Search Hiveround Investment Projects

Purpose

Discover live startup raises an investor could put money into on Hiveround, optionally filtered by free-text keyword, funding stage, and maximum raise size, and return each match as structured data (slug, name, stage, sector, raise amount, one-liner, listing URL). This is a read-only discovery task — it lists/searches public raises and does not request intros, watch projects, or move money. Hiveround is an agent-native marketplace: it ships an MCP server, an ECP JSON API, and markdown representations specifically so agents can read raises without scraping HTML.

Transport note (Browserless): This is a plain HTTPS JSON / JSON-RPC interface — the curl and GET examples below are canonical and run from any client. Only under restricted egress, route via browserless_function, which executes in a browser page context: page.goto('https://hiveround.com/') first, then page.evaluate a same-origin fetch POST to /api/mcp (same-origin, so it works without CORS). Never route API keys through the browser gratuitously — the anonymous read tools need none.

When to Use

• "Find new projects to invest in on Hiveround" / "What's raising right now?"
• "Show me prototype-stage AI startups raising under $500k."
• "Search Hiveround for fintech raises" or "list the newest live raises."
• Building an investor pipeline: enumerate candidates before doing diligence (get_project returns the full pitch.md per slug).
• Any time you need the structured raise feed rather than a human-readable page.

Workflow

The fastest, most reliable path is the Hiveround MCP server — the site explicitly tells agents to use it instead of crawling (/llms.txt). All read tools are anonymous (no API key).

Recommended: MCP server (POST /api/mcp, JSON-RPC 2.0)

1. Optionally confirm tools with tools/list. The read tools are:
   • list_projects — newest open raises (args: limit≤25, optional stage). No query needed.
   • search_projects — keyword search across name, one-liner, description, sector (args: query required, optional stage, max_raise_usd, limit≤25).
   • get_project — full listing by slug, including the founder's GitHub handle and the entire pitch markdown in description.
2. Call the tool. To search prototype-stage projects matching "AI":

   curl -X POST https://hiveround.com/api/mcp \
     -H "Content-Type: application/json" \
     -H "Accept: application/json, text/event-stream" \
     -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"search_projects","arguments":{"query":"AI","stage":"prototype","limit":10}}}'

   For an unfiltered feed, swap in {"name":"list_projects","arguments":{"limit":10}}.
3. Parse the response. The JSON-RPC envelope's result.content[0].text is itself a JSON string — parse it to get { "projects": [ … ] }. Each project has the fields in Expected Output below.
4. (Optional) drill in with get_project per slug to pull the full pitch markdown for diligence.

stage is one of idea | prototype | mvp | launched | revenue. max_raise_usd is a number in USD.

Alternative: ECP JSON / markdown over plain GET (no POST, no JS)

If you can't make a JSON-RPC POST, the same data is content-negotiable over GET — works through a simple HTTP fetcher and residential proxy:

• Structured JSON: GET https://hiveround.com/api/ecp/projects?q=AI&stage=prototype&max=500000 → application/ecp+json with a Collection whose items[] are full Project objects (same fields as MCP plus the pitch description).
• Markdown list: GET https://hiveround.com/projects?q=AI&stage=prototype with header Accept: text/markdown → a clean markdown digest of matching raises with listing links.
• GET https://hiveround.com/llms.txt is a hand-maintained summary of the marketplace + every live raise's one-liner — good for a quick overview, and llms-full.txt inlines the full corpus.

Browser fallback

Only needed if both HTTP paths are blocked. The human page mirrors the same query params, so you rarely need to script the form. Drive it with a single browserless_agent commands array (the session persists across calls, keyed by proxy/profile):

1. { "method": "goto", "params": { "url": "https://hiveround.com/projects", "waitUntil": "load", "timeout": 45000 } } — or jump straight to https://hiveround.com/projects?q=AI&stage=prototype&max=, since the query params drive the results directly.
2. If using the form: { "method": "type", "params": { "selector": "input[name=\"q\"]", "text": "AI" } }, then { "method": "select", "params": { "selector": "select[name=\"stage\"]", "value": "prototype" } }, then { "method": "click", "params": { "selector": "<the filter button>" } } (confirm the selector via snapshot if it misses).
3. { "method": "snapshot" } or { "method": "text", "params": { "selector": "body" } } to read the result cards. The result count renders as "N LIVE".
4. There is a "VIEW AS JSON" link on the page — following it lands you back on the ECP JSON above, so prefer that over scraping cards.

Note: a goto issues a GET, so it cannot call the MCP endpoint (which needs a POST). To exercise MCP under restricted egress, use browserless_function: page.goto('https://hiveround.com/') first (to establish the origin + network egress), then page.evaluate an async same-origin fetch('/api/mcp', { method: 'POST', headers: { 'Content-Type': 'application/json', 'Accept': 'application/json, text/event-stream' }, body: … }) and return r.json() — same-origin, so no CORS issue.

