Claude (Symbiotica)

Claude (Symbiotica) is a text-in, text-out LLM node that sits in your workflow as a prompt author, a caption or critique step, or a structured-extraction stage feeding an image node. One thing it is not: an image generator. "Claude draws nothing" is the pack's own description, and it's worth taking literally - this is the LLM-in-the-graph pattern, not a diffusion stand-in. You wire it up where you want a smart reader and writer, not where you want pixels.

The model choice is where this node gets interesting. The dropdown is a dynamic combo: pick a model and the node shows only the inputs that model actually accepts. Opus 5 and Fable 5 reason unconditionally, so they offer a reasoning_effort with no "off" option; Haiku 4.5 has no reasoning input at all; models that reject temperature alongside thinking simply don't show a temperature widget. It's a small ergonomic thing that saves you from the classic "I set a widget and the API 400s" surprise. Haiku is the cheap lever for bulk captioning, per the node's own tooltip.

How it works

The node takes a prompt, an optional system prompt, and up to twenty reference images, and returns Claude's answer as a STRING. Reference images are described by the prompt, not replaced by it - you tell Claude what to do with the pictures. The seed input is a lie on purpose: Anthropic takes no seed, so the widget exists purely to defeat ComfyUI's output caching - bump it (or let it auto-increment) and re-queueing actually re-runs the node.

Billing is the interesting part. On a box configured with the pack's Cloudflare AI Gateway environment variables (SYMBIOTICA_AIG_BASE etc.), every call routes through the gateway on the studio's own key - which is how this pack runs order pipelines headless with per-studio spend tracking. Anywhere else, it calls Anthropic directly with a key from the node widget, the Settings UI, or the ANTHROPIC_API_KEY / CLAUDE_API_KEY env var, in that order. For a beginner, ignore the gateway entirely: paste your key into the node or Settings and go.

The inputs that matter

• prompt - what to ask. Since reference images ride along, this is where you say "caption this in 50 words" or "extract the lighting and composition as JSON."
• model - the dynamic combo described above. Haiku for cheap bulk work, Sonnet for quality, Opus/Fable when you want the reasoning.
• system_prompt - standing instructions, sent as Anthropic's real top-level system field, not a fake user message.
• images - reference slots, filled as you wire them in (up to 20). Each slot can itself carry a batch.
• max_tokens - a budget, not a target. The node's docs are blunt: on thinking models this caps reasoning and answer together, so a small value cuts the answer off - and this node raises an error on a truncated reply rather than silently handing you a fragment.

The single output, text, is a plain STRING. Wire it to a text viewer, a save node, or another prompt stage.

Install

ComfyUI Manager → search Symbiotica → install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/symbiotica-ai/comfyui-nodes.git symbiotica
pip install -r symbiotica/requirements.txt

Restart ComfyUI. No models download - it's a cloud call. You need an Anthropic key (env var, Settings, or node widget). The pack loads without any key.

Common issues

• "Empty response from Claude model." The node deliberately raises on an empty reply rather than returning that literal string like ComfyUI's own node does - a refusal, a cut-off answer, an oversized input, and an empty reply are four separate errors with four separate fixes. Read the actual error message.
• Large reference images refused. The pack downsizes references to the model's own ceiling first (2576px on Opus 5/Sonnet 5/Fable 5, 1568px elsewhere) and refuses a batch over ~8MB - that's a Cloudflare gateway log limit, not a whim. Feed smaller images.
• Re-queueing returns the same answer. That's ComfyUI's cache; click the seed's refresh control so the node actually re-runs.

This is an API-wrapper node from a pack with no real community footprint yet - same category that shipped credential-stealing malware once (LLMVISION), so the standing advice applies: read the source before you paste a key in, and it's MIT-licensed and open, so you can.