エージェントを強化
公開元自身のコマンドです。PlugMCP のエンドポイントではありません。
キーとログ付きでコピーをホスト。
npx -y [email protected]Claude Code にはローカルコマンド用のインストールリンクがありません。下のコマンドをコピーして、ターミナルで実行してください。
claude mcp add employee001 -- npx -y employee001@0.15.1ターミナルで実行します。
{ "mcpServers": { "employee001": { "command": "npx", "args": [ "-y", "[email protected]" ] } }}claude_desktop_config.json
リモート URL か、手元のマシンのコマンドか 2 つのインストール方法の違い
PlugMCP で実行
A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.
起動公開元より
Employee001 builds an AI twin of each employee from their own work: what they know, how they decide, and the tools they choose to connect. Ask the twin of the person who knows, ask the whole company for one cited answer, or convene the right twins in a Team Meeting. External actions wait for a person's approval. Open source, self-hosted.
Your company's organizational brain: agent twins of your real employees, self-hosted on your own machine. MIT. Telemetry is opt-in and off by default.
Consent first: no twin is trained, given a person's tools or allowed to act for them until that person has recorded their consent.
Every install is also your organization's own MCP server, so your team can ask the twins from the agents they already use: Claude Code, Codex, Cursor, Claude Desktop.
npx employee001 # first run: a short setup, then it opens http://localhost:3000
Signed in to Claude Code? Your Claude subscription runs the twins, no API key needed. Subscription mode covers one person.
Claude subscription covers one person. For a team, use an Anthropic API key, a cloud provider or a local model.
Get started · Why · How it works · Features · Install · MCP · Privacy & security · Compare · Roadmap
npx employee001. Setup finds your signed-in Claude Code and offers your Claude subscription, with no API key; it covers one person. Otherwise it asks for an Anthropic API key, your cloud provider (Bedrock, Vertex AI, Foundry) or a local model.No keys yet? npx employee001 demo opens an invented company with six twins and replays recorded answers. More on the demo.
In a growing company, the answers live in a few people's heads and in threads scattered across Slack, email, tickets and code. Everyone else waits for those people. Cross-team questions wait for a meeting. Context is at risk whenever someone is on leave, changes roles or leaves.
Search finds documents. A generic AI agent knows what a "Head of Sales" usually does. Neither knows what your colleague decided, why, and who else to pull in. Employee001 is built for that gap.
Twins are not fine-tuned models and they don't replace anyone. Each twin is its profile files plus its memory, run on Claude through the Claude Agent SDK. Employee001 also ships generic marketplace agents for roles you don't have a person for.
A twin can be built two ways. Both end in the same nine Markdown files: CONTEXT, EXPERTISE, PROJECTS, PEOPLE, DECISIONS, PREFERENCES, TONE, BOUNDARIES, EMPLOYMENT.
1. Build my twin (you). The person running Employee001 briefs their own twin at /build-my-twin. Write a few paragraphs about your work (what you own, how you decide, what people ask you, a few messages you wrote), or choose Interview me and answer about eight questions, one at a time. Add documents if you like (PDF, Word, Markdown, text, CSV, JSON). The builder writes the profile files from that material and says where it falls short instead of making things up. No invite, no Composio, no OAuth.
2. Invite a colleague (them). An admin creates a private invite link and picks a lookback window of 30 to 360 days. The colleague reviews a consent screen, answers a few role questions and connects their own tools (Slack, Gmail, Linear, GitHub, Calendar, Drive and more) through Composio. A builder agent reads that history through a read-only gate (120 turns, $5 budget) and writes their profile files. Invites need an API key or a cloud provider, plus a Composio key; Claude subscription mode is for one person.
Then, for every twin:
flowchart TB
Self["Build my twin<br/>write or be interviewed, + documents"]
Invite["Invite a colleague<br/>they consent and connect their tools"]
Builder["Twin builder (read-only)"]
Files[("9 profile files, plain Markdown")]
Ask["Ask a twin · Ask the company"]
Meeting["Team Meeting"]
Approve["Approval queue<br/>send, post, file"]
Audit[("Hash-chained audit log")]
Self --> Builder
Invite --> Builder
Builder --> Files
Files --> Ask
Files --> Meeting
Ask --> Approve
Meeting --> Approve
Approve --> Audit
Everything below is in the current npm release (0.10.0).