Site-Specific Gotchas

• Agent-native by design. The homepage Link header and /.well-known/api-catalog advertise: MCP (/api/mcp), ECP (/.well-known/ecp, /ecp.json), an agent-skills index (/.well-known/agent-skills/index.json), an MCP server card (/.well-known/mcp/server-card.json), and /llms.txt. Start from /llms.txt if you're unsure of the interface — it names the MCP tools directly.
• MCP needs the streaming Accept header. The transport is streamable-http; include Accept: application/json, text/event-stream or the POST may be rejected.
• result.content[0].text is double-encoded. MCP tool results wrap the payload as a JSON string inside the JSON-RPC envelope — parse twice.
• search_projects requires query. Calling it without query errors; use list_projects when you want everything.
• Read vs. write auth. list_projects, search_projects, get_project are anonymous. request_intro, watch_project, update_watch, list_watches, and the intro-thread tools require a Bearer API key (hr_sk_*) generated at https://hiveround.com/mcp. This skill only uses the anonymous read tools.
• limit caps at 25 for list_projects/search_projects. There were only ~5 live raises total at capture time, so paging rarely matters, but the ECP GET endpoint also exposes page/page_size for larger catalogs.
• Keyword search is broad. search_projects/q= matches across name, one-liner, description, and sector — so "AI" returned 4 of 5 raises (including ones whose sector isn't literally "AI") because the term appears in their pitch bodies. Don't assume a hit means the sector field equals your query.
• sector and founder.* can be null. Anonymous founders return founder.handle/display_name/github_url = null; some listings have sector = null. Don't treat these as errors.
• Cloudflare fronts the site but does not block agents. robots.txt allows /projects and sets Content-Signal: ai-train=yes, search=yes, ai-input=yes for every named AI agent (incl. ClaudeBot/Claude-User). robots.txt disallows /api/, but /api/mcp and /api/ecp/* are the documented, advertised agent interfaces — that Disallow targets crawlers, not the intended programmatic clients. No captcha, login wall, or 4xx anti-bot pattern was observed across two traced iterations; a residential proxy (proxy: { proxy: "residential" } on the browserless_agent/browserless_function call) was sufficient and full stealth/verification was not required.
• No POST from a goto. Confirmed during iteration 1 — a plain navigation to /api/mcp only returns the server card (a GET). Use curl/fetch, or a browserless_function same-origin fetch POST, for the actual JSON-RPC call.

Expected Output

search_projects / list_projects (after unwrapping result.content[0].text):

{
  "projects": [
    {
      "slug": "seminara",
      "name": "Seminara",
      "one_liner": "Seminara is an AI-hosted session platform for education-led sales and onboarding through real-time voice interaction and orchestration.",
      "stage": "prototype",
      "sector": "Enterprise SaaS",
      "raise_amount_usd": 150000,
      "raise_instrument": "Open to discussion",
      "monthly_revenue_usd": null,
      "url": "https://seminara.online/",
      "logo_url": "https://.../project-logos/.../...png",
      "founder": { "handle": null, "display_name": null, "github_url": null },
      "posted_at": "2026-05-10T20:45:08.865415+00:00",
      "listing_url": "https://hiveround.com/projects/seminara"
    }
  ]
}

A convenient task-level shape to emit to a caller:

{
  "success": true,
  "method": "mcp",
  "query": "AI",
  "stage": "prototype",
  "max_raise_usd": null,
  "count": 4,
  "projects": [
    {
      "slug": "seminara",
      "name": "Seminara",
      "stage": "prototype",
      "sector": "Enterprise SaaS",
      "raise_amount_usd": 150000,
      "one_liner": "Seminara is an AI-hosted session platform…",
      "listing_url": "https://hiveround.com/projects/seminara"
    },
    {
      "slug": "elastova",
      "name": "Elastova",
      "stage": "prototype",
      "sector": "AI & Agents",
      "raise_amount_usd": 250000,
      "one_liner": "AI recovery agent for loose skin after major weight loss.",
      "listing_url": "https://hiveround.com/projects/elastova"
    },
    {
      "slug": "watta",
      "name": "watta",
      "stage": "prototype",
      "sector": null,
      "raise_amount_usd": 250000,
      "one_liner": "ai workout tracker for rowers…",
      "listing_url": "https://hiveround.com/projects/watta"
    },
    {
      "slug": "arispay-executive-summary",
      "name": "ArisPay",
      "stage": "prototype",
      "sector": "Fintech",
      "raise_amount_usd": 2000000,
      "one_liner": "The settlement layer for agentic commerce.",
      "listing_url": "https://hiveround.com/projects/arispay-executive-summary"
    }
  ],
  "error_reasoning": null
}

get_project adds a description field containing the full pitch markdown. On no matches, return count: 0 with an empty projects array (not an error). On failure, success: false with error_reasoning populated from the response.