Hiveround — Search Investment Projects

Purpose

Discover live startup raises on Hiveround — a marketplace where founders post their pitch as markdown and investors point agents at it. Return a list of currently-raising projects (slug, name, stage, sector, raise amount, one-liner, founder handle, listing URL) filtered by keyword, stage, and max raise, plus on-demand the full pitch.md body for any selected project. Read-only; no commitments, no intros sent, no auth required for the discovery path.

When to Use

• Daily / hourly monitoring of new live raises in a given sector or stage band.
• Building an investor pipeline: surface candidate projects, then call get_project for diligence.
• Cross-checking what's "out there" before reaching out to founders directly.
• Any read-only flow where you'd otherwise scrape the marketplace HTML — the site is explicitly designed to be queried by agents and exposes three structured paths (MCP, ECP JSON-LD, ECP Markdown) before you ever need a browser.

Workflow

Hiveround is an agent-first marketplace. The canonical path is the MCP server at https://hiveround.com/api/mcp — read tools (list_projects, search_projects, get_project) are anonymous, no API key required. A second equally-valid path is the Endpoint Context Protocol (ECP) HTTP layer, which serves the same data as JSON-LD (/api/ecp/projects) or agent-friendly Markdown (/api/ecp-md/projects) over plain GET. Use MCP when your runtime speaks JSON-RPC; use ECP when you only have an HTTP client. Do not drive the browser unless you need to sign in or post a project — the rendered /projects page is a thin Next.js shell over the same APIs and pays a ~50× cost premium for the same data.

1. (Optional) Confirm discovery surface

Every public route advertises the agent surface in Link response headers and at well-known paths. To bootstrap when nothing is hardcoded:

GET https://hiveround.com/.well-known/mcp/server-card.json   → MCP transport + tool list + auth
GET https://hiveround.com/.well-known/ecp                    → ECP route map
GET https://hiveround.com/.well-known/agent-skills/index.json → official Hermes skill manifest
GET https://hiveround.com/llms.txt                           → human/agent intro + live-raise summary

The server card declares anonymousReadAllowed: true for the read tools and pins the API-key prefix as hr_sk_ — verify both before assuming you need auth.

2. Search via MCP (recommended)

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": {
        "q": "ai",
        "stage": "prototype",
        "max_raise_usd": 300000,
        "limit": 24
      }
    }
  }'

The MCP search_projects tool is the only layer that honours all three filters together (keyword + stage + max-raise) — the ECP HTTP query string drops max_raise silently (see gotcha). tools/list returns the live JSON schema for every argument; call it once at skill startup if you want to pin types defensively.

For the newest-first sweep without filters, call list_projects with { "limit": N } (default newest first). For diligence on a specific project, call get_project with { "slug": "<slug>" } — that response includes the full pitch.md description plus founder handle + GitHub URL when the founder isn't anonymous.

3. ECP fallback (when you only have HTTP GET)

If your runtime can't speak JSON-RPC, the ECP layer serves the same data over plain GET. Two distinct URLs — pick by what your consumer wants:

GET https://hiveround.com/api/ecp/projects?q=<kw>&stage=<stage>     → JSON-LD Collection
GET https://hiveround.com/api/ecp-md/projects?q=<kw>&stage=<stage>  → Markdown digest
GET https://hiveround.com/api/ecp/projects/{slug}                   → single project JSON-LD
GET https://hiveround.com/api/ecp-md/projects/{slug}                → single project Markdown

The JSON-LD shape per item: { slug, name, one_liner, description, stage, sector, status, logo_url, raise: { amount_usd, instrument, monthly_revenue_usd }, founder: { handle, display_name, github_url, profile_url }, posted_at, updated_at, links: { self, html, collection, mcp } }. The top-level Collection has count, page, page_size (default 24), filters (echoes only the params the server actually applied — useful for verifying acceptance), and items[].

4. Compose results

For each project of interest, emit slug, name, one_liner, stage, sector, raise.amount_usd, founder.handle (or null for anonymous), and links.html (the user-facing listing URL). If the caller wants diligence material, attach the full markdown description from get_project / /api/ecp-md/projects/{slug}.

Browser fallback

Skip unless you're posting a new project, signing in, or the MCP + ECP endpoints are unreachable. If you must browse:

1. browse open https://hiveround.com/projects — the marketplace index.
2. browse get markdown body — extract the listing list as Markdown.
3. For a specific project: browse open https://hiveround.com/projects/<slug>, then browse get markdown body.

Never click a "request intro" button from the browser — that path expects an authenticated session and triggers an intro request. Use the MCP request_intro tool with an API key instead (out of scope for this read-only skill).

