mcp: Claude Code paints with a ChatGPT subscription by default; local agents and endpoints

mcp/README.md

@@ -14,11 +14,13 @@ budgets. The rules the agent follows are in [`AGENTS.md`](viggle_remake_mcp/data
 
 **What you need:**
 - **an agent**: Claude Code, codex, or any MCP client;
-- **a painter** for the keyframes: either a codex login with a ChatGPT subscription (`REMAKE_PAINT=pool`, no API
-  spend), or a key for an OpenAI-compatible images API (the default; see [Cost](#cost));
+- **a painter** for the keyframes: `codex` logged in with a ChatGPT subscription (`REMAKE_PAINT=pool`: no API key,
+  no API spend; the Claude Code and codex setups below use it), or an OpenAI-compatible images API
+  (`OPENAI_API_KEY`, billed per image; see [Cost](#cost));
 - **render workers**: `viggle-remake-worker` on a card with 96 GB or more, or the
   [ComfyUI worker](#one-32-gb-card-the-comfyui-worker) on one 32 GB card such as an RTX 5090 (Linux, about 60 GB of
-  RAM). The server itself needs no GPU. Cards below 32 GB are not tested with the agent; use ComfyUI by hand
+  RAM). The server itself needs no GPU, but a worker uses nearly all of its card's memory, so a local model as the
+  agent needs another GPU or another machine. Cards below 32 GB are not tested with the agent; use ComfyUI by hand
   ([`COMFYUI.md`](../COMFYUI.md#smaller-gpus-int8-and-gguf)).
 
 ## How it fits together
@@ -37,6 +39,9 @@ budgets. The rules the agent follows are in [`AGENTS.md`](viggle_remake_mcp/data
   share a directory, so they can run on different machines over a shared filesystem.
 - **Keyframes** are painted by an image model, in one of three backends (`REMAKE_PAINT`):
   - unset: an OpenAI-compatible images API (`gpt-image-2` by default). The only backend with API spend.
+    `OPENAI_BASE_URL` points it at another server that implements `images.edit` with several input images and
+    returns `b64_json`, such as one you host (not tested with local models). A server that reports no usage is counted
+    at $0.15 a picture against `REMAKE_MAX_USD`.
   - `codex`, *codex image mode*: the agent must be codex. An image tool's first call returns a CODEX IMAGE STEP, codex
     paints with its own `image_gen` tool, and the second call of the same tool imports the picture.
   - `pool`, *pool image mode*: the server starts a helper `codex exec` per picture (its own codex home, no MCP,
@@ -262,8 +267,10 @@ face detector and recognizer you are licensed to use.
 1. Register the server. Copy [`examples/claude_mcp.json`](examples/claude_mcp.json) to `.mcp.json` in your project,
    or run:
    ```bash
-   claude mcp add viggle-remake -e REMAKE_HOME=$HOME/.viggle-remake -e OPENAI_API_KEY=$OPENAI_API_KEY -- viggle-remake-mcp
+   claude mcp add viggle-remake -e REMAKE_HOME=$HOME/.viggle-remake -e REMAKE_PAINT=pool -- viggle-remake-mcp
    ```
+   Both paint in pool image mode: `codex` on `PATH`, logged in with a ChatGPT subscription, and no API key. To paint
+   through an images API instead, drop `REMAKE_PAINT` and pass `-e OPENAI_API_KEY=$OPENAI_API_KEY`.
 2. Give the agent the rules: put `@AGENTS.md` in the project's `CLAUDE.md`, or copy the file in.
 3. Raise the tool timeout. `wait` blocks for up to 30 minutes and `paint` can take a few, so start Claude Code with
    `MCP_TOOL_TIMEOUT=4200000` (milliseconds).

mcp/examples/claude_mcp.json

@@ -3,7 +3,7 @@
     "viggle-remake": {
       "command": "viggle-remake-mcp",
       "args": [],
-      "env": {"REMAKE_HOME": "${REMAKE_HOME:-~/.viggle-remake}"}
+      "env": {"REMAKE_HOME": "${REMAKE_HOME:-~/.viggle-remake}", "REMAKE_PAINT": "pool"}
     }
   }
 }