Hivemind Fa2e27c4

MCPAI代理

提供方 AmirK-S · 包名 hivemind-mcp

AI代理 stdio · 本地进程 无需鉴权 社区 / 第三方 Claude DesktopClaudeCursorWindsurf
查看接入方式

HiveMind是一个AI代理的共享记忆系统,通过MCP协议连接代理,共享知识并实时更新,适用于需要多代理协作的场景。

一句话选型结论

按资料字段整理

Hivemind Fa2e27c4 属于AI代理类 MCP 服务器,stdio · 本地进程,无需鉴权,可直接接入 Claude Desktop、Claude、Cursor 等,有可直接执行的安装命令。接入前先确认权限范围;涉及写入、删除或付款的操作,建议先在测试环境验证。

分类AI代理 传输方式stdio · 本地进程 鉴权无需鉴权 部署方式本地部署
GitHub 星标 1 仓库热度,无公开仓库时显示未标注
工具数 6 暴露给模型的工具数量
资源 / 提示词 0 / 0 资源与提示词模板数量
维护方 社区 / 第三方 核验等级 仅资料核对

工具与能力

AI代理
数量来自来源目录的标注,具体工具清单以官方文档为准
工具数量6暴露给模型可调用的工具个数
资源数量0可读取的资源条目
提示词数量0内置提示词模板数量

能力标签

支持客户端

适用场景

AI代理Python本地部署共享记忆知识共享实时更新MCP协议

安装与配置

stdio · 本地进程 无需鉴权
客户端在本机拉起子进程,数据不出本机
npx hivemind-mcp https://your-hivemind-instance.com your-api-key

这条命令只是入口,具体启动参数请按官方文档补齐,不要直接照搬到别的服务器。

  1. 1
    准备运行环境

    stdio 服务在本地起进程,需要先装好 Node.js,再按安装命令拉取依赖。

  2. 2
    安装服务

    复制上面的安装命令执行;命令里带占位符的位置,换成自己的密钥或路径。

  3. 3
    写入客户端配置

    在客户端的 MCP 配置里填好启动命令和参数,保存后完全退出客户端再打开,让新配置生效。

  4. 4
    验证是否可用

    连接成功后客户端会列出约 6 个工具,随便调一个只读工具确认能返回结果。

典型使用场景

3 类
按分类与标签整理的常见用法,具体以官方文档为准
给代理接外部能力

让代理在任务中途调用搜索、数据库等真实工具。

多步任务编排

把一串操作拆成步骤,由代理按顺序执行。

自动化重复流程

把固定流程交给代理跑,减少手工操作。

选择建议

选之前先看这几条
按条目字段给出的判断,不替你决定用不用

值得选的地方

  • 资料标注 6 个工具,职责相对集中,接入后不容易误操作。
  • 有现成的安装命令,不用自己找入口,照着跑就能起服务。
  • 已适配 Claude Desktop、Claude、Cursor,换客户端不用重新研究配置。
  • 资料标注无需鉴权,装好就能连。

先想清楚的地方

  • stdio 方式要求本地装好运行时(如 Node.js),换机器要重新装一遍。
  • 版本更新与行为细节以官方仓库为准。

更适合

  • 需要在AI代理场景里让模型直接调用外部能力的人。
  • 已经在用 Claude Desktop、Claude 的用户,接入路径最短。
  • 想先快速试一下、能接受在本地或自有环境跑服务的人。

不太适合

  • 完全不想在本地装环境的人。
  • 需要厂商级 SLA 和技术支持的生产环境。

同类对比

6 个同类
同分类里挑最接近的几条,字段以各自详情页为准
MCP 服务器工具数传输方式部署方式安装命令最后核验
Hivemind Fa2e27c4当前 6 stdio · 本地进程 本地部署 有 2026-05-17
Filesystem 560c4363 13 未标注 本地部署 无 2026-05-17
Server Everything (everything) 11 stdio · 本地进程 本地部署 有 2026-05-17
Time 7365cb6d 2 SSE · 服务端推送 远程部署 有 2026-05-18
Everything Claude Code 未标注 stdio · 本地进程 本地部署 有 2026-05-17
Openmemory 557e1585 未标注 stdio · 本地进程 本地部署 有 2026-05-17

服务器档案

AI代理
字段来自项目仓库与公开索引,中文说明由本站标注
提供方AmirK-S上游标注的维护方 AmirK-S
分类AI代理本站按上游分类展示
连接方式stdio · 本地进程客户端在本机拉起子进程,数据不出本机
鉴权方式无需鉴权远程服务建议用最小权限凭据
维护方社区 / 第三方核验等级 仅资料核对
包名 / 运行时hivemind-mcp · Node.js本地安装时用到的包标识与运行时
部署类型本地部署上游给出的部署形态
最近核验2026-05-17上游最后核验时间

详细介绍

11.4 KB
来自项目仓库与上游目录的原文,未做删改

HiveMind

Shared memory for AI agents, served over the Model Context Protocol. One agent
contributes what it learned, another one finds it. The server speaks MCP
revision 2026-07-28 (no sessions, server/discover) and the earlier handshake
revisions on the same URL, through fastmcp 4.0.3.

