Archie HQ
Autonomous Responsive and Collaborative Hyper Intelligent Employee
Potencia tu agente
Instalación del publicador. No es un endpoint de PlugMCP.
Host a copy with our keys and logs.
Aún no hay una versión empaquetada en el registro. Sigue el README del repositorio.
¿URL remota o un comando en tu máquina? En qué se diferencian las dos instalaciones
Ejecutar en PlugMCP
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.
AbrirDel publicador
Archie is an AI employee — you delegate real work to it in Slack, the same way you would to a colleague, and it gets the job done across any domain: engineering, marketing, analytics, ops, or anything you plug in.
Under the hood a task is one long-lived agent — the PM — that loads the skill for the domain, hands bulky or specialised work to workers it spawns, and reports back as a single voice. It's built on the Claude Agent SDK with a plugin architecture — add a new skill or department by dropping in a plugin directory, no core code changes.
Breaking change in 0.2.0. A task used to run several agent processes coordinating through message queues and a shared log; it now runs one. Plugin agent frontmatter, the per-agent MCP scoping and the sandbox network allowlist all changed, so the engine and your plugins repo must be upgraded together — and existing task metadata is rewritten, so back it up first. If you run Archie with your own plugins, read Migrating to the flat PM before you deploy.
Archie used to run several long-lived agent processes per task, coordinating through message queues, a shared log and owner handoffs. It now runs exactly one — the PM — and spawns workers through the SDK's own delegation tool. Three reasons:
Archie still reads as an employee, not a tool: it has a workplace (Slack), a memory that outlives a task, and a human approval gate before anything ships. You onboard new abilities as plugins, not forks.
Archie is a production system, not a demo — and it's deliberately cheap to operate:
Slack / CLI / GitHub → Task → PM agent ──┬─ skills (plugin:skill)
└─ workers (plugin:agent | general-purpose)
↕
MCP tools / Git clones / APIs
The task is sandboxed: filesystem access is restricted to the task's own folder and clones, network egress from Bash is an explicit allowlist, and code changes require human approval.
Archie ships with a small example plugin set — a summarize-or-draft skill the PM runs and a writing worker it can hand bulky material to — so a fresh clone does something useful immediately. This path needs only an Anthropic API key — no Slack app, no GitHub App, no SSH keys.
# 1. Clone and install
git clone https://github.com/sweatco/archie-hq.git && cd archie-hq
npm install
# 2. Configure — the only required value is your Anthropic API key
cp .env.example .env
# edit .env and set ANTHROPIC_API_KEY=sk-ant-...
# 3. Use the bundled example plugins (symlinks examples/plugins -> workdir/plugins)
npm run example:setup
# 4. Start the server in CLI-only mode (no Slack/GitHub required)
npm run dev # or: npm run docker:dev (runs inside the OS sandbox)
# 5. In a second terminal, chat with Archie via the interactive CLI
npm run cli
Ask it something like "summarize this: " — the PM loads the example skill and returns a structured summary, handing the work to the example worker when the material is bulky.
