swarm-mcp
An MCP server that exposes Swarm's fact reconciliation to AI agents, allowing them to query scope, evidence gaps, and review requirements without being able to issue a final verdict.
README
swarm-mcp
An MCP stdio server that exposes Swarm's read + reconcile facts to agent clients (Claude Desktop, Cursor) — so an agent mid-task can ask Swarm "what's my scope?", "what evidence is missing?", "what should the reviewer not miss?" — without being allowed to declare itself done.
What it is — and what it is not
swarm-mcp is a thin adapter over the swarm CLI's --json contract. It spawns swarm <cmd> --json
with fixed arguments and reshapes the output into MCP tools, resources, and prompts. It does not
import swarm-cli's internals, run a model loop, write durable artifacts, or issue a verdict.
- Reconcile-only. Every tool result carries
noVerdictIssued: true; resources serve the CLI's--jsondata verbatim (also no verdict, just unwrapped). swarm-mcp surfaces facts (coverage gaps, out-of-scope changes, empty-evidence Pass rows, self-report mismatches) and a derived human-attention list; a human or an independent reviewer owns the Pass / Fail / Unverified / Blocked result. An empty or weak Evidence cell reads Unverified regardless of a clean reconcile. - Root-confined. It only reads inside a configured workspace root; every client-supplied input is
validated before any subprocess runs. File paths are realpath-confined (no
.., no absolute escapes, no symlink escapes); ids/stems must be a single safe segment; a git base must be a flag-free ref. - Many libraries, not a framework. It couples to swarm-cli only through the public
--jsoninterface, so swarm-cli keeps its minimal footprint and each piece stays useful on its own.
Run it
// Claude Desktop / Cursor MCP config
{
"mcpServers": {
"swarm": {
"command": "swarm-mcp",
"args": ["--workspace", "/path/to/your/swarm-workspace"]
}
}
}
Config: --workspace <path> / SWARM_WORKSPACE (the workspace root); --swarm-bin <path> / SWARM_BIN
(the swarm binary, default swarm on PATH). Requires the swarm CLI
installed.
The swarm-mcp command above resolves to this package's bin. To install from source until a published
build is available:
git clone https://github.com/jcosta33/swarm-mcp && cd swarm-mcp
pnpm install && pnpm build && npm link # exposes `swarm-mcp` on PATH (runs the built dist/)
Node: a published/built install needs Node ≥ 18.18; running from a source checkout (no dist/) needs
Node ≥ 22.6 (it strips types at runtime).
v0 surface (read-only)
- Tools (10). Reconcile/check:
swarm_get_status,swarm_check_workspace,swarm_check_file,swarm_scan_task,swarm_reconcile_review,swarm_validate_review_packet. Parsed-artifact loaders:swarm_get_task,swarm_get_spec,swarm_get_review,swarm_get_checks. - Resources (7). Fixed:
swarm://workspace,swarm://status,swarm://checks. Templated:swarm://tasks/{id},swarm://specs/{id},swarm://reviews/{id},swarm://findings/{id}. - Prompts (5).
swarm_task_briefing,swarm_before_done(the implementer — may not approve its own work),swarm_review_assistant(an independent reviewer — falsify, don't trust),swarm_evidence_rule,swarm_finding_candidate. The before-done / review-assistant asymmetry is deliberate: no prompt grants verdict authority.
Develop
pnpm install
pnpm gate # typecheck + lint + coverage (thresholds enforced) + build
Status: v0 — the full read + reconcile surface over the swarm CLI's --json contract. It couples to
swarm-cli only through that public contract (recorded in
swarm ADR-0085),
so swarm-cli keeps its minimal footprint and each piece stays useful on its own.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。