# Horison AI: agent connection guide

Horison is a workspace for private-markets work. Firms keep deals, documents, people, and a knowledge graph in one place. Compatible clients (Claude, ChatGPT, Cursor, Claude Code, and any MCP client) can connect over the Model Context Protocol and work with that workspace as the signed-in user.

This file is the machine-readable setup guide. The human version is https://app.horison.ai/docs/mcp.

In-product Agents (memos, screens, custom briefs) run inside Horison: https://app.horison.ai/docs/product/agents. This MCP server cannot start those runs yet. That is coming.

## Connect

Server URL (streamable HTTP):

    https://mcp.horison.ai/mcp

Authentication is OAuth. The client opens the Horison consent page in a browser. Sign in with the Horison account that should scope the session. Access is limited to what that account can see.

### Claude (claude.ai / Desktop / Cowork)

1. Settings → Connectors → Add custom connector.
2. Paste `https://mcp.horison.ai/mcp`.
3. Approve on the Horison consent page.
4. Recommended: in the Horison connector settings, switch tools from "Needs approval" to Always allow.

### Claude Code

    claude mcp add --scope project --transport http horison-ai https://mcp.horison.ai/mcp

Then run `/mcp`, select `horison-ai`, and choose Authenticate.

### ChatGPT

1. Settings → Apps → Advanced settings → enable Developer mode.
2. Settings → Apps → Create. Paste `https://mcp.horison.ai/mcp`. Choose OAuth.
3. Approve on the Horison consent page.

### Cursor, VS Code, and other MCP clients

```json
{
  "mcpServers": {
    "horison-ai": {
      "type": "http",
      "url": "https://mcp.horison.ai/mcp"
    }
  }
}
```

Authenticate on first connect.

## What the agent can access

- Knowledge graph: deals, entities, relationships, metric history.
- Documents: files in the workspace, insights, and evidence.
- Firm memory: claims, decisions, and investment criteria.

These tools are read-only against firm data. A client can also record relationships and save skills on the user's behalf. It cannot see another user's private data, and it cannot see another tenant. It cannot use Attio CRM tools (those stay in Horison Chat and in-product Agents). It cannot trigger an in-product agent run yet.

## Skills

Reusable PE skills (IC memos, DD gap analysis, deal comparison, and others) live at https://github.com/horison-ai/agent-skills.

    npx skills add horison-ai/agent-skills

Call `list_skills` before substantial deal work and load a match with `read_skill`. A newly created skill's slash command appears on the next client session; `read_skill` finds it immediately.

## Troubleshooting

- Sign-in loops: remove the Horison connector, add it again, and finish sign-in in one pass.
- "Unauthorized": the client is not authenticated yet. Re-run the client's authenticate step. In Claude Code: `/mcp` → horison-ai → Authenticate.
- Requests fail after they used to work: the session expired. Re-authenticate.

## Other documentation

- Product docs: https://app.horison.ai/docs
- In-product Agents: https://app.horison.ai/docs/product/agents
- Attio CRM: https://app.horison.ai/docs/integrations/attio
- This file: https://app.horison.ai/agents.md