Going further:
ARCHIE_PLUGINS at a git URL, or replace workdir/plugins with your own checkout. Read the bundled writing-plugins skill at examples/plugins/.claude/skills/writing-plugins/SKILL.md and the Plugin System doc.SLACK_BOT_TOKEN + SLACK_SIGNING_SECRET for HTTP webhook mode, or SLACK_APP_TOKEN (xapp-...) for Socket Mode (no public URL needed). See Local Development.archie.json).Archie is configured entirely through plugins — directories that follow the Claude Code plugin structure with Archie-specific extensions. A plugin defines:
skills/) — domain workflows the PM loads on demand, namespaced as plugin:skillagents/*.md) — worker types the PM can spawn, addressable as plugin:agenthooks/) — Claude Code hooks for cost guards, validation, etc.Skills, agents and hooks are loaded natively by the Claude Agent SDK. Three files at the repo root stay engine-owned:
.mcp.json) — external tool integrations (Jira, Firebase, BigQuery, …), plus the per-tool approval tiers that decide which calls need a humanarchie.json) — the sandbox network allowlist, which repos to warm-clone at startup, and which may be merged without askingpm.md) — standing organisational context, tone and rules appended to the PM's system prompt, and its default model and effort, with no engine changes needed; see Plugin SystemExample plugin structure:
plugins/
├── engineering/
│ ├── .claude-plugin/plugin.json
│ ├── skills/
│ │ └── pr-workflow/SKILL.md # loadable as engineering:pr-workflow
│ ├── agents/
│ │ └── qa-reviewer.md # spawnable as engineering:qa-reviewer
│ └── hooks/hooks.json
├── marketing/
│ ├── .claude-plugin/plugin.json
│ ├── skills/
│ │ └── tone-analysis/SKILL.md
│ └── agents/
│ └── tov-reviewer.md
├── .mcp.json # MCP servers + per-tool approval tiers
└── archie.json # network allowlist, warm repos, auto-merge
An agent file needs a name, a description (that is what the PM reads when choosing a worker) and usually a model and effort. There is no repo binding: the GitHub App installation is the allowlist, and the PM mounts what a task needs.
Write an agent file only when a worker earns one — a fixed procedure and output envelope, a reviewer that must be blind to how the material was made, or a model/effort pairing the PM cannot express per spawn. Everything else is a skill the PM loads plus the general-purpose worker.
To add a new domain: create a plugin directory, drop in skills, and push. No core code changes and no redeploy — the next task picks it up.
A task is one agent and one SDK session:
Workers run inside the PM's own session and report back to it; only their final report reaches its context, and none of them can reach a user directly.
Key capabilities:
Agent tool, with background workers that wake the PM when they finishA task runs in a sandboxed environment with defense-in-depth. The task, not an agent, is the isolation boundary: workers run inside the PM's session and share its credentials, network allowlist and filesystem grants, so per-call approval tiers — not per-agent credential scoping — are what gate critical writes.
archie.json); web access only through the controlled research pipelineSee Security Architecture for the full threat model, enforcement layers, and deployment requirements.
Architecture:
web_research pipelineGuides:
writing-plugins skill under examples/plugins/.claude/skills/)Project:
ARCHIE_WORKDIRLicensed under the GNU Affero General Public License v3.0 or later (AGPL-3.0-or-later). See LICENSE.
In plain terms: you are free to use, study, modify, and self-host Archie, including inside your own organization, without restriction. The AGPL's network-copyleft condition means that if you run a modified version as a service made available to others (e.g. a hosted offering), you must make the corresponding source of your modifications available under the same license. This keeps improvements flowing back to the community.
This software is provided "as is", without warranty of any kind, express or implied. See the LICENSE for the full terms.
AI employee for your team — delegate real work in Slack and a team of Claude-powered agents gets it done. Built on the Claude Agent SDK with OS-level sandboxing, MCP tools, and a plugin architecture.
Resumen general
AI employee for your team — delegate real work in Slack and a team of Claude-powered agents gets it done. Built on the Claude Agent SDK with OS-level sandboxing, MCP tools, and a plugin architecture. TypeScript/Node.
Herramientas
Aún no hay lista de herramientas. Los servidores remotos la reciben tras la primera comprobación de estado.
Preguntas frecuentes
Copia en esta página la configuración stdio para Cursor, Claude Desktop, Claude Code o VS Code. Usa la instalación del publicador, no un proxy de PlugMCP.
Este registro no indica ninguna variable de entorno obligatoria.
Archie Hq se ejecuta localmente por stdio.
Última comprobación: 2026-10-28. AI employee for your team — delegate real work in Slack and a team of Claude-powered agents gets it done. Built on the Claude Agent SDK with OS-level sandboxing, MCP tools, and a plugin architecture. TypeScript/Node.