Nine providers.
Swappable per tool.
The old paradigm is out. The new paradigm is AI. AI automation is InTouch AI. Here's the proof of it: AI isn't a connector bolted onto the side. It's the architecture. InTouch was built with AI at the center, and the vault, scheduling, RBAC, and audit sit behind it — which is exactly why nine providers drop in as first-class tools instead of one hard-wired vendor lock.
Anthropic Claude, OpenAI, Mistral, Groq, DeepSeek, xAI Grok, Google Gemini, and Ollama (local) — all eight ship as native task types in every edition. Hugging Face rounds out the set as a 9th, job-only provider. Each is a first-class tool — not an abstraction layer papering over differences. Use Claude's extended thinking where it helps, OpenAI's structured outputs where they help, Gemini's multi-modal where relevant, Groq for sub-second LPU inference, DeepSeek/Mistral for OpenAI-compatible cost-down options, xAI Grok for the Grok roster, and Ollama for anything that needs to stay inside your network. A general AI-native engine does what a single-vendor wrapper does — and the wrapper can never grow what InTouch already has at its core.
Each Provider, With Its Distinctive Surface
Config-era tools pick a model for you and call it a feature. InTouch puts every surface on the table and lets the job choose — because the engine is AI-native, not AI-retrofitted.
Anthropic Claude (5 tool variants)
- Claude Agent — multi-turn with tool use
- Claude Batch — async batch API for high-volume
- Claude Document — PDF and document reading
- Claude Tools — direct tool-use without agent loop
- Claude (base) — simple prompt-completion
Valid model IDs: claude-opus-4-6, claude-sonnet-4-6, claude-haiku-4-5-20251001.
OpenAI
- Chat completions with function calling
- Structured outputs / JSON mode
- Embeddings (for the RAG pipeline)
- Fine-tuning management (for fine-tuned models)
- Image generation (DALL-E)
- Whisper speech-to-text, TTS text-to-speech
Valid model IDs: gpt-4o, gpt-4o-mini.
Google Gemini
- Text and multi-modal (image, document)
- Function calling
- Context caching for long-context prompts
Valid model IDs: gemini-2.5-flash, gemini-2.5-pro.
Ollama (Local)
- Any model the operator installs locally (
llama3.2,mistral,gemma2, your own fine-tunes) - Zero restrictions, zero API cost, no data leaving the network
- Ideal for regulated domains (finance, legal, medical) and air-gapped deployments
Model names are operator-set — no fixed dropdown.
Mistral, Groq, DeepSeek, xAI
- Mistral —
mistral-large-latestand the Mistral roster viaapi.mistral.ai/v1 - Groq — sub-second LPU inference;
llama-3.3-70b-versatileand the Groq-hosted catalogue viaapi.groq.com/openai/v1 - DeepSeek —
deepseek-chatviaapi.deepseek.com/v1 - xAI — Grok via
api.x.ai/v1(defaultgrok-2)
All four use the OpenAI-compatible chat completions surface. Same task definition, swap the tool name and credential to switch provider.
Hugging Face (Job-only)
Hugging Face Inference is wired in as a job task type for batch — not as an Assistant chat backend. Useful when you want to call a specific HF model from a scheduled job (classification, embeddings, summarization) without standing up an inference server.
The Defaults Are the Right Defaults
This is the contract, now intelligent: tell it what to do, and what to do when it doesn't work. The "doesn't work" clause stopped being a dumb rule and became an assessment — it reads the failure, refreshes an expired token, smart-retries, and surfaces the one sentence that matters. Your API key never appears in any of it.
Safety Preamble
Every AI call prepends a safety preamble resisting prompt injection and enforcing output format. Override per tool when you need bare prompts; on by default.
Rate-Limit Handling
429s retried with exponential backoff and jitter. No hand-rolled retry loops. Invisible when it's working, diagnostic when it's not. It broke, here's why, it fixed it.
Audit Log
Every call logs provider, model, input tokens, output tokens, response. Budget tracking via SQL query on the audit table. Credentials stay AES-256 in the vault — referenced by name, never in a script, never exposed even to the AI itself.
"If Anthropic Refuses, Swap to Ollama"
Anthropic (and OpenAI, and Gemini) have content policies. When a domain crosses a refusal line — financial advice, legal synthesis, medical reasoning — swap the tool reference to Ollama with a local model. Zero restrictions, zero API cost, no data leaving the network. The job definition doesn't change; only the tool name and credential do.
That same freedom is a dial. Run a workflow on pure AI while you're still earning trust in it, then graduate it toward fully deterministic — zero-AI-cost, identical every time, fully audited — once it has proven itself. Nobody else gives you the dial, because nobody else put the AI at the center to begin with.
Pick the Right AI for the Job
Not the one your SaaS vendor chose for you. That's the difference between an engine built AI-native and a config-era tool with AI stapled on — and you can't staple on a center. Free Personal edition, all 8 native providers included (plus Hugging Face as a 9th, job-only).