Claude Code's multi-agent orchestration—Dynamic Workflows, Agent Teams, and Subagents—has always been Anthropic's domain. That changes today with cc-fleet, an open-source project from ethanhq that lets any OpenAI- or Anthropic-compatible API join your Claude Code session as a first-class worker. DeepSeek, GLM, Kimi, Qwen, Groq, Together, Fireworks, local vLLM deployments—even your existing Codex subscription—can now run inside Claude Code's orchestration layer without requiring a separate Claude subscription of their own.

How It Works

The project installs via a single shell command and registers itself as a Claude Code plugin. Once in place, cc-fleet intercepts the orchestration layer and routes agent() calls to whatever provider you've configured instead of defaulting to Anthropic's models. Your main session stays completely untouched—its OAuth credentials never leave your environment while third-party workers run on their own API keys. The project takes key protection seriously: tokens are fetched at request time only, never written to environment variables, command-line arguments, or shell history, and stored at 0600 permissions on disk.

Three Delegation Lanes

cc-fleet exposes three distinct ways to distribute work across providers. Dynamic Workflows use a JavaScript API identical to Claude Code's native Workflow tool—parallel() fans out tasks, pipeline() chains phases, and each agent() node accepts a provider option so different vendors run in the same orchestration graph. Budget caps (in USD or tokens) can gate spend mid-run, and finished leaves hit cache on replay. Agent Teams spawn long-lived Claude processes in tmux panes, letting teammates collaborate across turns with native TeamCreate and SendMessage primitives—mix providers freely within one team, then hand follow-ups between them. This lane requires tmux (no Windows support yet) but supports park/restore so you can hide a pane without losing context. Subagents are the lightest option: synchronous, one-shot dispatches ideal for parallel fan-outs across files or modules, with slim-ro mode limiting tools to read-only inspection so providers can analyze code safely.

Provider Management and Key Rotation

The TUI (launched via ccf) handles all provider setup. Common vendors ship as presets—DeepSeek, Kimi, Qwen, Groq, Together, Fireworks—with endpoints pre-filled; anything else uses the Custom option. Each provider entry supports default/strong/fast model slots tagged with context length and reasoning effort, so you write "use the strong model" instead of hardcoding IDs. Multi-key rotation (off/round_robin/random) spreads quota across a single provider's keys to avoid rate limits. For Codex users, device-code OAuth login plumbs your ChatGPT subscription into cc-fleet as another provider—though the project warns this may conflict with OpenAI's terms.

The TUI Board

After launching ccf, pressing Tab reveals the Agents Board—a live view of every Workflow, Team, and Subagent organized by project and session. Drill down to see a Workflow's run → phase → leaf progress tree, a team's teammate inbox with pending messages, or individual subagent jobs with their prompts, responses, and spend. Act directly from the board: x stops a leaf or phase, r reruns it, p pins against cleanup, c clears finished entries. The UI follows your system light/dark theme.

Key Takeaways

  • cc-fleet breaks Claude Code's Anthropic-only constraint for Dynamic Workflows, Agent Teams, and Subagents
  • Third-party workers run as real claude processes with their LLM backend swapped—no emulation layer
  • API keys never touch env vars, argv, or shell history; stored at 0600 on disk
  • Interactive ccf run lets you use Claude Code hands-on without a personal Anthropic subscription
  • Agent Teams require tmux and experimental flag CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 (no Windows support yet)

The Bottom Line

This is the kind of project that makes vendor lock-in advocates nervous. cc-fleet doesn't hack Claude Code—it just removes the artificial barriers keeping you inside Anthropic's pricing model. If you've got API credits sitting with DeepSeek or a Codex subscription gathering dust, cc-fleet turns them into workers for a native-feeling multi-agent setup. Worth watching as the ecosystem matures.