Show HN: I built a tiny search engine for my notes
312 points · 84 comments
01SIMPLE
DeepDeck is an open-source, local AI workbench.
The desktop client for DeepSeek Harness.
Workspaces, modes, models, and the prompt all fall into one natural flow. DeepDeck keeps existing Harness settings and extensions compatible; after guided first-run setup, it is ready for work.
Now available: Browser + WebMCP
02WEBMCP
Available now
Let agents learn a site. Keep what works.
You set the goal. The Agent explores and uses the website, then saves verified workflows as WebMCP tools for future tasks. No need to teach it every click: existing products become easier for agents to work with.
Download latestFor Apple Silicon and Intel Macs.
REAL TASK · FIGMA
Figma Agent declined the task. DeepDeck + WebMCP did it.
Create a new page and make the OpenAI logo. For this task, Figma Agent said it couldn't edit vector shapes or paths yet. DeepDeck + WebMCP completed it.
The task: create a new page and make the OpenAI logo
Vector editing not supported yet
Figma Agent suggested using the Draw tool in the toolbar.
Editable vector logo created
A 360 × 360 vector graphic on a newly created OpenAI page.
A record of this task. DeepDeck used WebMCP alongside browser tools during execution.
Use the website. Save the experience.
Asking an Agent to build WebMCP is like asking it to use a website and preserve what it learned: how to search, read results, or edit and check a draft. That experience lives in inspectable, executable tools future tasks can reuse. You provide the goal; the Agent discovers the method, verifies the result, and keeps the operations that work.
- Explore the site
- Try the workflow
- Verify the result
- Save as tools
- Reuse next time
What is WebMCP?
WebMCP is a proposed web API: websites expose JavaScript functions or HTML forms as tools with natural-language descriptions and structured parameters. Agents can discover what a site does, understand the inputs it needs, and call tools in the current page, sharing the user's interface and browser context.
Explore the WebMCP projectWebMCP vs. ordinary computer use
With ordinary computer use, an Agent reads the page, finds controls, then clicks and types step by step. WebMCP declares website features as tools, so the model knows what it can do, which arguments to provide, and how to call them.
| Compare | Ordinary computer use | WebMCP |
|---|---|---|
| Discovery | Infer what controls do, the page state, and the workflow from screenshots or page structure. | Discover explicit capabilities through tool names, descriptions, and input schemas. |
| Execution | Observe → locate a control → click or type → observe again, one step at a time. | Choose a tool → supply arguments → read its result. The tool handles the corresponding page operations. |
| Reuse | Typically repeat the UI steps; reuse requires separately saving a script or workflow. | DeepDeck saves verified operations as tools, so later tasks can reuse that experience. |
| Time and tokens | Repeated page reads and model decisions add context and interaction overhead. | For tasks covered by tools, fewer page reads, control lookups, and interaction rounds can save time and tokens. |
| Best suited to | Exploring a website, working without existing tools, or handling a one-off page interaction. | Using tools a site already offers, or saving recurring workflows as tools for later tasks. |
Same task: search for a keyword and read the results
- Ordinary computer use
- Read the page → find the search field → click and type → submit → read the page again → extract results
- WebMCP
- Call the search tool with a keyword → read its result; call a results-reading tool if the tool design requires it
DeepDeck uses both approaches together: the Agent first explores and verifies the website through browser interaction, then saves working operations as WebMCP tools. Later tasks reuse those tools, with browser interaction available for anything they do not cover.
Building tools first takes exploration and verification. Time and token savings depend on the site, tool design, and task; site changes may require tool updates.
- 01
Reuse existing WebMCP
Open a website and find its Website tools in Site Agent’s WebMCP tab. Switch to Use, describe your task, and the Agent can call the capabilities the site already provides.
- 02
Quickly add WebMCP to a website
No tools yet? Open WebMCP Builder or switch to Builder and describe your goal. The Agent explores the site, tries its workflows, and saves verified operations as tools. No step-by-step manual, site source changes, or separate MCP server required.
Discover tools the website already offers
On openai.com, DeepDeck automatically discovers search_openai and lists it under Website, ready for the Agent to use.
The Agent explores, uses, and verifies
On X, set a goal such as search, post reading, or draft editing. Builder explores the live page, tries the workflows, and checks which operations accomplish the task.
