Four runtimes. Pick per agent.
A project is not locked to one engine. Each agent inside it picks its own — our compiler loop, Codex, Claude Code, or a key you already have — and that choice is set per agent, not for the whole project.
Start freePlanx
Our own loop — the default
A compiler assembles the system prompt per turn from a library of instruction modules, chosen by what the task actually looks like, then our loop runs the tool calls, the delegation, and the capability checks. This is the runtime every other feature on this site — capability narrowing, per-endpoint connector access, the delegate tree — is built around directly.
Models: Anthropic, OpenAI, Google, or your own OpenAI-compatible endpoint
Codex
OpenAI's own agent, run as-is
This is the real Codex CLI, with its own tool loop and its own context compaction — not a prompt made to imitate it. We drive it over its actual protocol and give it a workspace, rather than reimplementing what it already does well.
Models: OpenAI models by default; a custom OpenAI-compatible endpoint via a wire override
Claude Code
Anthropic's own agent, run as-is
Same idea as Codex, for Anthropic's CLI. Its own tool loop, its own session and compaction behavior, driven over the real Claude Code protocol rather than approximated.
Models: Anthropic models only — that is what the runtime itself supports
Bring your own endpoint
Any key you already have
Point an agent at any OpenAI or Anthropic-compatible endpoint you already pay for. It runs on the Planx loop by default, so it gets the same capability narrowing and delegation as everything else.
Models: Whatever you provide — self-hosted, a third-party provider, an internal proxy
Why run their real binary instead of imitating it
Codex and Claude Code already solved their own tool loop
Reimplementing what a coding agent's own compaction and tool execution already does well is a rewrite with no upside. We drive the real thing over its real protocol instead — a workspace, a capability token, and the actual CLI. When a task calls for our own loop instead — narrower capability control, per-endpoint connector access, or delegating across a whole team — Planx is what those features are written against directly.
One team, several runtimes
A lead agent on any runtime can delegate a subtask to a specialist on a different one — each narrower than the one that started it, in the same run.