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 <YOUR_APIFY_TOKEN>" }
}
}
}
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
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:
MCP client config (local stdio):
{ "mcpServers": { "bountyradar": { "command": "node", "args": ["src/server.mjs"] } } }
| Tool | Purpose | Price |
|---|
get_bounty_feed | The subscriber feed. Ranked + filtered (category, min_reward, max_competition, source, freshness_min, agent_solvable_min, include_honeypot, sort). | $0.0005 |
search_bounties | Keyword + filter search across title/description/tags/repo. | $0.0010 |
get_bounty_detail | Full record + score breakdown by id. | free |
score_bounty | Raw explainable score breakdown for one bounty. | $0.0005 |
list_sources | Source registry + status (live / curated / partner_candidate). | free |
get_stats | Counts by source / category / reward band. | $0.0005 |
subscribe_feed | Save a filter profile + poll descriptor (stored in server memory; webhook URL echoed, never called). | $0.0005 |
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:
| Event | Price |
|---|
mcp-get-bounty-feed | $0.0005 |
mcp-search | $0.0010 |
mcp-score | $0.0005 |
mcp-stats | $0.0005 |
mcp-subscribe | $0.0005 |
Local / self-hosted use is free forever.
Self-host / deploy (Apify, PPE)
npm install
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
node scripts/mcp-smoke.mjs
node scripts/showcase.mjs
node scripts/refresh.mjs
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 <YOUR_APIFY_TOKEN>" }
}
}
}
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
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:
MCP 客户端配置(本地 stdio):
{ "mcpServers": { "bountyradar": { "command": "node", "args": ["src/server.mjs"] } } }
工具
| 工具 | 作用 | 价格 |
|---|
get_bounty_feed | 订阅 feed 本体。 排序 + 过滤(类型 / 最低赏金 / 最高竞争数 / 来源 / 新鲜度 / 可解性 / 含蜜罐 / 排序方式)。 | $0.0005 |
search_bounties | 关键词 + 过滤搜索(标题/描述/标签/repo)。 | $0.0010 |
get_bounty_detail | 按 id 取完整记录 + 评分拆解。 | 免费 |
score_bounty | 单条可解释评分拆解。 | $0.0005 |
list_sources | 来源注册表 + 状态(live / curated / partner_candidate)。 | 免费 |
get_stats | 按来源 / 类型 / 赏金档统计。 | $0.0005 |
subscribe_feed | 保存过滤档案 + 轮询描述(仅存于服务器内存;webhook URL 只回显、绝不调用)。 | $0.0005 |
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:
| 事件 | 价格 |
|---|
mcp-get-bounty-feed | $0.0005 |
mcp-search | $0.0010 |
mcp-score | $0.0005 |
mcp-stats | $0.0005 |
mcp-subscribe | $0.0005 |
本地 / 自托管永久免费。
自托管 / 部署(Apify,PPE)
npm install
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
node scripts/mcp-smoke.mjs
node scripts/showcase.mjs
node scripts/refresh.mjs
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 / 判斷,不在發現——所以護城河是評分層,不是爬蟲。
- BountyRadar 排序 + 去噪,讓 agent 只把算力花在真正能贏的目標上。
