EffectCraft AI Chat

An AI chat tab inside EffectCraft — the open-source, pure-Rust After Effects alternative. Chat with a local model, or with whatever agents you already have: Codex on a ChatGPT plan, Devin over ACP, OpenCode, any OpenAI-compatible API, any CLI.

No fork of EffectCraft, no recompiling, no changes to the app. It's a ScriptUI panel — the app's own extension mechanism — plus a small local bridge that holds the connections.

Community project. Not affiliated with ArtCraft or Adobe. EffectCraft is MIT/Apache-2.0; this repo is independent.

Quick start (panel + local model)

1. Ollama running with a model: ollama serve, ollama pull llama3.2:3b.

2. In EffectCraft: Settings ▸ Scripting & Expressions ▸ Allow Scripts to Write Files and Access Network — on. (Scripts can't use the network without it.)

3. Copy the panel in:

   # macOS settings folder; Windows: %APPDATA%\EffectCraft; Linux: ~/.config/effectcraft
   mkdir -p "$HOME/Library/Application Support/EffectCraft/Scripts/ScriptUI Panels"
   cp "extensions/AI Chat.jsx" "$HOME/Library/Application Support/EffectCraft/Scripts/ScriptUI Panels/"

   (Or use File ▸ Scripts ▸ Install ScriptUI Panel….)

4. Restart EffectCraft → Window ▸ AI Chat.jsx → the tab docks anywhere.

Connections (the bridge)

The panel's scripting Socket API is plain TCP with no TLS and it can't spawn processes, so it talks to a local bridge over HTTP on 127.0.0.1:8765. The bridge is what reaches everything else — cloud APIs over HTTPS, and CLIs:

python3 bridge/agent-bridge.py --config bridge/agents.json   # copy agents.example.json first

Kind	What it connects to	Example config
ollama	Local models (also the panel's direct mode)	{"kind":"ollama","url":"http://127.0.0.1:11434","model":"llama3.2:3b"}
command	Any CLI agent	["codex","exec","--skip-git-repo-check","{prompt}"], ["opencode","run","{prompt}"]
acp	ACP agents over stdio	["devin","acp"], ["opencode","acp"]
openai	OpenAI-compatible APIs	{"kind":"openai","base_url":"https://api.openai.com/v1","model":"gpt-4o-mini","key":"pass:openai"}

Keys come from pass (pass:name) or the environment (env:NAME) — never stored in the panel or the repo. Add or remove agents in agents.json, hit Refresh in the panel, and the picker updates.

The four adapters above are verified working (see Testing): a local model, Codex with a ChatGPT plan, Devin over ACP, and OpenCode — each answered through the bridge on the maintainer's machine.

Pick a model

In the EffectCraft AI Chat panel, click Settings (or Settings ▸ AI Chat) and set Agent Model — Agent Default, Claude Opus, Claude Sonnet, Claude Haiku, GPT or Gemini. The panel sends the choice with each chat request and the bridge resolves it per agent kind:

• ollama / openai: the request's model overrides the configured model.
• command: the model is substituted into the agent's model_args (["-m", "{model}"], appended to the argv). An agent without model_args answers "can't switch models".
• acp: the model is matched against what the agent offered in session/new — an exact value match, else the first option whose value or name contains it (opus → the first Claude Opus) — and applied with session/set_config_option.

The reply reports the model that answered, shown in the panel's status line (e.g. ready · claude-opus-5-5-medium).

How it works

EffectCraft ── ScriptUI panel ── plain HTTP ──▶ 127.0.0.1:8765  (agent-bridge.py)
                                                   ├── ollama      → local model
                                                   ├── command     → codex / opencode / ...
                                                   ├── acp         → devin acp (JSON-RPC over stdio)
                                                   └── openai      → HTTPS + key from pass/env

• The panel is one .jsx file using EffectCraft's documented ScriptUI + Socket APIs. It can't see the project yet — it's a chat, not an editor (see Roadmap).
• The bridge is one Python file, standard library only. It binds 127.0.0.1 and never logs message contents.

Testing

python3 tests/run-panel-e2e.py

Starts the bridge with a test config, then runs the real panel file inside EffectCraft's own script engine (headless, with the network gate set in-session): it checks that the panel builds, discovers the bridge's agents, gets a reply from Ollama directly, and gets a reply from the bridge. 7 checks, all green on macOS with EffectCraft 0.8.0.

Layout

Path	What
extensions/AI Chat.jsx	the panel — install into Scripts/ScriptUI Panels/
bridge/agent-bridge.py	the connections: ollama / openai / command / acp
bridge/agents.example.json	copy to agents.json and edit
tests/	headless end-to-end test
scripts/install-cli.sh	installs effectcraft-cli (for the tests and MCP)

Roadmap

• Streaming replies (today a request blocks until the agent finishes).
• Let the agent edit the project: the scripting API can, the panel just doesn't ask it to yet.
• A one-click installer for the panel.

License

MIT. See LICENSE.