MCP
Use MCP to let AI assistants read and act on work context through tools.
Endpoint and authentication
Use the hosted MCP endpoint when a client asks for the server URL:
https://mcp.onehorizon.ai/mcpAuthentication uses OAuth 2.0. Public clients use PKCE directly, so most MCP clients do not need a client ID or client secret. The assistant acts with the workspace access of the person who connects it.
Use MCP when an assistant should answer questions, prepare updates, create work, or update work from inside the assistant. Use the REST API or JavaScript SDK when you are building a backend service or CI job.
What assistants can do
MCP tools can read work recaps, team recaps, planned work, completed work, blockers, initiatives, bugs, task details, documents, teams, workspaces, identity, and taxonomy.
They can also create or update initiatives, bugs, feature requests, and todos when the authenticated user has access.
MCP prompts can generate work summaries, standup prep, handoff notes, initiative summaries, and bug triage prep.
Common requests include:
- "Show my planned work for this week."
- "Give me a team recap for last week."
- "What's blocking the frontend team?"
- "Capture this as completed work."
- "Prepare bug triage notes for the open bugs."
- "Write handoff notes for the tasks I finished today."
For execution work, MCP is usually the context layer. Use spec-driven development to keep the source work clear, then use AI Handoff, Use in Terminal, or Local Workers when the assistant needs to move from context into implementation.
Client setup
| Client | Setup notes |
|---|---|
| ChatGPT | Requires ChatGPT Plus or higher with Developer Mode enabled. Add a custom app in ChatGPT Apps with the One Horizon MCP endpoint and OAuth authentication. Developer Mode currently works in regular chats, not Project chats or Deep Research. |
| Claude Code | Add the server with claude mcp add --transport http onehorizon https://mcp.onehorizon.ai/mcp, or import it from Claude Desktop with claude mcp add-from-claude-desktop --scope user. Use /mcp inside Claude Code to inspect connected servers. |
| Claude Desktop | Add a remote HTTP MCP server named onehorizon, authenticate through the OAuth prompt, then start a new conversation before using tools. |
| Cursor, Windsurf, JetBrains IDEs, and Codex | Add the MCP server from the tool's MCP settings and authenticate through the browser. If the tool supports project-level MCP config, keep workspace-specific setup in the project config. |
| n8n and automation clients | Use MCP when the automation wants assistant-style tools. Use Webhooks when the automation should react to workspace events. |
If a client supports both a user-level and project-level MCP configuration, choose user-level for personal work context and project-level when the connection should travel with a repository.
Access rules
MCP runs with the authenticated user's workspace access. Team IDs supplied to create or update tools are validated against teams the user can access.
The assistant cannot see workspaces, teams, documents, or tasks the signed-in user cannot access. It also cannot bypass workspace permissions through a tool call.
Troubleshooting
If tools do not appear, start a fresh assistant chat after connecting the server. Some MCP clients cache available tools per conversation.
If OAuth does not complete, confirm the client is using the hosted endpoint exactly and that browser popups are not blocked. Public MCP clients should not require a manual client secret.
If a write fails, check the user has access to the target workspace and team. For permission lookup material, see Permissions.
If a client reports intermittent tool or HTTP errors, retry from a new chat with fewer active MCP servers. Large enabled tool sets can slow or confuse some clients.