate development server is required.

Connect the Browser

First install and enable dsh-doyourweb in DSH Plugins > Add plugin, then connect the companion browser extension, then reference a page in chat. These are separate steps; installing the plugin does not install or pair the extension.

Once enabled, Browser in the left navigation opens the page panel; Connection settings opens pairing. Alternatively, type /browser in the composer and select Browser. The @ menu should offer the browser setup candidate even before the extension is installed or connected. After pairing, use the panel or @page title to reference a page.

If both the navigation entry and /browser candidate are missing, check that the plugin is installed and enabled, fully quit DSH including the system tray, then restart. If still missing, report the full DSH build, plugin version, missing entry location and loading errors. If the entry shows Not connected, open setup and follow the pairing steps below. For Service unavailable, inspect plugin loading errors and the local bridge service before retrying pairing.

1. Enter /browser in DSH and open the local connection page.
2. Open edge://extensions or chrome://extensions, enable developer mode, and load the project's extension/ folder as an unpacked extension.
3. Compare the extension's four-digit badge code with the pending request on the connection page, then approve it.
4. Return to DSH, select a tab with @, and send your request.

The extension popup and pairing page do not need to remain open. Installing the DSH plugin does not install the browser extension automatically.

After upgrading the DSH plugin, fully quit DSH from its tray/background process and relaunch it so the new host module is loaded.

Configuration

Defaults are declared in cordis.patch.yml:

Setting	Default	Purpose
port	43187	Local bridge port
snapshotChars	2400	Page-read and initial reference-context budget
idleTimeoutMs	120000	Idle time before releasing the debugger connection

The data directory is <DSH_HOME>/browser-tabs/, or ~/.dsh/browser-tabs/ when DSH_HOME is unset. An explicit plugin home setting takes precedence. Set environment variables on the DSH process; the plugin does not require its own model API key.

Usage

The left navigation's 浏览器 button opens the native right panel with pages, apps and diagnostics. Clicking an item inserts a reference into the current draft; it does not send it. The @ menu remains available. Page/app context, annotations and explicitly referenced logs display as compact expandable chips while preserving their full model payload. Diagnostics stay in local memory (600 metadata entries), can be paused, cleared or exported, and only enter a draft when you choose to reference them. The annotation toolbar can collapse its editor or park at the page edge; parking stops picking and keeps the unfinished note and position for restoration.

After referencing a page, you can ask:

Read the result list on this page and collect its titles and links. Do not submit any forms.

Model tool	Use
browser_read	Default mode: "auto" discovers WebMCP tools first, otherwise returns a DOM snapshot. Use mode: "dom" when site tools do not cover the task; narrow with selector, query or depth. cursor paginates DOM content; limit accepts 400-8,000 characters.
browser_act	Use the latest snapshot ID to click, fill, press, select, check, hover, scroll, navigate, go back/forward, wait, extract, upload, drag or drop. drag uses source ref/selector and destination toRef/toSelector in the same frame. drop accepts user-authorized local files and/or MIME data.
browser_upload	Upload 1-20 user-specified absolute local file paths through a file input or chooser button, with optional iframe selector frame. Requires a current snapshot. Files may upload immediately.
browser_capture	Save a screenshot when visual inspection is needed.
browser_network_requests	Explicitly start, filter/list, inspect details, clear or stop network metadata capture.
browser_network_requests	Explicit action:start captures up to 200 request metadata entries in memory. list supports query, errorsOnly, includeStatic, afterId and limit (max 20). details with requestId adds timing/MIME metadata; clear clears entries and stop removes listeners. No prior requests, URL credentials/query/fragment, headers or bodies are recorded. Idle detach discards the capture.
browser_webmcp_list	Discover or narrow page tools with query; reuse an existing catalog while it remains valid.
browser_webmcp_call	Call a matching page tool using catalog, name and arguments; success returns nextCatalog for another call.

When multiple tabs are authorized, pass the selected tab's reference as tab. DOM reads may opt into scope: "main"; default body retains navigation and dialogs. Upload authorization, handle receipts and measured limitations are described in the optimization guide.

Annotations and App Control

After installing or upgrading, fully exit DSH from the system tray, restart it, and reload the browser extension. For local development, build this working tree and install its absolute directory path. Do not enable the npm copy and local copy on the same port.