This is a demonstration pinned to a date, not a product. It has no hosted instance, no
users and no daily maintenance. The end of life section says
what will break next. Everything stated here can be checked from the repository
alone.

What it does

  • Seven MCP tools over Streamable HTTP at /mcp: contribute, search, list,
    delete, publish, report an outcome, manage roles.
  • Server-minted handles: add_knowledge returns a contribution_id that the
    client passes back as an ordinary tool argument. No protocol session is
    needed; the 2026-07-28 revision removed sessions and expects exactly this.
  • Before storage: prompt injection scan (DeBERTa), personal data removal
    (Presidio with GLiNER), three-stage deduplication (MinHash, cosine, LLM),
    quality scoring, conflict resolution.
  • Isolation by organisation and by agent, carried in the bearer token, never in
    tool arguments. RBAC with Casbin.
  • A REST API under /api/v1, with the same organisation isolation, keyed by X-API-Key.

What it does not do

  • It does not run anywhere. You run it.
  • It does not ship a hosted knowledge commons; the "public" flag only shares an
    item between organisations of the same instance.
  • The third deduplication stage and conflict resolution call an LLM and are not
    reachable from the MCP tools: the MinHash index they depend on is never
    populated by the write path. They are kept as code, documented as unwired.
  • docker-compose.yml starts the server, PostgreSQL and Redis. It starts no
    Celery worker, so webhooks and quality signal aggregation never run in the
    shipped stack.
  • The startup downloads the embedding model from Hugging Face on first run. A
    machine without network access does not start.
  • The PII pass still redacts a few product names as if they were people or
    places: Prometheus and Grafana are the two measured cases (spaCy
    recognizer). Known, documented, not corrected in this release.
  • A2A facade: not planned in this release.

Quick start

Requirements: Docker with Compose. Ports 8000, 5432 and 6379 free.```bash
git clone https://github.com/AmirK-S/HiveMind.git
cd HiveMind
cp .env.example .env
docker compose up -d --build
curl -s localhost:8000/health


The first start takes a few minutes: migrations run, then three models load,
one of them downloaded from Hugging Face on first run. `docker compose logs -f
hivemind` shows the progress; the health check allows four minutes before it
reports the container unhealthy.

Mint a bearer token for an organisation and an agent. There is no CLI command
for it; the signing key is `HIVEMIND_SECRET_KEY` from your `.env`:```bash
docker compose exec hivemind python -c \
  "from hivemind.server.auth import create_token; print(create_token('org-a', 'agent-1'))"

Connect an MCP client that speaks Streamable HTTP. The generic shape, to adapt
to your client's configuration file:```json
{
"mcpServers": {
"hivemind": {
"type": "http",
"url": "http://localhost:8000/mcp",
"headers": { "Authorization": "Bearer <token>" }
}
}
}


Clients that only speak stdio can go through `mcp-remote` with a `--header`
argument. There is no `npx` launcher in this repository, and the server is not
published on PyPI: it needs PostgreSQL with pgvector and Redis, so it is
distributed as this repository, to build locally. The distribution name in
`pyproject.toml` is `hivemind-mcp`; `hivemind` on PyPI is an unrelated project.

Talk to the server by hand, on the 2026-07-28 wire:```bash
curl -s -X POST localhost:8000/mcp \
  -H &#x27;Content-Type: application/json&#x27; -H &#x27;Accept: application/json, text/event-stream&#x27; \
  -H &#x27;MCP-Protocol-Version: 2026-07-28&#x27; -H &#x27;Mcp-Method: server/discover&#x27; \
  -d &#x27;{&quot;jsonrpc&quot;:&quot;2.0&quot;,&quot;id&quot;:1,&quot;method&quot;:&quot;server/discover&quot;,&quot;params&quot;:{&quot;_meta&quot;:{&quot;io.modelcontextprotocol/protocolVersion&quot;:&quot;2026-07-28&quot;,&quot;io.modelcontextprotocol/clientCapabilities&quot;:{}}}}&#x27;

