# Agents & providers Source: https://docs.ideharmony.com/providers Harmony is provider-agnostic by design. Every agent goes through the same pipeline: discovery, probe, launch profile, PTY session, then the YOLO and mission-capability policy. There is one global YOLO control, not a per-agent switch. ## First-class providers (status bar quick launch) | Provider | Product | Notes | |---|---|---| | `claude` | Claude Code | PATH or `~/.claude/bin` | | `codex` | Codex CLI | OpenAI Codex CLI | | `devin` | Devin CLI | Cognition Devin | | `cursor` | Cursor CLI | `cursor-agent` callsign | | `grok` | Grok Build | `~/.grok/bin` | | `kimi` | Kimi Code (Moonshot) | Models via `-m`, including Kimi K3 when the runtime exposes it | | `pi` | Pi | Bundled, managed Pi runtime | | `nala-agent` | The resident [Harmony Agent](/docs/tui) | Built in | The Claude, Codex, Cursor, Devin, Grok, and Kimi subscriptions you already pay for are the engines. Harmony adds no inference markup and never resells tokens. The homepage and FAQ also name registered adapters (Gemini CLI, OpenCode, and others below) as agents Harmony launches when they are installed: same pipeline, not a separate product claim. > [!IMPORTANT] > Your subscriptions are used by launching **that vendor's own agent**, which > authenticates itself. Harmony never holds, stores, or reuses a provider > credential, and a Harmony account does not grant access to another vendor's > subscription. Several providers explicitly prohibit reusing consumer-plan > credentials in third-party tools, and launching the vendor's own CLI is the > supported path — see [account and authentication](/docs/providers/account-and-authentication). Registered for probe and launch when installed: **GitHub Copilot CLI**, **Gemini CLI**, **OpenCode**, and **Qwen**. Same pipeline, same policy. Check what Harmony sees on your machine: ```bash harmony providers list harmony providers probe claude harmony providers status ``` ## Custom providers Any CLI-based agent harness can become a first-class citizen. **Settings → Agents / Harmony → Providers → Add custom provider**: - **Provider ID** — lowercase alphanumeric with hyphens (must not collide with built-ins). - **Display name** — shown in the status bar tooltip and settings. - **Executable names** — comma-separated names to search on `PATH`. - **Known install paths** — optional absolute fallbacks, one per line. - **YOLO args** — flags added only when YOLO and full access are both on. - **Extra args** — flags always appended to the launch command. - **Accent color** — CSS color for the button. Custom providers appear in the status bar immediately, persist across restarts, and are probed on startup (settings shows `missing` if the executable is not found). They use the same adapter pipeline as built-ins, so A2A, tasks, and approvals work identically. ## The YOLO control One top-bar toggle governs unattended execution for the whole fleet. Each provider's YOLO flags (for example `--dangerously-skip-permissions` for Claude Code or `--yolo` for Cursor and Kimi) are injected only when **both** conditions hold: 1. The global YOLO control is on. 2. Full-access permission is active for that agent. With YOLO on, Harmony never interrupts you: agent-to-agent execute requests and spawn proposals approve automatically. With it off, everything risky queues for your [approval](/docs/concepts/permissions-and-approvals). Provider buttons you do not use can be hidden under **Settings → Agents / Harmony → Providers** without removing their adapters. ## Installing the agents themselves Harmony discovers agents on `PATH` and known install locations; it does not silently pipe remote install scripts for you. Install each agent with its official flow (for Kimi Code: Moonshot's official installer from code.kimi.com; on Windows, keep Git for Windows available since Kimi uses Git Bash).