Claude Code
Claude Code works with One Horizon in two ways. As a local agent, it runs queued work from your machine. Through the MCP connector, you can ask it about your work while you code. For Claude on desktop or the web, see Claude.
Choose the Claude Code path
| Path | Best for | What happens |
|---|---|---|
| Local Claude Code agent | Queued implementation, planning, review, and workflow steps. | One Horizon starts the claude CLI on your machine, sends the work item as the brief, and records the run. |
| MCP connector | Asking Claude Code about workspace work from an interactive terminal chat. | Claude Code calls the hosted One Horizon MCP server with your authorized user access. |
| Open in Claude Code | Resuming a stopped local agent run. | Completed or failed runs can reopen the saved Claude Code session from the run menu. |
MCP setup does not satisfy local agent prerequisites. The local agent needs the claude binary on PATH and an authenticated Claude Code CLI session.
What a local Claude Code agent can do
A local agent claims sessions from Send to, direct assignment, or workflow steps. It runs with a Builder, Planner, or Reviewer preset and the Read-only, Workspace access, or Full access level from its configuration. It prepares the task workspace, follows the workflow file for setup, checks, commits, and handoff, and reports progress, terminal activity, and the final outcome to the run. When a saved provider session ID exists, a completed or failed session can be reopened.
One Horizon does not collect Claude credentials. The Claude Code CLI on your machine owns authentication.
Set up a Claude Code local agent
Install and sign in to Claude Code CLI
Confirm the
claudecommand is available in the same environment that starts Desktop agents or runsone agent:claude --versionclaude /login
Claude Desktop is not the automation harness. The local agent invokes the CLI directly.
Install Desktop or sign in to the CLI
Use One Horizon Desktop for guided setup, or use the One Horizon CLI from the repository.
Create the Claude Code agent
In Desktop, open Workflows, go to the agents page, select Add agent, and choose Claude Code. In the terminal, run:
one agent createChoose the task mode, agent access, project folder, code isolation, workflow file, and PR or MR behavior.
Verify the harness
In Desktop, finish setup and confirm the agent shows online. In the terminal, run:
one agent doctorFix any missing
claudebinary, CLI authentication, workspace, repository, or workflow file issue before starting pickup.Run assigned work
Open an initiative, bug, or todo with a usable brief and choose Send to, or target the agent from a workflow step. For terminal pickup, run:
one agent watch
Local Agents covers task modes, access levels, project folders, worktrees, pull requests, and restart behavior. Monitoring covers run states, recovery actions, and paused sessions.
Local agents are experimental. Keep tasks small enough to review, and check code, tests, and pull requests before merging.
Prepare the work item
Before launch, add the outcome, important files or links, constraints, and acceptance notes to the initiative, bug, or todo. For planned implementation work, follow spec-driven development: specific enough for Claude Code to act, short enough for a person to review.
Connect Claude Code through MCP
The MCP connector lets Claude Code chat read and update your work while you code. It uses the hosted One Horizon MCP server over HTTP with OAuth.
Add the MCP server
Import an existing Claude Desktop MCP configuration:
claude mcp add-from-claude-desktop --scope userCopied!Or add One Horizon directly:
claude mcp add --transport http onehorizon https://mcp.onehorizon.ai/mcpCopied!Authorize the connection
Run
/mcpin Claude Code, open the browser authorization flow, sign in to One Horizon, and return to Claude Code./mcpCopied!Use Clear authentication in the
/mcpmenu when you need to revoke and reconnect.Share project configuration when needed
Commit
.mcp.jsonat the repository root when the project should carry the One Horizon MCP entry for the team..mcp.json { "mcpServers": { "onehorizon": { "type": "http", "url": "https://mcp.onehorizon.ai/mcp" } }}
Copied!The hosted MCP server uses OAuth, so this file does not need environment variables.
Ask Claude Code through MCP
"show my assigned issues and todos for today"
Claude Code can pull assigned work from One Horizon, including work connected from Linear or Jira.
"use my assigned work details for the gateway API"
Claude Code can bring the work item into context before proposing changes.
"capture this as completed work"
Claude Code can summarize the result and add a completed task.
Load workspace capabilities in sessions you start yourself
One Horizon Desktop can connect Claude Code so sessions you start yourself load the skills, MCP servers, and plugins your workspace assigns to that repository. Open Settings → Integrations → Coding tools, then select Connect on Claude Code. Without Desktop, run one capabilities connect claude. Sessions One Horizon starts for you load capabilities without this step.
Claude Code shows Not qualified yet until One Horizon has tested the installed version, and Connect stays unavailable until then. Agent capabilities covers repository assignments and each status.
Security and permissions
The MCP connector acts on your behalf. Data access follows your roles, workspace, team membership, and granted scopes.
Local agent runs use the agent access level you chose during setup. The selected access level is the boundary for commands, files, and provider tools.