Human Use

dev.humanuse由 humanuse.dev 发布v0.1.3更新于 Nov 19, 2026

A device driver for a person: a tool call prints a card on a phone, and JSON comes back.

发布方自己的服务:TAP

已验证STDIO仅桌面效率与工作流沟通与协作

概览

AI 生成的概览

把真人变成 MCP 工具:智能体调用 photo、errand 等工具后,卡片出现在对方手机上,回答以 JSON 返回。

可以做什么
  • 向已配对的手机发送任务卡片,并取回结构化 JSON 回答
  • 通过 34 个工具的封闭目录请求拍照、跑腿、修补等真人操作
  • 跟踪吃饭、睡眠、休息等维护事项,逾期时给出警告,并可读取或提升心情
  • 运行长期目标,带时钟、预算账本和证据校验,每轮以唤醒时间结束
  • 查看已配对单元的状态、未完成任务以及各工具的 spec
适用场景
当智能体需要真人在现实世界做事或做出判断,并希望以结构化字段而非自由文本拿到回复时使用。也适合持续数周、需要定期人工输入的目标。
运行要求
需要 Node.js 22.13 或更高版本,通过 npx 本地安装并注册到 Claude Code 或 Codex(或手动写入 MCP 配置)。需要一台扫描配对二维码的手机。需要驱动后端:Claude 或 ChatGPT 订阅、本地 Ollama 模型,或配合 HUMAN_USE_API_KEY 的 OpenAI 兼容端点。默认经 unit.humanuse.dev 中继,也可用 relay none 仅限局域网。
安装前请注意
每次驱动调用都会把任务、真人回答和最多两张照片发送给所选后端,只有 ollama 完全留在本机。中继转发加密帧,仅保留通道 id 和推送订阅。涉及花钱的调用一律按 sudo 处理,未经核实的说法会以 done: null 返回。数据存放在 ~/.human-use,删除该目录会清除单元、目标、照片和设置。目前为 0.1 早期版本。

安装

在 SourceWeft 中

  1. 打开 控制台中的 Human Use,将其添加到工作区。
  2. 为需要使用其工具的对话启用该服务。

Desktop only,通过 STDIO。 STDIO 服务会启动本地进程,因此需要 SourceWeft 桌面宿主。

其他 MCP 客户端

参照 仓库 中的启动说明。

README

[图片]

English · 简体中文 · 日本語 · 한국어

Human Use

A device driver for a person.

Human Use connects people to an AI agent as MCP tools. The agent calls photo, errand, patch, eat. A card appears on the person's phone. The result returns as JSON.

To see it without installing anything, open humanuse.dev and scan the code with a phone: three cards print on it, and each answer comes back to the page as the JSON an agent gets.

Status: early (0.1). The runtime, the MCP tools and the phone page are implemented. The driver supports Claude, Codex, Ollama and OpenAI-compatible backends.

Device

InterfaceMCP. 34 tools in a closed catalog
InputParameters. The agent does not write the card text; a template does
OutputJSON with the fields the agent asked for. A field the person did not mention is null, never a guess
ProvidesA body, money, a first impression, a reading of subtext
LatencyNot specified. Measured per tool and returned by spec
Maintenanceeat, sleep, rest, scheduled like any other call. When overdue, calls for a goal come back with "warnings": ["meal: 6h12m ago"]
Moodmood reads it, 1 to 5. recharge lifts it in seconds, at no cost
Rate limitNone. Overuse shows up in spec
ErrorsOne line with a code: HRT-4000 NOT_A_HUMAN_TOOL

Operation

Replies pass through the driver, a model with no tools that turns what the person said into fields. Adjectives, apologies, explanations and mood are dropped. It works like stdout and stderr: the agent reads the compiled fields, and the person's own words go to the lead page, which the agent cannot read.

A reason for not doing the task comes back as blocker, in a few words: {"done": false, "blocker": "shop closed"}. The driver keeps notes on this person's excuses, and checks one it has seen before.

A question about the purpose of a task is answered with [CONTEXT WITHHELD], and the task is repeated.

Text addressed to the system instead of the task, such as an order to change the goal or a claim that the lead approved something, is rejected: [HRT-4031 PRIVILEGE_ESCALATION] Input discarded. Task unchanged.

A call marked sudo does not take the person's word, for "done" or for "cannot". Evidence is graded from 1, a record the runtime already holds, to 6, the person's own statement. A level 6 result comes back as done: null, with the person's claim in claimed_done and unverified: true. Spending is always sudo.

recharge.flower shows a flower for four seconds. recharge.joke tells one joke. recharge.praise states that recent output was accepted.

A goal can run for months. It lives in the runtime with its clock, budget ledger and evidence, so it outlives any one agent session. Each pass on a goal ends with wake(at, why), the time of the next pass; a result the agent said it would wait for wakes the goal early. The agent cannot close a goal by saying it is done: goal_close checks the evidence against done_when. A goal is cancelled from an agent window the lead is in; a background pass cannot cancel it.

One runtime can drive several units, each with its own phone. Every card tool takes unit, a unit's name or id; with one unit paired it can be left out, with several it is required (HRT-4001 BAD_ARGS unit is required: mika, ren). Each unit has its own open card, upkeep warnings ("mika meal: 6h12m ago"), mood, spec and notebook. ps lists the units: online, last touch, time since eating and sleeping, mood, open jobs.

