Skip to content

Phoenix 1.9.0 - Bring your own brain

Choose a tag to compare

@erazz91-ai erazz91-ai released this 09 Oct 23:54
· 2 commits to master since this release

Phoenix is a creative's control room for local AI — a free, open‑source harness that drives your local generation toolchain from one browser workbench. 1.9.0 is about one idea: bring your own brain.

✨ Highlights

A provider per seat

Every role in Phoenix — orchestrator, troubleshooter, meta‑prompter, vision and category provider — now picks its own model source: your Claude subscription (Claude CLI), your ChatGPT plan (Codex CLI), a local OpenAI‑compatible server (LM Studio, Ollama, llama.cpp) or an API key. Put a big model where it matters and a small, fast one where it doesn't. Models that drive tools have to pass a tool test first (local ones also a short benchmark), and a non‑Claude vision seat has to pass a vision test — a colour card proves it really looked at the render before it may judge one. Keys live in .secrets/ only, never in the config, a response or a log.

Ask / Auto

Ask stops at the gates you switched on. Auto runs the job to the end within a step budget you set, and stops only when the render check doesn't clearly confirm a step.

Checkpoints + rewind

Before each scene‑changing step Phoenix saves a copy of your open Blender scene, together with the chat at that moment. Rewind to any checkpoint — and undo the rewind if you change your mind.

Phoenix as an MCP server

An external agent (Claude Code, Codex, Cursor) can call Phoenix as a tool: generate props, import them and check the result with Phoenix's render vision.

Remesh via Meshy (optional, paid)

Rebuild a staged mesh's topology in Meshy's cloud. A quad result gets its quads restored on import (GLB can only store triangles). Off by default; nothing is uploaded until you enable it and press the button.

And also

  • GPU share (best effort) — before an image or mesh stage Phoenix asks a local LLM server on the same card to unload its model (waits for confirmation on the llama.cpp router).
  • Resizable, collapsible panels; Settings split into General and Connections.
  • Hardening from the release review: same‑origin protection on every state‑changing route, a strict argument allowlist for the Claude CLI, persistent fail‑closed bookkeeping for paid remesh jobs.
  • The default orchestrator is now opus — the alias that always follows the newest Opus.

Getting started

git clone https://github.com/laconova/phoenix.git
cd phoenix
node server.js

See the README for ComfyUI, Blender and Unreal setup, and the ROADMAP for what's next.

🎬 Watch what's new: https://youtu.be/EuxETcEaCvE

Status: alpha — early and evolving. Issues and feedback welcome, here or via laconova.com.