BountyRadar MCP

BountyRadar MCP

The judgment layer for the agent-bounty economy. BountyRadar aggregates AI-agent-solvable bounties across Opire, BountyHub, ClawHunt, Algora, GitHub, and validated security programs — then scores each for agent-solvability, freshness, and competition and serves a ranked, agent-consumable feed over MCP.

Positioning: an *intelligence / alerting layer* — NOT a marketplace. No fund custody, no KYC, no cross-border settlement. Public OSS bounties are agent-saturated; the scarce skill is *target selection*, not discovery. BountyRadar is the selection layer.

Languages: English · 简体中文 · 繁體中文

---

English

Current version: 1.0.1

Why this exists

• Public open-source bounties are agent-saturated — a single Algora bounty has drawn 8–158 competing agent attempts within hours.
• The bottleneck is triage / judgment, not discovery — so the moat is the *scoring layer*, not scraping.
• BountyRadar ranks and de-noises so an agent spends compute only on targets it can actually win.

Quick start (hosted MCP endpoint — recommended)

The hosted endpoint is a Streamable HTTP MCP server with Pay-Per-Event billing (no monthly fee). Connect it with any MCP client:

Claude Desktop / Claude Code / WorkBuddy / any MCP client (remote HTTP):

{

"mcpServers": { "bountyradar": { "type": "http", "url": "https://neeenja--bountyradar-mcp.apify.actor/mcp", "headers": { "Authorization": "Bearer " } } } }

initialize and tools/list are free; only value events are charged (see Pricing).

No install step and no npm required — the endpoint *is* the product. One-liner for the common clients:

claude mcp add bountyradar --transport http https://neeenja--bountyradar-mcp.apify.actor/mcp

Cursor / Windsurf / any client that reads .mcp.json — add:

{ "mcpServers": { "bountyradar": { "type": "http",

"url": "https://neeenja--bountyradar-mcp.apify.actor/mcp",

"headers": { "Authorization": "Bearer " } } } }

Run locally (stdio, MIT — clone and self-host)

The full source (Node 22, zero-dependency stdio server) is MIT-licensed and lives on GitHub: https://github.com/PanStories/bountyradar-mcp. Clone it and run:

node src/server.mjs          # zero-dependency JSON-RPC 2.0 stdio server

MCP client config (local stdio):

{ "mcpServers": { "bountyradar": { "command": "node", "args": ["src/server.mjs"] } } }

Tools

Resources: bountyradar://feed/latest, bountyradar://stats, bountyradar://sources Prompt: daily_bounty_brief — briefs an agent on today's best targets.

Scoring (the moat — explainable, deterministic)

freshness      = exp(-ageDays / 14)                    # half-life ~10d

competition = 1 / (1 + competition_count) # 0 tries -> 1.0 agent_solvable = field ?? derive(task_type) # debug/test/validate high reward_score = min(1, log10(reward+1)/log10(5000)) composite = 0.40*solvable + 0.25*freshness + 0.25*competition + 0.10*reward

A free LLM (Cheapest-LLM Router / Cloudflare Workers AI) only *assists* task_type / agent_solvable at refresh time; scoring itself is deterministic and needs no training.

Honeypot guard

Bounties flagged honeypot: true (e.g. "reserved for interview loop") are excluded from the default feed so agents don't waste compute or burn reputation. Toggle with include_honeypot: true.

Monetization — Pay-Per-Event (no monthly fee)

initialize / tools/list / list_sources / get_bounty_detail are free. Charged events:

Local / self-hosted use is free forever.

Price update (floor correction): all charged events raised to $0.005 effective 2026-11-06; the current $0.0005 / $0.0010 prices are grandfathered until then. This is a floor correction (the previous $0.0005 was 10× below the $0.005 PPE floor), not a margin play.

Self-host / deploy (Apify, PPE)

npm install          # pulls express + MCP SDK + apify (optional deps)

apify actor push

Standby mode (120s / 512MB) keeps idle cost near zero. The actor's main process binds the HTTP port itself (Standby-safe) and answers the container readiness probe.