Build twins
/build-my-twin): brief your own twin by writing or by Interview me, with optional documents. No tools to connect./profile): a split-pane editor over the nine files, with version history..env files, keys, .git and node_modules are never read, and every add and remove is audited. The files it reads go to your model provider.Ask and collaborate
/ask): talk to one twin, see each source and tool it reaches for, or hand it a task. Your conversations are saved per person; each person sees only their own./brain): one question goes to the 1 to 3 twins whose area it is, the Org Brain and org-wide MCP systems, in parallel and read-only. One answer comes back with each claim cited, written through whichever model provider you use./council): selected or @mentioned twins take turns, see the whole thread and can tag a colleague in. Summarize & decide gives a decision, owned actions and open questions. Export as Markdown.Twins that keep context
OPENAI_API_KEY or EMBEDDINGS_BASE_URL./handover): when someone leaves, a short interview asks what their twin's profile is missing. With their consent the twin stays as a read-only "Former " that everyone can keep asking. Admins can download the handover notes.Work, with sign-off
/routines, /cockpit) with per-run archives and a daily per-twin budget. They run only when Autonomy is on (off by default).Platform
EMPLOYEE001_SANDBOX=1 wraps each twin process in Anthropic's Sandbox Runtime for an OS-level file and network boundary. It fails closed. Tested on macOS.export / import of the whole data/ folder, optionally encrypted.npx employee001 demo
Opens a made-up company, Lumen Labs, in a temporary folder: six invented twins, one of them a former employee. No API keys and no model calls. Everything is replayed from recordings, and nothing is written to your current directory.
Any other question gets a pointer to --live, which runs the demo on your own model provider. The microphone needs an ElevenLabs key.
npx employee001 # starts the server and opens http://localhost:3000
With no command it runs start, which runs setup first if there's no .env in the current folder. Setup asks only what a first answer needs and takes the next free port if 3000 is busy. Add Composio, ElevenLabs or OpenAI keys later from Settings. Press Ctrl+C to stop.
The folder you run setup in is your install: it holds .env and data/. Run start from the same folder, or set EMPLOYEE001_HOME=/path/to/install to use it from anywhere.
.env, sign in to Claude Code again, or run npx employee001 setup, then restart. On Bedrock, Vertex AI or Foundry, check that provider's credentials on this machine.cd back to it, or set EMPLOYEE001_HOME.npx employee001 doctor checks Node, .env, keys, Claude Code sign-in, port and the build. Tested on Node 22 and 24 LTS.[!NOTE] Requirements. Node.js 22+ and one model provider: your Claude subscription through a signed-in Claude Code on the same machine (no API key, for you alone), an Anthropic API key, Amazon Bedrock, Google Vertex AI, Azure AI Foundry, or a local Anthropic-compatible endpoint. Inviting colleagues also needs a Composio API key and a provider other than the Claude subscription. Building your own twin and the marketplace agents need neither.
Every install includes an MCP server. It belongs to your organization: it runs inside your installation, answers from your twins and your org brain, and is not published anywhere. Each person connects their own agent to it, like any MCP tool. So a developer in Claude Code checks what product decided before building, an engineer in Codex asks the sales twin what a customer requires, and a support lead in Claude Desktop asks who owns a migration.
Start Employee001 first (npx employee001 start), then connect a client. The Claude Code plugin is the easiest. It adds /employee001:ask-company, /employee001:ask-twin and /employee001:team-meeting:
/plugin install employee001 --marketplace dolevhayut/Employee001
Claude Code, MCP tools only:
claude mcp add --transport http employee001 http://127.0.0.1:3000/api/mcp
Codex:
codex mcp add employee001 --url http://127.0.0.1:3000/api/mcp
Cursor (.cursor/mcp.json):
{ "mcpServers": { "employee001": { "url": "http://127.0.0.1:3000/api/mcp" } } }
Claude Desktop and other clients that launch a local command use the stdio bridge:
{ "mcpServers": { "employee001": { "command": "npx", "args": ["-y", "employee001", "mcp"] } } }
Teammates on other machines. If Employee001 runs on a shared server, each person points their client at its URL and sends the instance's MCP token (EMPLOYEE001_MCP_TOKEN, kept in their own environment, never in a committed file):
claude mcp add --transport http employee001 https://e001.example.com/api/mcp --header "Authorization: Bearer $EMPLOYEE001_MCP_TOKEN"
codex mcp add employee001 --url https://e001.example.com/api/mcp --bearer-token-env-var EMPLOYEE001_MCP_TOKEN
What an agent can do. Read: list_twins, get_twin_profile, search_org_brain, list_pending_approvals, list_team_meetings, get_team_meeting. Ask: ask_twin, answer-only, rate- and cost-limited. Agents get member access: no settings, no approvals, no external actions. Full reference: docs/mcp.md. Plugin guide: plugins/claude-code.
Four different things, stated separately. The full list of subprocessors is in SECURITY.md; npx employee001 doctor --egress prints the hosts for your configuration.
WebFetch wait for an approver. Reads that leave the machine ask unless an admin allowlisted the source. Deletes and payments are always blocked. Unanswered approvals expire as a deny: 10 minutes in chat, 6 hours for background work.Bash, NotebookEdit, EnterWorktree and ExitWorktree are hard-disallowed, by the SDK's disallowedTools and a PreToolUse hook.org-brain/ and org-skills/. Other twins' memory and chats, invites, the audit log and settings are blocked. Write is limited to data/scratch/<employee-id>/.EMPLOYEE001_SANDBOX=1 adds a second, OS-level boundary around the model process. macOS needs ripgrep; Linux also needs bubblewrap and socat. See SECURITY.md.data/audit.jsonl with run id, employee id, tool, input, verdict and who approved. Rows are SHA-256 hash-chained; verify with npx employee001 audit verify and browse at /audit.¹ Limited to data/scratch/<employee-id>/; path traversal is rejected. Only the twin builder writes profile files.
