agent4.io — tenant MCP server

The tenant control plane for agent4.io, exposed over MCP. Point a coding agent at it and it can build and operate real customer-facing agents end to end — create knowledge bases, upload documents, write skills, configure an agent, publish a share link, and read usage — without anyone opening the dashboard.

Listed in the official MCP registry as io.agent4/agent4-tenant.

This repository is the server's public description: the manifest, the tool list, and how to connect. The server itself is hosted — there is nothing to install or run, and no source here.

server.json is a mirror of the manifest published to the MCP registry. It is generated from the platform repo — don't edit it here; the change would be silently overwritten on the next sync.

---

Connect

Endpoint: https://api.agent4.io/v1/mcp (streamable HTTP) Auth: X-API-Key header, using your tenant API key — agent4.io dashboard → Tenant & API, or the console → Settings → Security. It is shown once at creation.

{
  "mcpServers": {
    "agent4": {
      "type": "http",
      "url": "https://api.agent4.io/v1/mcp",
      "headers": { "X-API-Key": "tk_live_..." }
    }
  }
}

The key scopes every call to your own tenant. There is no cross-tenant access.

Don't have a tenant yet? Sign up free — the free tier includes 5M tokens/month and 5 end users, with unlimited agents.

---

What it is for

agent4.io is a multi-tenant RAG agent platform for consulting-led businesses — law firms, immigration and study-abroad agencies, lenders, insurers, tutoring companies, hospitality groups, B2B software teams. You load your own material and the platform runs agents that answer prospects on your website, in Telegram, and on WhatsApp, 24/7, in seven languages.

Two decisions shape the whole product, and they show up in this API:

• Grounded by default. Agents answer from the tenant's own documents. Web search is not wired in. When the answer is not in the material, the agent says so rather than guessing.
• Explicit refusal boundaries. Every agent template carries a hand-written boundary — never quote a rate, never predict a visa outcome, never assess a child. The failure mode that hurts these businesses is a confident wrong answer, not a missing one.

---

Tools

Agents create_agent · update_agent · get_agent · list_agents · diagnose_prompt · improve_prompt

Knowledge bases create_knowledge_base · update_knowledge_base · get_knowledge_base · list_knowledge_bases · add_knowledge_file · add_knowledge_text · search_knowledge_base

Skills create_skill · update_skill · get_skill · list_skills

Deployment create_share · list_shares — website widget, Telegram, WhatsApp

End users list_end_users · get_end_user · list_user_sessions

Storylines (stateful, graph-orchestrated agents) create_storyline · update_storyline · get_storyline · list_storylines · validate_storyline · publish_storyline · unpublish_storyline · export_storyline · import_storyline · clone_storyline

Page context (what the agent knows about the page it is embedded on) upsert_page_context · resolve_page_context · list_page_contexts · page_context_stats

Ops & docs tenant_info · usage_stats · list_tools · list_mcp_servers · search_agent4_docs

search_agent4_docs is worth knowing about: it searches the product documentation, the integration cookbook and the generated REST reference, so an agent can look up how to do something before doing it.

---

The other ways in

This MCP server is one of three entry points, and they are not interchangeable — pick by what you are building.

	What it is	Use it when
This MCP server	~40 tools on a remote endpoint	You want an agent to operate the platform — create agents, load knowledge, publish share links
agent4-io skill on ClawHub	The how-to: recipes, conventions, worked examples, and an /agent4-io command	You want your coding agent to know how to use the tools well, not just that they exist
REST API — agent4.io/api	The full HTTP surface, generated from the live OpenAPI spec	You are writing an application rather than driving an agent

Install the skill and the MCP server together: the skill assumes this endpoint is connected, and tells the agent which recipe to follow for a given job.

Machine-readable references: /api.json · /api.md · /llms.txt

The skill's recipes are authored as the Cookbook on the website, which is their single source — the ClawHub package is generated from it.

---

Links

Website · Docs · API reference · Pricing · MCP registry entry

Operated by SUCHKA sp. z o.o., Warsaw, Poland — contact@agent4.io