Playwright MCP
toolLets an agent drive a real browser.
Over MCP or a plain HTTP API, an agent can check what you have already saved, with the sources attached, before it spends time on a web search.
Sopie isn't released yet. The waitlist is free.
Example: an agent asks first
AgentDo we already have something for browser automation?
Sopie: yes, Playwright MCP
An agent without memory reaches for the web every time. With Sopie, it checks your library first.
You ask Claude Code, Cursor or Codex to add browser tests. On its own, it goes straight to the web and picks whatever shows up first.
One plain question: do we already have something for this? Sopie looks for an entry, not for a page of links.
If there is an entry, the agent gets it with its sources and its neighbours. If there is not, it searches the web as usual.
Without Sopie: straight to the webThe agent asks Sopie firstIt works from your entry
ExampleTaskAdd browser tests to the dashboard
MCPsopie.search("browser automation")
Searching the web
Without SopiePlaywright MCP
toolLets an agent drive a real browser.
Send a plain question. Get back a structured entry the agent can act on, with every source attached.
Illustrative only. Tool names, endpoints and fields are not final, and Sopie is not released yet.
Over MCP
Add Sopie to your agent as an MCP server. It shows up as a tool the agent can call before it goes looking elsewhere.
// the agent calls a tool on the Sopie server { tool: "search_entries", arguments: { query: "browser automation", kinds: ["tool", "technique"] } }
Over HTTP
For scripts and tools that do not speak MCP, the same question goes over a plain HTTP API and comes back as JSON.
// the same question, as a plain request POST /v1/search Authorization: Bearer YOUR_KEY Content-Type: application/json { "query": "browser automation" }
What comes back
The agent gets what the tool is, where it came from and what it connects to, so it can act without another search.
// one entry, with its sources { name: "Playwright MCP", kind: "tool", sources: [ { type: "github" }, { type: "reel" }, { type: "docs" } ], needs: ["Node.js"], skill_file: "playwright-mcp/SKILL.md" }
Take any entry out as plain text your agent can load. Your library stays yours, in an open format.
One folder, one SKILL.md. A short description up top, then what the tool is for and where it came from.
--- name: playwright-mcp description: Lets an agent drive a real browser. --- # Playwright MCP Use it when a task needs a real browser: open pages, click, type, read the screen. Needs Node.js Sources GitHub repo, Reel, docs page
Plain text in an open format. Move it, edit it, keep it.
Any client that speaks MCP can connect the same way. Everything else can use the HTTP API.
Add Sopie as an MCP server and let Claude Code check your entries before it searches.
Over MCPConnect the same server in Cursor, so your saved tools show up while you code.
Over MCPPoint Codex at Sopie and it can ask what you already have before it writes something new.
Over MCPAnything that speaks the Model Context Protocol can connect and ask the same questions.
Over MCPFor scripts and tools without MCP: send a question over HTTP and read the entry back as JSON.
Over HTTPJoin the waitlist and we will write to you when Sopie is ready for you.
Thanks. We will write to you when Sopie is ready for you.