Save the experience as reusable tools
This X example saves verified operations as 23 tools under Built with DeepDeck. Enabled tools load when you return, with source inspection and version rollback available.
Switch to Use and describe your task
Ask “compose a hello world x post” and the Agent calls WebMCP to prepare a Hello, world! 👋 draft. The post is not published; filling and submitting are separate actions.
03APP SUPPORT
Supports Apps.
A DeepDeck App is an installable extension with its own window, settings, and AI conversations. Hacker News Reader is a complete example.
Reading, search, sign-in, and AI Actions such as Explain and Summarize all live inside the same App.
04APP SHOWCASE
Apps in action.
An App brings one focused way of working into DeepDeck: a dedicated window for the interface, a Workspace for context, and an Agent at the moments that need reasoning.
READING + RESEARCHINSTALLABLE
Hacker News Reader
Browse live feeds, search stories, follow complete discussion trees, and ask the Agent to explain or summarize any story, comment, or selection.
- Live feeds
- Full threads
- Explain / Summarize
LONG-RUNNING INVESTIGATIONINSTALLABLE
Video Sherlock
Turn long-form video investigation into a trackable case, watch work progress, and return to key moments on an evidence timeline.
- Visible progress
- Evidence timeline
- Structured artifacts
Each App can own its settings, credentials, canonical Harness Sessions, and a source-bound Creator Workspace.
05INSTALL
Install an App in minutes.
No terminal required. DeepDeck inspects the source, package identity, and build command first, then installs only after you approve the plan.
- 01Download DeepDeck
Install the latest release and complete first-run setup.
- 02Open Apps settings
Go to Settings → Apps and find the install control.
- 03Choose and paste a source
Pick either open-source App below and paste its Git address.
- 04Inspect and confirm
Choose Inspect source, review the build plan, then Confirm install and restart when prompted.
APP SOURCE
Choose an open-source App and paste its Git address into Install an App plugin
After restart, open the newly installed App from the Apps sidebar.
BUILD YOUR OWN
Develop from source—or make an App of your own.
Settings → Apps → New App generates a complete Cordis Host / Client starter. Developers can also build and mount a public App source manually.
Read the Apps protocolDeveloper: manual install example
git clone https://github.com/jo32/dsh-hackernews-reader.git
cd dsh-hackernews-reader
pnpm install --frozen-lockfile
pnpm check && pnpm test && pnpm build
dsh plugin --profile web add "$PWD"Run the final command against a DeepDeck-managed web profile so the required App runtime is present.
06VIBE APP
Vibe an App.
Vibe Coding is DeepDeck's AI-assisted app-building workflow. Open it from Hacker News settings to bind the source workspace; describe a change, and the Agent writes code while Bun and Cordis rebuild and hot-reload it.
- 1Open CreatorThe source workspace is bound automatically
- 2Describe the ideaThe Agent edits the App directly
- 3Build and hot-reloadNo DeepDeck restart required
THAT'S DEEPDECK.Work simply.
Work simply.
Turn ideas into Apps.
Work, browse, and build. One desktop.
Download DeepDeck07CHANGELOG
What’s new in DeepDeck.
New capabilities, meaningful improvements, and how to use them. Development previews and published features are labeled separately.
V1.0.38 / BROWSER / WEBMCP
Browser + WebMCP is now available.
Use the Site Agent and WebMCP Builder in v1.0.38 to turn reading, search, editing, and sign-in workflows into reusable website tools.
- Signed and notarized installers for Apple Silicon and Intel Macs, with in-app updates for existing users.
- Builder discovers inputs, drafts, and sign-in controls. The Agent reads, edits, fills, and checks, with submission as a separate action.
- Detect draft and target changes before editing. The Agent continues through native browser input when needed and verifies the page result.
BROWSER / WEBMCP
A dedicated Agent for your websites.
Browser and WebMCP Builder bring the conversation alongside the page, turning discovered website capabilities into tools you can use again.
- Each website keeps its own Agent conversation, with Use and Builder in the same thread.
- Discover existing WebMCP tools first. Builder creates missing tools with typed parameters; Apply verifies registration and saves them for that site.
- Cover sign-in, search, and draft editing. The Agent reads, edits, fills, and checks the result, with filling separate from submitting.
This entry describes the source preview before its installer release.