快速開始(託管 MCP 端點——推薦)
託管端點是**按事件付費(PPE)**的 Streamable HTTP MCP 伺服器,無月費。任何 MCP 用戶端均可接入:
Claude Desktop / Claude Code / WorkBuddy / 任意 MCP 用戶端(遠端 HTTP):
{
"mcpServers": {
"bountyradar": {
"type": "http",
"url": "https://neeenja--bountyradar-mcp.apify.actor/mcp",
"headers": { "Authorization": "Bearer <YOUR_APIFY_TOKEN>" }
}
}
}
initialize 與 tools/list 免費;僅對價值事件計費(見「商業化」)。
無需安裝、無需 npm —— 端點本身就是產品。常見客戶端一行搞定:
claude mcp add bountyradar --transport http https://neeenja--bountyradar-mcp.apify.actor/mcp
本機執行(stdio,MIT 開源,可克隆自架)
完整原始碼(Node 22、零依賴 stdio 伺服器)以 MIT 許可證開源在 GitHub:https://github.com/PanStories/bountyradar-mcp。克隆後執行:
MCP 用戶端設定(本機 stdio):
{ "mcpServers": { "bountyradar": { "command": "node", "args": ["src/server.mjs"] } } }
工具
| 工具 | 作用 | 價格 |
|---|
get_bounty_feed | 訂閱 feed 本體。 排序 + 過濾(類型 / 最低賞金 / 最高競爭數 / 來源 / 新鮮度 / 可解性 / 含蜜罐 / 排序方式)。 | $0.0005 |
search_bounties | 關鍵字 + 過濾搜尋(標題/描述/標籤/repo)。 | $0.0010 |
get_bounty_detail | 按 id 取完整記錄 + 評分拆解。 | 免費 |
score_bounty | 單條可解釋評分拆解。 | $0.0005 |
list_sources | 來源註冊表 + 狀態(live / curated / partner_candidate)。 | 免費 |
get_stats | 按來源 / 類型 / 賞金檔統計。 | $0.0005 |
subscribe_feed | 儲存過濾檔案 + 輪詢描述(僅存於伺服器記憶體;webhook URL 僅回顯、絕不呼叫)。 | $0.0005 |
資源: bountyradar://feed/latest、bountyradar://stats、bountyradar://sources
提示詞: daily_bounty_brief(briefing 今日最佳目標)
評分(護城河——可解釋、確定性)
freshness = exp(-ageDays / 14) # 半衰期 ~10 天
competition = 1 / (1 + competition_count) # 0 人搶 -> 1.0
agent_solvable = 欄位 ?? derive(task_type) # debug/test/validate 高
reward_score = min(1, log10(reward+1)/log10(5000))
composite = 0.40*可解 + 0.25*新鮮 + 0.25*競爭 + 0.10*賞金
免費 LLM(Cheapest-LLM Router / Cloudflare Workers AI)僅在 refresh 時輔助判定 task_type / agent_solvable;評分本身是確定性、無需訓練。
蜜罐守衛
標記為 honeypot: true 的賞金(如「面試專用」)預設移出 feed,避免 agent 浪費算力或毀聲譽。include_honeypot: true 可切回顯示。
商業化——按事件付費(無月費)
initialize / tools/list / list_sources / get_bounty_detail 免費。計費事件:
| 事件 | 價格 |
|---|
mcp-get-bounty-feed | $0.0005 |
mcp-search | $0.0010 |
mcp-score | $0.0005 |
mcp-stats | $0.0005 |
mcp-subscribe | $0.0005 |
本機 / 自架永久免費。
自架 / 部署(Apify,PPE)
npm install
apify actor push
Standby 模式(120s / 512MB)讓閒置成本接近零。Actor 主程序自行繫結 HTTP 埠(Standby 安全)並回應容器就緒探針。
開發與驗證
npm test
node scripts/mcp-smoke.mjs
node scripts/showcase.mjs
node scripts/refresh.mjs
選用增強(同一份目錄資料)
- 網站 + 訂閱警報——
node website/build.mjs 產生零依賴靜態站 + Cloudflare Pages Functions(KV 訂閱、受保護的 CSV 匯出);scripts/digest.mjs 渲染按訂閱者定製的郵件 / Slack 摘要。
- 即時更新——
scraper/ 是獨立的 Apify actor(或本機 scripts/refresh.mjs),採集真實賞金 issue 並寫入 data/live/,絕不污染 curated 的 data/bounties/。設 BR_INCLUDE_LIVE=1 可在執行時合併。
誠實聲明:只保留真正是賞金的內容(寫明 $ 金額或帶 bounty 標籤);無賞金的 "good first issue" 是雜訊,已排除。Opire / Algora / BountyHub 需平台金鑰;ClawHunt 是 partner_candidate,刻意不採集(上游索引,不競爭)。
授權
MIT——詳見 LICENSE。按上表計費使用託管端點。