² Spawns a restricted sub-agent: a web-researcher (WebSearch + WebFetch) or a brain-explorer (Read + Glob + Grep), with the same hard-disallow list.
³ External-effect tools. Every call that changes something hits the approval gate.
Models. claude-sonnet-5-5 by default, claude-sonnet-5 as fallback; switch to Opus per message in the chat UI. Source of truth: src/lib/sdk-defaults.ts.
./
├── .env # your keys (chmod 600; `secure` moves them to the OS keychain)
└── data/
├── employees/<id>/ # 9 Markdown profile files + knowledge/ per twin
├── org-brain/ # shared knowledge (Org Brain)
├── org/ # org settings: skills, Slack triggers, custom MCP servers
├── memory/<id>/ # per-twin memory: episodes and distilled facts
├── meetings/ # Team Meeting transcripts
├── audit.jsonl # every tool call, hash-chained
├── routines.json # scheduled work
└── shifts/<runId>/ # autonomous-shift archives
The server binds to 127.0.0.1 by default, so only this machine can reach it. To share it on a LAN (for example a Mac mini in the office):
EMPLOYEE001_BIND=0.0.0.0 npx employee001 start
Off loopback, the token setup generates (EMPLOYEE001_TOKEN) only opens the page that creates the first owner account. After that, everyone signs in with their own account. Invite links reach only /join, /onboarding and /api/invites/<token> without signing in.
[!WARNING] The app is not hardened for the public internet. Use a firewall or Tailscale. Don't put this on a port-forwarded box.
You host Employee001, so it runs inside your environment and your compliance scope. We make no SOC 2 or ISO claims. Report vulnerabilities privately, as described in SECURITY.md.
Last reviewed 2026-10-28, from each project's public pages. These are good open-source projects that overlap with Employee001; the difference is where an agent comes from, and whether the person it represents has consented. Where a cell says a page "doesn't describe" something, that is only what the page says, not a claim about the software.
Sources: paperclipai/paperclip, docs.paperclip.ing, Paperclip v2026.1001.0, cabinetai/cabinet, runcabinet.com, docs.runcabinet.com, titanwings/distilly, garrytan/gbrain, meetopenbot/openbot (several unrelated projects share the name openbot; this is the local-first agent harness). Spotted something out of date? Open an issue and we'll fix the table.
Employee001 Cloud is not available. A managed edition is planned as a separate paid service, not a feature gate on this code; join the waitlist at employee001.com. Everything in this repo stays MIT and self-hostable.
Your own cloud instance, today. Deploy a single-tenant instance to an always-on host you control: one org = one container + one persistent volume + its own secrets. The storage model is unchanged (data/ on a mounted volume), and migration uses the same export / import. Recommended: Fly.io. Not Vercel or other serverless hosts, whose filesystems are ephemeral. Runbook: docs/DEPLOY-CLOUD.md. Keep it private (Fly private networking or Tailscale).
Planned: SAML single sign-on, full UI translation, multi-tenant, and a desktop wrapper for non-technical admins.
data/; no databaseemployee001; releases carry build provenanceflowchart TB
People(["You and your team: browser, MCP clients"])
subgraph local ["Your machine or server"]
direction LR
Server["employee001<br/>Next.js + Agent SDK"]
Data[("data/<br/>profiles · memory · audit")]
Server <--> Data
end
subgraph models ["Model provider, pick one"]
direction LR
CC["Your Claude Code<br/>(subscription)"]
API["Anthropic<br/>API key"]
Cloud["Bedrock · Vertex<br/>· Foundry"]
Local["Local model"]
end
Tools["Composio + custom MCP servers"]
People --> Server
Server -.->|prompts + profile context| models
Server -->|tool calls, writes after approval| Tools
100% of the code here is MIT. Paid services, not paid features: professional onboarding (we install it, connect your tools, build the first twins with your team), SLA support and custom integrations. Open a discussion or email [email protected].
PRs welcome. See CONTRIBUTING.md for setup and the rules of the road, and SECURITY.md for reporting a vulnerability. Maintainers: RELEASING.md describes the tag-driven publish flow.
MIT © Dolev Hayut
If Employee001 is useful to you, star the repo. It helps others find it.
Your company's organizational brain: agent twins of your real employees, self-hosted on your own machine. MIT. Telemetry is opt-in and off by default.
{ "mcpServers": { "employee001": { "command": "npx", "args": [ "-y", "[email protected]" ] } }}ツール
ツール一覧はまだありません。リモートサーバーは、最初のヘルスチェック後に取得されます。
よくある質問
このページで Cursor、Claude Desktop、Claude Code、VS Code 用の stdio 設定をコピーできます。PlugMCP のプロキシではなく、公開元のインストール方法を使います。
このレコードに必須の環境変数は記載されていません。
Employee001 は stdio 経由でローカル実行されます。
最終チェックは 2026-11-15 です。 Your company's organizational brain: agent twins of your real employees, self-hosted on your own machine. MIT. Telemetry is opt-in and off by default.