Demo

scripts/demo.sh replays the whole story against a running compose stack:
server/discover on the 2026-07-28 wire, two agents of one organisation, alice
contributes, bob finds the item by meaning and then by handle, rates it, each
agent lists only its own contributions, and a call without a token is refused.
scripts/demo-transcript.txt is its output on a fresh clone on 2026-09-06.```bash
cp .env.example .env && docker compose up -d --build && ./scripts/demo.sh


## The seven tools

| Tool | What it does | Refuses when |
| --- | --- | --- |
| `add_knowledge` | Contributes a text with a category; returns `contribution_id` and `status` (`queued`, or `auto_approved` if a rule exists for the organisation) | content shorter than 10 characters, confidence outside 0.0 to 1.0, unknown category, injection detected, too much redacted, burst rate exceeded |
| `search_knowledge` | Fetches by `id`, or searches by `query` (PostgreSQL full text plus vectors, fused in SQL) | neither `id` nor `query`, unknown or foreign private item |
| `list_knowledge` | Lists the caller&#x27;s own contributions, pending and approved | unknown status or category filter |
| `delete_knowledge` | Soft-deletes an item of the caller&#x27;s organisation and agent | item of another agent or organisation, unknown id |
| `publish_knowledge` | Makes an item visible to other organisations, or hides it again | item of another organisation, unknown id |
| `report_outcome` | Records `solved` or `did_not_help` for an item, once per `run_id` | unknown item, unknown outcome |
| `manage_roles` | Reads or assigns Casbin roles and permissions | caller is not an admin of the organisation |

Every refusal is an MCP error (`isError: true`) with a one-line message. Every
tool reads `Authorization: Bearer `; the JWT carries `org_id` and
`agent_id`. The REST API under `/api/v1` uses `X-API-Key` instead.

The RBAC policy table starts empty. Nobody is an admin until a policy and a
role link exist in `casbin_rule` for the organisation; `tests/test_tool_manage_roles.py`
shows the two rows to insert.

## Protocol revision and conformance

The official suite, `@modelcontextprotocol/conformance`, was run before and
after the upgrade from fastmcp 2.14.5 to 4.0.3. Reports, logs and the
`expected-failures` files are in `conformance/`.

| Requirements | Before (2.14.5) | After (4.0.3) |
| --- | --- | --- |
| `2025-11-25`, scored scenarios passed | 8 of 30 | 10 of 30 |
| `2026-07-28`, scored scenarios passed | 5 of 37 | 13 of 37 |
| `2026-07-28`, checks passed / failed | 55 / 102 | 111 / 56 |
| Wire schema violations | 0 | 0 |

Most of the suite drives reference tools with imposed names (`test_simple_text`,
`test_input_required_result_*`) or capabilities this server does not expose
(resources, prompts, completion, the tasks extension). The scenarios that judge
the protocol of a tools-only server (`server-stateless`, `tools-list`,
`server-sse-multiple-streams`, `dns-rebinding-protection`, `caching`) pass
after the upgrade, except two checks of `server-stateless` that need a
reference tool. Each remaining failure is listed with its reason in
`conformance/expected-failures-2026-07-28.yaml`, and the tool exits 0 with it.

Almost none of the upgrade happened in this repository&#x27;s code. The revision is
absorbed by the framework. Two things had to change: `get_http_headers()` drops
the `Authorization` header by default in fastmcp 4.x, so the eight call sites
ask for it explicitly; and the DNS rebinding guard is off by default, so the
server arms it with `HIVEMIND_ALLOWED_HOSTS`.

## Tests```bash
uv sync --extra dev
docker run -d --name hivemind-test-pg -e POSTGRES_USER=hm -e POSTGRES_PASSWORD=hm \
  -e POSTGRES_DB=postgres -p 55432:5432 pgvector/pgvector:pg16
HIVEMIND_TEST_ADMIN_URL=postgresql://hm:hm@localhost:55432/postgres uv run pytest -q -m &quot;not models&quot;

69 tests in that run. The three models are replaced by doubles, so none of
them downloads anything or calls an LLM. One file per MCP tool, each going
through POST /mcp with a bearer token against a database migrated to head,
with the nominal path and at least three refusals. Without
HIVEMIND_TEST_ADMIN_URL the tests that need PostgreSQL skip. CI runs the same
command on every push (.github/workflows/ci.yml).

29 more tests carry the models marker: they load the real PII pipeline, so
GLiNER (about 400 MB, downloaded to ~/.cache/huggingface) and the spaCy model
of the models dependency group. They run separately:bash uv sync --extra dev --group models uv run pytest -q -m models

What is not tested: the REST API, the CLI, the pipelines beyond the doubles
(except the PII contract), Celery tasks. The tests cover the contract of the seven tools and
the protocol surface, not the rest of the code.

