All integrations
AI providers: BYOK, hosted & local
AI — /api/ai/*, /api/local/*
Document extraction is not a one-model problem. The best model for a clean digital invoice is different from the best model for a crumpled receipt scan, and the best provider this quarter may not be the best next quarter. Dynamite Docs keeps the workspace constant and the model swappable: bring your own keys to any of 24+ providers, use hosted routing, or run fully local via the Ollama companion.
Try it free or read the API & webhooks docs.
Bundled AI is a pricing and privacy decision made for you
- Most extraction tools bundle one AI vendor, so their model choice decides your per-page cost and your privacy posture, all at once.
- A tool that only offers one hosted model cannot route a scanned statement to a strong vision model and a clean PDF to the free deterministic path.
- Sensitive documents, tax IDs, and contracts should never be routed to a model your team did not approve.
- Lock-in is a feature you pay for twice: once in price, once in the workflow that stops working if the vendor changes terms.
How provider routing works
01 — Connect a key (or use hosted)
Bring your own keys for Gemini, Mistral, OpenAI, Anthropic, Groq, Cloudflare, OpenRouter, and more, or any OpenAI-compatible endpoint. Prefer zero setup? Hosted routing picks the best available provider automatically.
02 — Set routing rules
AI Settings exposes rules: Sensitive Documents mode blocks training-permitted providers by default, plus no-training-only and local-only policies.
03 — Extract, per document
Clean digital PDFs parse deterministically from their text layer, no AI call at all. Scans and photos route to the vision model you chose. Every field is confidence-scored as usual.
04 — Pay your provider, not us
BYOK inference carries no markup. You pay your provider’s rate. On Pro and Ultra, BYOK processing is unlimited and never counts against your monthly allowance.
The provider surface
24+ providers are auto-registered; keys are encrypted and used only to route your own requests:
- Built-in providers — Gemini, Mistral, OpenAI, Anthropic, Groq, Cloudflare, Cohere, xAI (Grok), and more
- Hosted routing — best available provider picked automatically
- Cloud & enterprise — Azure OpenAI, Amazon Bedrock, Google Vertex AI
- OpenAI-compatible endpoints — any provider, any endpoint, or your own
- Local Ollama companion — fully offline extraction on your machine
- Routing policies — sensitive, no-training-only, free-only, local-only
- Key security — AES-256-GCM encrypted, used only for your requests
- Deterministic path — clean digital PDFs parse for free, no AI
A realistic example
The routing decision on a mixed batch:
| Document | Path | Why |
|---|
| Clean digital invoice PDF | Deterministic parse | Text layer reads perfectly, no AI cost, identical every time |
| Scanned bank statement | Your vision key | Best model for the worst ink, at your provider’s rate |
| Client contract (sensitive) | Sensitive mode / local | Training-permitted providers blocked; local Ollama if needed |
Model choice is the cost, accuracy, and privacy decision
The per-page cost of extraction is not a fixed fee. It is the price of a specific model, and the right model differs by document. A clean, text-layer PDF needs no model at all; Dynamite Docs parses it deterministically, identically every time, for free. A scanned statement needs a vision model strong enough to read bad ink. A contract with commercially sensitive terms needs a provider you explicitly trust. One bundled model cannot serve all three well, and a tool that forces it is quietly making your cost and privacy decisions for you.
BYOK is the counterweight. Connect your own keys and the workspace routes your requests through your provider account, at your provider’s published rate, with zero markup. On Free trial and Hobby plan your own-key draws share the monthly allowance; on Pro and Ultra BYOK processing is unlimited. That structure is deliberate. It keeps the hosted product honest while making heavy, own-infrastructure workloads cheap at volume.
Provider keys are encrypted server-side with AES-256-GCM and exist for one purpose, routing your own extraction requests. The app never stores them in plaintext, and the client never sees them. Sensitive Documents mode and the routing rules in AI Settings go further, letting you block entire classes of provider (anything training-permitted, for example) so a wrong autopilot choice cannot send a tax ID to a model you did not approve.
Local is the fully offline option. The Ollama companion runs extraction entirely on your machine on port 8756, with every route protected by a signed local token and loopback-only origins. For air-gapped work or client agreements that forbid cloud processing, local mode keeps the full infer → correct → remember loop with no cloud contact at all.
What provider routing does and does not
Does
- Routes extraction through the provider and model you choose, hosted or BYOK, with no markup on inference.
- Encrypts provider keys and uses them only to route your own requests.
- Offers routing policies (sensitive, no-training-only, free-only, local-only) that lock extraction to providers you trust.
Does not
- Does not sell or resell your documents to AI vendors. Your documents are never used to train foundation models.
- Does not let hosted routing silently bypass your Sensitive Documents or no-training policies.
- Does not store provider keys in plaintext or expose them to the client.
Why model choice is a baseline feature
Extraction quality and per-page cost are properties of the model, and only you should choose the model that sets those trade-offs.
BYOK with zero markup means heavy extraction workloads pay the market rate for inference, not a reseller’s margin.
Sensitive-mode and local routing make the privacy decision enforceable policy, not a hope.
Questions, answered
Which providers are supported?
24+ providers are built in, Gemini, Mistral, OpenAI, Anthropic, Groq, Cloudflare, OpenRouter, Cohere, xAI (Grok), and more, plus any OpenAI-compatible endpoint (including Azure OpenAI, Amazon Bedrock, and Google Vertex AI) and a local Ollama companion.
Is BYOK unlimited?
On Pro and Ultra, yes. BYOK processing is unlimited and never counts against your monthly allowance. On Free trial and Hobby plan, your own-key draws share the same monthly bucket as hosted.
Do you add markup to inference?
No. When you bring your own key, you pay your provider’s rate directly; there is no markup on inference.
Can I run extraction fully offline?
Yes. The local Ollama companion runs the extraction loop on your machine with token-signed, loopback-only routes, no cloud contact.
Document workflows it feeds
Open app, no payment details.