Orca

Orca (source) is a coding-agent orchestrator: Claude Code, Codex, Cursor CLI, OpenCode, Qwen Code and dozens more, each in its own git worktree, in parallel. A coordinator agent splits the task between workers, reviews diffs and sends comments back to the agent; there are remote runtimes (SSH, cloud, VM) and a mobile app.

Orca itself is neither a provider nor a model: it launches your CLI agents, and each carries its own authorization. Point those agents at Mixen and every worker runs from a single rouble balance — a Claude/ChatGPT subscription does not scale to ten parallel workers, a Mixen key does.

Installation

Ready builds — on orca.dev: .dmg for macOS, .exe for Windows, an AppImage for Linux. Via package managers:

brew install stablyai/orca/orca   # macOS (Homebrew)
# AUR: stably-orca-bin            # Arch Linux

Orca is open source (MIT), github.com/stablyai/orca; it updates often, with daily releases.

Setup

Two paths, both work:

1. Configure the agent itself — Orca picks up its config as is (Claude Code, for example, reads ~/.claude with no extra setup):

• Claude Code: ANTHROPIC_BASE_URL=https://api.mixen.ai plus an mxn-… key in ~/.claude/settings.json;
• Codex CLI: OPENAI_BASE_URL=https://api.mixen.ai/v1;
• OpenCode, Qwen Code — by their own guides.

2. Per-agent environment in Orca: Settings → Agents → Per-agent launch defaults — every agent has editable Command, Arguments and environment. Put the variables there (ANTHROPIC_BASE_URL, ANTHROPIC_AUTH_TOKEN or OPENAI_BASE_URL) and every launch of that agent gets them, including the coordinator’s workers.

After that pick the agent in the dropdown of any worktree terminal — Orca launches it with the worktree as the working directory.

The economics of parallel workers

• Every worker is a full agent session with its own context; run as many as the balance allows.
• Prompt caching works per session: a repeating prefix (system prompt, project rules) is billed at ~10% of the input price from the second turn on.
• For bulk routine subtasks pick a cheap model — the coordinator does not have to be pricier than the workers.

Recommended models

Role	Model	Input / output per 1M
Workers: routine, classification, small edits	GLM 5.3 Flash (z-ai/glm-5.3-flash)	0.50 / 20.60 ₽
Workers: edits in familiar code	GLM 5.3 (z-ai/glm-5.3)	74.80 / 235.20 ₽
Coordinator, review, hard tasks	Claude Sonnet 5.5 (anthropic/claude-sonnet-5.5)	213.80 / 1069.20 ₽

Full tables with cache columns — in the catalog.

Troubleshooting

• The agent in Orca hits its own account, not Mixen — the env did not arrive: check Per-agent launch defaults (path 2) or the agent’s own config (path 1); restart the agent session after editing.
• Diff review is slow — the reviewer inherits the agent’s model; move review to a session with a cheap model.
• Rate limits — one per key: 120 RPM shared by all workers; for bulk runs spread workers across several keys.

The rest of the integrations — in the overview.