Site-Specific Gotchas

• ECP HTTP filters only honor q and stage. Verified by sending max_raise=300000, sector=Fintech, limit=2&offset=2 — none echo back in the filters object on the response, and count is unchanged. The Markdown-route also accepts q/stage but not max_raise. For max-raise / sector filtering, use the MCP search_projects tool (server card explicitly documents these as supported there). The HTTP filters echo is your verification surface — always assert it contains every param you sent before trusting the result set.
• Anonymous reads work on every read tool. Server card sets anonymousReadAllowed: true. The hr_sk_ Bearer key is only required for write tools (request_intro, watch_project, update_watch, list_watches, submit_commitment, acknowledge_commitment, update_commitment_status, send_intro_message, list_commitments, list_intros, read_intro_thread). Read tools — list_projects, search_projects, get_project — work without any headers beyond Content-Type and Accept. Don't waste a sign-in flow for a read-only skill.
• MCP transport is streamable-http (post-only). A GET to /api/mcp returns the server-info envelope (handy for one-shot health checks: server.name, version, tools[], auth). Actual tool calls require POST with JSON-RPC 2.0 body. Include Accept: application/json, text/event-stream to support both single-response and streamed-event responses — the server is allowed to choose.
• Content negotiation on the public /projects route is unreliable from headless clients. The ECP discovery doc says public routes serve Markdown when the caller isn't a browser (Sec-Fetch-Dest: document absent, no text/html in Accept), but real-world HTTP clients (curl, bb fetch, default Next.js fetch) often send Accept patterns that resolve to HTML anyway. Don't rely on content-negotiation against /projects — call the explicit /api/ecp-md/projects or /api/ecp/projects routes, which are content-typed by route, not by header.
• Founder fields can be entirely null when the listing is "anonymous". Both founder.handle and founder.display_name come back as null (not omitted — the keys are present). Treat null as the "no founder linked" branch when emitting; don't fall back to the raise's GitHub URL or the project's domain. Observed on seminara and elastova; populated on watta, hiveround, arispay-executive-summary.
• raise.amount_usd is a raw integer. Listings display "$150k" / "$2.0M" in UI text, but the JSON field is the integer dollar amount (e.g. 150000, 2000000). Format on emit, never re-parse the display string from the Markdown digest — it's already rounded and elides edge cases like odd amounts.
• raise.instrument is a free-text string, not an enum. Observed value across all 5 live raises: "Open to discussion". Don't assume SAFE/priced-round/convertible classifications — surface the raw string and let the agent reason.
• raise.monthly_revenue_usd is almost always null at prototype stage. Present in the schema but populated rarely. Filter on raise.amount_usd for headline raise size; don't use monthly_revenue_usd as a maturity proxy without checking presence.
• Pagination params are page and page_size on the Collection, not limit/offset. limit/offset are silently dropped on the ECP HTTP layer (confirmed — count: 5 returns unchanged with limit=2&offset=2). Use page=2&page_size=24 instead. Default page_size=24. The MCP list_projects / search_projects tools accept limit (per the server card examples in /mcp) — these are different parameter conventions; check the layer you're hitting.
• stage values observed: prototype, launched. Server card doesn't enumerate the full set. Treat anything else returned as a forward-compatible additional value (don't hard-error).
• sector is free-form per posting. Observed: Enterprise SaaS, AI & Agents, Fintech, and absent on watta (the field is omitted entirely, not null). When grouping by sector, treat missing as "Unspecified" rather than excluding.
• status: "open" does NOT appear on the listing-card markdown — only on the JSON-LD shape. The /api/ecp-md/projects digest leaves status off the per-item block (you have to read the prototype · Enterprise SaaS · anonymous · raising $150k summary line and infer). For programmatic status checks, use the JSON-LD route.
• Cloudflare fronts the origin. Every response carries Cf-Cache-Status, Cf-Ray, etc. The dynamic OG-image route (/opengraph-image) is unreliable and frequently returns 502 — don't rely on it for thumbnails. The static listing pages cache fine; the dynamic image route does not.
• No rate-limit observed at low volume, but the /llms-full.txt corpus is 100KB+ and re-fetching it tight-loop is uncivil. Pull it once per session, cache locally; let MCP/ECP handle the incremental queries.
• The /api/mcp GET envelope only lists tool names, not schemas. For actual argument types, call tools/list over POST or read /.well-known/mcp/server-card.json. The server-card lists 14 tools (more than the public landing page's 10) — submit_commitment, acknowledge_commitment, update_commitment_status, list_commitments exist but require auth and are out of scope for read-only search.

Expected Output

{
  "query": { "q": "ai", "stage": "prototype", "max_raise_usd": 300000 },
  "source": "mcp:search_projects",
  "count": 3,
  "page": 1,
  "page_size": 24,
  "projects": [
    {
      "slug": "seminara",
      "name": "Seminara",
      "one_liner": "AI-hosted session platform for education-led sales and onboarding through real-time voice interaction and orchestration.",
      "stage": "prototype",
      "sector": "Enterprise SaaS",
      "status": "open",
      "raise": {
        "amount_usd": 150000,
        "instrument": "Open to discussion",
        "monthly_revenue_usd": null
      },
      "founder": {
        "handle": null,
        "display_name": null,
        "github_url": null,
        "profile_url": null
      },
      "listing_url": "https://hiveround.com/projects/seminara",
      "mcp_get_project_slug": "seminara"
    },
    {
      "slug": "watta",
      "name": "watta",
      "one_liner": "AI workout tracker for rowers. Photograph the erg screen — we read every number off it, score the effort 0–100, and roll training into clubs and leaderboards.",
      "stage": "prototype",
      "sector": null,
      "status": "open",
      "raise": {
        "amount_usd": 250000,
        "instrument": "Open to discussion",
        "monthly_revenue_usd": null
      },
      "founder": {
        "handle": "stevemilton",
        "display_name": "Steve Milton",
        "github_url": "https://github.com/stevemilton",
        "profile_url": "https://hiveround.com/u/stevemilton"
      },
      "listing_url": "https://hiveround.com/projects/watta",
      "mcp_get_project_slug": "watta"
    }
  ]
}

When the caller wants the full pitch for diligence, attach a pitch_markdown field per project, sourced from get_project / /api/ecp-md/projects/{slug}:

{
  "slug": "watta",
  "pitch_markdown": "# watta.\n\nai workout tracker for rowers. take a photo of the erg screen — we\nread every number off it, score the effort 0–100, and roll your\ntraining into clubs and leaderboards.\n\nlive on the app store ..."
}

When the search returns nothing:

{
  "query": { "q": "biotech", "stage": "launched", "max_raise_usd": 100000 },
  "source": "mcp:search_projects",
  "count": 0,
  "projects": []
}