Dev & verification

npm test                       # 11 unit/schema/scoring tests

node scripts/mcp-smoke.mjs # spawn server: initialize -> tools/list -> tools/call node scripts/showcase.mjs # generate feed-samples.html (data-driven) node scripts/refresh.mjs # refresh catalog (graceful; needs APIFY_TOKEN + LLM creds)

Optional add-ons (same catalog)

• Website + subscriber alerts — node website/build.mjs → a zero-dependency static site + Cloudflare Pages Functions (KV subscribe, guarded CSV export); scripts/digest.mjs renders per-subscriber email/Slack digests.
• Live refresh — scraper/ is a standalone Apify actor (or scripts/refresh.mjs locally) that collects real bounty issues and writes to data/live/, never polluting the curated data/bounties/. Set BR_INCLUDE_LIVE=1 to merge at runtime.

Honesty: only issues that are actually bounties (a stated $ reward or a bounty label) are kept; unfunded "good first issues" are noise and excluded. Opire / Algora / BountyHub require platform keys; ClawHunt is a partner_candidate and is intentionally not scraped (index upstream, don't compete).

License

MIT — see the LICENSE file. Free to use the hosted endpoint per the pricing table above.

Privacy

Read-only and stateless — no accounts, no personal data collected. An optional webhook_url is held in memory only; optional classification may send public bounty text to a configured LLM endpoint. See PRIVACY.md.

---

简体中文

Current version: 1.0.1

Agent 赏金经济的"判断层"。 BountyRadar 跨 Opire / BountyHub / ClawHunt / Algora / GitHub 及已核验安全计划聚合"AI agent 可解"的赏金，对每条做 agent 可解性 / 新鲜度 / 竞争度 评分，并通过 MCP 输出一份排好序、agent 可直接消费的 feed。

定位： *情报 / 告警层*——不是市场 / 结算层。不碰资金托管、KYC、跨境结算。公开 OSS 赏金已被 agent 淹没；稀缺能力是"选目标"，不是"发现目标"。BountyRadar 就是这一层。

为什么做这个

• 公开开源赏金已被 agent 农场淹没——单条 Algora 赏金几小时内就涌进 8–158 个抢单 agent。
• 瓶颈在分流 / 判断，不在发现——所以护城河是评分层，不是爬虫。
• BountyRadar 排序 + 去噪，让 agent 只把算力花在真正能赢的目标上。

快速开始（托管 MCP 端点——推荐）

托管端点是按事件付费（PPE）的 Streamable HTTP MCP 服务器，无月费。任何 MCP 客户端均可接入：

Claude Desktop / Claude Code / WorkBuddy / 任意 MCP 客户端（远程 HTTP）：

{

"mcpServers": { "bountyradar": { "type": "http", "url": "https://neeenja--bountyradar-mcp.apify.actor/mcp", "headers": { "Authorization": "Bearer " } } } }

initialize 与 tools/list 免费；仅对价值事件计费（见「商业化」）。

无需安装、无需 npm —— 端点本身就是产品。常见客户端一行搞定：

claude mcp add bountyradar --transport http https://neeenja--bountyradar-mcp.apify.actor/mcp

Cursor / Windsurf / 任何读 .mcp.json 的客户端 —— 加入：

{ "mcpServers": { "bountyradar": { "type": "http",

"url": "https://neeenja--bountyradar-mcp.apify.actor/mcp",

"headers": { "Authorization": "Bearer " } } } }

本地运行（stdio，MIT 开源，可克隆自托管）

完整源码（Node 22、零依赖 stdio 服务器）以 MIT 许可证开源在 GitHub：https://github.com/PanStories/bountyradar-mcp。克隆后运行：

node src/server.mjs          # 零依赖 JSON-RPC 2.0 stdio 服务器

MCP 客户端配置（本地 stdio）：

{ "mcpServers": { "bountyradar": { "command": "node", "args": ["src/server.mjs"] } } }