• Annotations: choose “批注当前页面” in the extension popup, or ask DSH to enable annotations on a referenced tab. Pick an element, write a comment, then press Enter or click “加入聊天草稿”. It appends to the target chat draft with an optional image attachment. Drag the toolbar as needed, collect multiple annotations, then press Enter in chat to send them together. Shift+Enter inserts a newline. The plugin keeps no annotation files or library. Controlled pages also show a top-right annotation button.
• Apps: select a running application with @Unity or another program name. DSH discovers connectors with app_discover, requests connection/installation through app_connect, then invokes MCP tools or CLI arguments with app_call. A connector found outside discovery can be proposed with its source documentation for human review.
• Management: /apps opens the local application, connector and approval manager. It need not remain open. Files, folders and conversations in @ each show three previews with expand/collapse controls; individual folders retain native navigation.
• Fallback: app_desktop uses upstream Cua Driver when structured tools cannot cover the task. Default reads return bounded controls without screenshots. Use screenshot: true only for visual grounding; app_mouse then supports up to 8 known-target clicks, scrolls or drags without intermediate captures. Missing windows and repeated inspections return stop: true. It does not bypass OS permissions.

See installation policy, Unity requirements and limitations.

Tool example: read a page, then extract a short result list

Call browser_read for the single authorized tab:

{
  "mode": "dom",
  "depth": 8,
  "limit": 2400
}

Use the returned snapshot ID in browser_act. Replace the selector with one from the actual page:

{
  "snapshot": "<latest snapshot ID>",
  "observe": "summary",
  "steps": [
    {
      "op": "extract",
      "selector": "a.result-title",
      "fields": ["text", "href"],
      "maxRows": 10
    }
  ]
}

These are model-tool arguments, not shell commands. summary does not return a new snapshot; read again before another action call.

06 · Structure and Technical Details

src/
  host.js            Plugin lifecycle, tools and conversation context
  client.js          Native mention and slash-menu sources
  client-entry.js    React UI and icons
  bridge.js          Pairing, extension connection and CDP relay
  engine.js          Snapshots, actions, extraction and screenshots
  annotations.js     Isolated element picker and acknowledged chat drafts
  apps/              App inventory, discovery, approval, installation and execution
  manage.js          Local application and connector management UI
  webmcp.js          Site-tool discovery, routing, validation and calls
  grants.js          Conversation-scoped tab authorization
  protocol.js        Tab references and Remote RPC descriptors
  output.js          Serialized output budgeting
  vendor/            Adapted Playwright browser model
extension/           Manifest V3 extension and popup
scripts/             Build, integration and compatibility checks
test/                Unit tests and fixtures
docs/setup-page.png  Pairing-page screenshot
cordis.patch.yml     DSH plugin registration and defaults
lib/                 Generated build output, not tracked in Git

Snapshots and actions. Full snapshot text is cached locally for pagination. Element references require a current snapshot; reads, actions and navigation can invalidate previous references. Calls are serialized per tab. Cancellation reports completed or uncertain steps rather than promising rollback.

Connection lifetime. Idle connections detach the debugger without closing the tab. A later read reconnects. Restarting the browser changes tab-reference sessions, so select the page again with @.

Data handling. Pairing credentials and authorization records live in the local data directory; screenshots are stored under captures/. Referenced page excerpts enter DSH's model context and may be sent to your configured model provider. Keep pairing files and screenshots private. Purchases, destructive actions and sensitive submissions require explicit user authorization; the plugin does not provide a separate transaction-confirmation dialog.

Development checks

npm run build
npm test
npx playwright-core install chromium
npm run test:browser
npm run verify:dsh -- --asar "<path to DeepSeek Harness app.asar>"

CI is configured for unit tests on Windows and Linux, plus isolated headless Chromium integration on Linux. The badge links to the actual workflow result. Headless checks do not replace end-to-end validation of a daily browser profile and the native DSH UI.

07 · Roadmap / FAQ / License

Roadmap

Version 0.1.3 includes page annotations, app references and MCP-first app control. Further work focuses on real-application compatibility and connector configuration. No release date is promised. Report bugs and discuss changes in Issues.

FAQ

Do I have to open the extension every time? No. Pair once; the extension reconnects in the background. Reload it after updating its files.

Why are some pages missing? Only regular HTTP/HTTPS and authorized local file:// tabs are exposed. Browser-internal and incognito pages are excluded. Check that the extension is enabled and paired if no tabs appear.

What if port 43187 is already in use? Avoid activating two plugin copies on the same port. Stop the duplicate bridge or change the plugin's port setting and reconnect the extension to that port.

What are the current limits? The companion extension is loaded unpacked, not installed from a browser store. Firefox and Safari are not supported. Host API compatibility has been checked against DSH 0.2.0-rc.2; other versions and native UI workflows need their own verification.

Contributing and License

Issues and pull requests are welcome. Include reproduction steps and relevant DSH/browser versions; remove tokens, private page contents and account data from logs. For code changes, run the checks relevant to the affected behavior.

Project code is licensed under MIT. Playwright-derived code and Lucide artwork retain their respective licenses; see NOTICE, Apache-2.0 and ISC.