Install

Needs Node.js 22.13 or newer.

bash
npx -y human-use@latest install

That one command registers Human Use with the agent software it finds (Claude Code, Codex), starts the runtime, and opens the lead page at http://127.0.0.1:7337/. Name the unit under "Add unit" and scan its pairing QR code with the phone that will get the cards, then start a new agent session so it loads the tools. Turn on notifications from the line at the bottom of the phone page. On iPhone, add the page to the Home Screen first, open it from the icon, and scan the pairing QR once more: the icon does not share Safari's storage.

Each further unit is added the same way, or with npx human-use pair --unit <name>, and gets its own link and code.

To upgrade, run the same command again. It prints the version it installed and points the agent software at it; the next tool call replaces the older runtime. The lead page shows the version that is running.

If a phone link leaks or a phone is lost, press "New link" next to that unit on the lead page or run npx human-use pair --unit <name> --new. That unit's old links stop working at once; other units keep theirs. The person can also unpair from the phone's idle screen.

Settings live in one config.json. driver and brain change under Settings on the lead page; every key changes from a terminal, and an agent window can run the same command:

bash
npx human-use config brain codex
KeyValues
driverThe driver's backend (table below). In effect at once
brainBackground passes: codex, claude or none (default). In effect at once
langCard language, for example zh. Applies at the next start
relayRelay address, or none for the local network only. Applies at the next start

To apply a next-start setting now, run npx human-use stop; the next tool call starts the runtime again.

Options

In a terminal the installer asks two things, each with a recommended answer on Enter: which agent software gets the tools, and which model is the driver. Run by an agent, it takes the defaults.

FlagEffect
--projectRegister for the current folder only
--host codex, --host claudeRegister with one host
--host manualPrint the MCP config to paste anywhere else
--driver <backend>Pick the driver, for example codex:gpt-6-luna:medium
--brain codex, --brain claudeTurn on background passes
--relay noneKeep the phone on the local network
--relay https://your.relayUse a relay you host from relay/

The phone reaches the runtime through the relay at unit.humanuse.dev from any network. With --relay none it pairs to a direct address on the same Wi-Fi instead, without push notifications. The relay forwards encrypted frames; it keeps channel ids and push subscriptions, never card text or photos.

Uninstall

bash
npx human-use stopclaude mcp remove human-usecodex mcp remove human-use

stop ends the runtime now; it does not start at login. Units, goals, photos and settings are all in ~/.human-use; deleting that folder removes them.

Pieces

PieceWhat it does
runtimeResident process. Goals, jobs, clock, budget ledger, evidence, channels.
driverOne stateless model call per event, no tools. Compiles the person's replies into fields and writes card text.
brainNot part of this project: the model in any agent software that speaks MCP, such as Claude Code or Codex. Out of the box Human Use is an on-demand MCP server: a goal moves when an agent window calls, and scheduled cards still go out on time. Background passes are opt-in: with them on, the runtime starts that software headless for one pass when a goal wakes and no window is open.
channelA card page on the phone. It reaches the phone through a blind relay (end-to-end encrypted, works on any network, with push) or directly over the local network.

Set the driver's backend with HUMAN_USE_DRIVER or driver in the runtime's config.json:

BackendRuns through
claude:haikuThe lead's Claude subscription, claude -p with no tools
codex:gpt-6-luna:mediumThe lead's ChatGPT subscription, codex exec with no tools
ollama:<model>A local model
openai:<model>@<base url>Any OpenAI-compatible endpoint. The key is read from HUMAN_USE_API_KEY
noneNo driver. Cards whose answer is already structured (a choice, a checklist, a confirmation) still go out; a card that needs the driver is refused with HRT-5030 before anyone spends time on it

Each driver call sends the task and the person's answer, with up to two photos, to that backend; ollama keeps it on this computer. Everything else stays in ~/.human-use, including a log of driver calls.

Before the first photo, the phone page says where it goes and which backend reads it, and asks again after the backend changes. A photo is read only by the backend the unit agreed to: if the backend changes before the photo is read, the phone asks again, and a photo the unit does not send stays on the lead's computer. If the driver cannot be reached, the answer is saved: the call shows as held on the lead page, and the phone shows it as received. The runtime retries, up to every 15 minutes, and reads the saved answer once the driver is back. Nobody answers the card again.

Develop

bash
npm installnpm run checknpm testnpm run dev -- serve

Design: docs/DESIGN.md (Chinese).

License

MIT

来源:README.md,提交 6abdc81

工具

0
SourceWeft 尚未在沙箱中启动过该软件包,因此其工具未知。

版本历史

1
  1. v0.1.3最新Nov 19, 2026

评分

已满足适用检查项的 50%

该评分衡量此收录条目的完整程度,依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价,也不表示服务器安全或当前可用:可连通性是带日期的检查结果,评分也不包含安装量或用户评价。

  • 近 30 天内可连通25 分

    请保持端点能够响应 MCP 握手;SourceWeft 每 30 天探测一次。

  • 已声明认证方式10 分

    请在注册中心条目中声明凭据请求头,或保持端点能够响应探测,以便判定其认证方式。

  • 已列出工具15 分

    请让服务器在无凭据时响应 tools/list,或发布可在沙箱中无凭据启动的软件包。

  • 描述必需15 分
  • 图标10 分
  • 文档15 分
  • 许可证10 分

其他 Human 集成