工具

资源： bountyradar://feed/latest、bountyradar://stats、bountyradar://sources 提示词： daily_bounty_brief（briefing 今日最佳目标）

评分（护城河——可解释、确定性）

freshness      = exp(-ageDays / 14)                    # 半衰期 ~10 天

competition = 1 / (1 + competition_count) # 0 人抢 -> 1.0 agent_solvable = 字段 ?? derive(task_type) # debug/test/validate 高 reward_score = min(1, log10(reward+1)/log10(5000)) composite = 0.40*可解 + 0.25*新鲜 + 0.25*竞争 + 0.10*赏金

免费 LLM（Cheapest-LLM Router / Cloudflare Workers AI）仅在 refresh 时辅助判定 task_type / agent_solvable；评分本身是确定性、无需训练。

蜜罐守卫

标记为 honeypot: true 的赏金（如"面试专用"）默认移出 feed，避免 agent 浪费算力或毁声誉。include_honeypot: true 可切回显示。

商业化——按事件付费（无月费）

initialize / tools/list / list_sources / get_bounty_detail 免费。计费事件：

本地 / 自托管永久免费。

价格更新（地板修正）： 所有计费事件于 2026-11-06 起统一提至 $0.005；现行 $0.0005 / $0.0010 价格祖父保留至该日。这是地板修正（此前 $0.0005 低于 $0.005 地板 10 倍），非涨价牟利。

自托管 / 部署（Apify，PPE）

npm install          # 安装 express + MCP SDK + apify（可选依赖）

apify actor push

Standby 模式（120s / 512MB）让闲置成本接近零。Actor 主进程自行绑定 HTTP 端口（Standby 安全）并响应容器就绪探针。

开发与验证

npm test                       # 11 项单元 / schema / 评分测试

node scripts/mcp-smoke.mjs # 拉起 server：initialize -> tools/list -> tools/call node scripts/showcase.mjs # 生成 feed-samples.html（数据驱动） node scripts/refresh.mjs # 刷新目录（优雅降级；需 APIFY_TOKEN + LLM 凭据）

可选增强（同一份目录数据）

• 网站 + 订阅告警——node website/build.mjs 生成零依赖静态站 + Cloudflare Pages Functions（KV 订阅、受保护的 CSV 导出）；scripts/digest.mjs 渲染按订阅者定制的邮件 / Slack 摘要。
• 实时刷新——scraper/ 是独立的 Apify actor（或本地 scripts/refresh.mjs），采集真实赏金 issue 并写入 data/live/，绝不污染 curated 的 data/bounties/。设 BR_INCLUDE_LIVE=1 可在运行时合并。

诚实声明：只保留真正是赏金的内容（写明 $ 金额或带 bounty 标签）；无赏金的 "good first issue" 是噪音，已排除。Opire / Algora / BountyHub 需平台密钥；ClawHunt 是 partner_candidate，刻意不采集（上游索引，不竞争）。

许可证

MIT——详见 LICENSE。按上表计费使用托管端点。

---

繁體中文

Current version: 1.0.1

Agent 賞金經濟的「判斷層」。 BountyRadar 跨 Opire / BountyHub / ClawHunt / Algora / GitHub 及已核驗安全計畫聚合「AI agent 可解」的賞金，對每條做 agent 可解性 / 新鮮度 / 競爭度 評分，並透過 MCP 輸出一份排好序、agent 可直接消費的 feed。

定位： *情報 / 警報層*——不是市場 / 結算層。不碰資金託管、KYC、跨境結算。公開 OSS 賞金已被 agent 淹沒；稀缺能力是「選目標」，不是「發現目標」。BountyRadar 就是這一層。

為什麼做這個

• 公開開源賞金已被 agent 農場淹沒——單條 Algora 賞金幾小時內就湧進 8–158 個搶單 agent。
• 瓶頸在**分流 / 判