Every coding agent. One chat. Every tool.
Connect Claude Code, Codex, or Cursor. Harness equips them with MCP servers and skills — and lets you swap their entire toolset mid-conversation, in one click.
Works with
Use the agent you already pay for.
No lock-in, no markup. Connect with your own credentials — a login or an API key — and the agent runs in an isolated cloud sandbox, billed to your account. Switch agents whenever you like.
Claude Code
Anthropic's coding agent
Codex CLI
OpenAI's coding agent
Cursor
Cursor's CLI agent
OpenCode
Open-source agent
Gemini CLI
Google's coding agent
Built on the Agent Client Protocol
Harness speaks ACP, so it works with any agent on the ACP registry — not just the ones above. New compliant agents drop in as they ship.
Switch contexts for code review without losing your place.
A harness bundles a set of MCP servers and skills. Swap from one to another mid-conversation and your agent's entire toolset changes in a click — the chat keeps going, context intact.
- One agent, many tool configurations
- Credentials brokered server-side — never in the sandbox
- No restart, no re-explaining — the transcript carries over
Tools available right now
The control plane for your agents. Not another model wrapper.
Bring your own credentials
Connect Claude Code, Codex, or Cursor with your own login or API key. Each runs in an isolated cloud sandbox — Harness never sees your model bill.
Rapid MCP context switching
Swap your agent's entire toolset mid-conversation. Go from a research stack to a GitHub-ops stack in one click — the context carries over, the tools change instantly.
One chat for all of them
The same polished UI for every agent: streaming reasoning, terminal output, file diffs, and inline approvals. No editor, no per-agent setup.
A catalog of MCP servers
GitHub, Notion, Linear, DeepWiki, Context7, and any custom URL. OAuth handled and refreshed — credentials brokered server-side, never in the sandbox.
Skills from skills.sh
Bundle battle-tested playbooks — code review, debugging, web search — into a harness. Your agent imports your team's conventions on connect.
Sandboxed & approvals-first
Agents execute in isolated Daytona sandboxes with a real terminal and git. Sensitive commands surface an inline approval card before they run.
From your agent to your tools in three steps. No SDK, no editor.
Connect an agent
Bring Claude Code, Codex, or Cursor — or start with Harness's built-in models, no setup. Credentials are encrypted and write-only.
Equip a harness
Bundle the MCP servers and skills your agent needs. OAuth into GitHub, Notion, or Linear in one popup. Save it as a reusable harness.
Chat & switch
Send a message and watch tool calls, diffs, and terminal output stream in. Swap harnesses any time — the conversation keeps going.
Connect your first agent.
Bring Claude Code, Codex, or Cursor — or start with Harness's built-in models. Equip it with tools in minutes.