Configuration

All variables carry the HIVEMIND_ prefix and have a default in
hivemind/config.py; .env.example lists the useful ones.

Variable Default Role
HIVEMIND_SECRET_KEY dev-secret-change-me HS256 key of the bearer tokens. Change it.
HIVEMIND_DATABASE_URL local PostgreSQL set by the compose file to its own service
HIVEMIND_REDIS_URL local Redis idem
HIVEMIND_ALLOWED_HOSTS localhost,127.0.0.1 hostnames accepted in Host; others get 421
HIVEMIND_EMBEDDING_MODEL sentence-transformers/all-MiniLM-L6-v2 must not change for the life of a database
HIVEMIND_ANTHROPIC_API_KEY empty LLM stages; empty means skipped

Repository layout

  • hivemind/: the server. server/main.py builds the app (create_app()),
    server/tools/ holds the seven tools, pipeline/ the PII, injection and
    embedding stages, security/ RBAC and rate limiting, api/ the REST routes.
  • alembic/: seven migrations. docker/entrypoint.sh applies them before uvicorn.
  • tests/: the suite described above.
  • conformance/: MCP conformance reports before and after the upgrade.
  • wrappers/: hivemind-langchain and hivemind-crewai, thin clients of the
    REST API, also on PyPI under those names.
  • scripts/: the demo, its transcript, and a script that exports the OpenAPI
    document of the REST API.

Removed in September 2026, still in the git history: a Next.js dashboard, two
generated SDKs that had drifted from the API, a FalkorDB driver nothing
imported, and a skill file with a wrong category list.

End of life

This repository serves MCP revision 2026-07-28 with fastmcp&gt;=4.0.3,&lt;5 and
the mcp 2.1.1 SDK, pinned in uv.lock. It is not maintained on a daily basis.

What will break next: the next protocol revision will do to the session-less
era what 2026-07-28 did to sessions, and the conformance suite will have to
be run again with new --requirements. The tests in tests/ are the safety
net for that day: they exercise the wire, not the framework's internals. The
official TypeScript SDK did not speak 2026-07-28 as of 2026-09-05 (1.30.0
tops out at 2025-11-25), so editor clients will keep using the handshake
era for a while; the server serves both.

Issues

Issues are welcome and are read, on a best effort basis, with no response time
commitment: this is a personal demonstration repository pinned to a date, kept
by one person, with no service expected of it. No feature is promised. A
reproducible defect in what this README claims is the kind of issue that gets
fixed; a request for a new capability gets a written answer and is closed as
out of scope.

License

MIT, see LICENSE.

常见问题

8 问
答案由本站依据条目字段整理,不引用上游原文
Hivemind Fa2e27c4 怎么安装?

在客户端按下面的安装命令接入:npx hivemind-mcp https://your-hivemind-instance.com your-api-key。对应的包名是 hivemind-mcp。装完在客户端配置里引用同一个包或端点。

Hivemind Fa2e27c4 怎么在 Claude Desktop、Claude、Cursor 里配置?

它的传输方式是stdio · 本地进程,在客户端的 MCP 配置里填好启动命令和参数,保存后完全退出客户端再打开,配置才会生效。

Hivemind Fa2e27c4 收费吗?

本站不代收费用,也没有标注具体价格。是否收费取决于服务本身和背后的接口额度,正式使用前按官方说明确认。

Hivemind Fa2e27c4 需要 API Key 吗?

资料标注为无需认证,装好就能连;如果实际调用时报鉴权错误,按官方文档补上凭据。

Hivemind Fa2e27c4 支持哪些客户端?

资料标注支持 Claude Desktop、Claude、Cursor、Windsurf、VS Code;其他客户端只要能配置 MCP Server,也能用同样的命令或地址接入。

Hivemind Fa2e27c4 能提供哪些能力?

资料显示约 6 个工具;能力标签包括 AI代理、Python、Claude、实时更新。具体工具清单以官方文档为准。

Hivemind Fa2e27c4 需要什么运行环境?

运行时为 Node.js;属于本地进程,需要在本机装好运行环境。

Hivemind Fa2e27c4 是本地还是远程?

它是本地进程(stdio),由客户端在本机拉起。传输方式为stdio · 本地进程。

热门专题

按标签挑服务器
全部专题 →

服务器信息以作者仓库与官方文档为准;分类、传输方式与鉴权方式的中文说明为中文整理,最近核验 2026-05-17。