MinTK
MinTK sees and proves. It never silently changes.
MinTK sits between your applications and the AI providers or your company's own cloud. A request that names a model is forwarded byte-for-byte or refused, never edited. Every response carries SHA-256 evidence of what arrived and what was forwarded, every request is recorded in a signed, per-tenant audit trail you can verify offline, and usage is measured from what the provider reports, never estimated. MinTK is ready for a proof of concept in your environment, and the validation status below shows exactly what has been tested so far.
Requests that name a model are forwarded byte-for-byte. Only Smart Route may choose a model, and it says so.
SHA-256 hashes of what arrived and what was forwarded, on every response.
A signed, per-tenant audit chain you can verify offline with mintk audit verify.
Requests containing API keys, private keys or tokens are blocked before they leave your network.
What MinTK does for every request
Forward or refuse
Requests that name a model are forwarded byte-for-byte or refused with a clear error. MinTK does not guess.
Evidence
Each response returns SHA-256 hashes of the request received and the request forwarded.
Signed audit trail
Audit events are HMAC-chained per tenant. Editing, deleting or reordering a line fails verification.
Measured usage
Token usage is taken from what the provider reports, never estimated.
Credential blocking
Requests that contain API keys, private keys or tokens are stopped before egress.
Retired-model warnings
Callers are warned when they request a model its provider has retired. The model is never changed.
Smart Route, when you ask for it
Applications that send model "mintk-auto" let MinTK pick the best model across your connected providers. It falls back automatically if a provider fails, explains every choice, and marks the response so the change is never hidden.
Your providers
OpenAI, Anthropic, Google Gemini, DeepSeek, xAI Grok, Groq and Mistral, plus OpenRouter for pass-through.
Your company cloud
Azure OpenAI, Google Vertex AI (Gemini and Claude) and Amazon Bedrock (Claude), signing in with the cloud's own identity.
Local models
Ollama models as a free, private option and the final fallback when cloud providers fail.
Two ways to run it
Enterprise
- Runs behind your ingress or VPC, or with mTLS.
- Tenant API keys, identity asserted by your SSO gateway, or mTLS.
- Provider keys from your secret manager, or callers bring their own.
- Per-tenant audit chain, exported to your SIEM through stdout.
- Docker image and a GKE deployment with US and India overlays.
Desktop
- Listens on 127.0.0.1 only.
- Native macOS app with a dashboard: Overview, Try a prompt, Keys, Connect apps, Audit.
- Provider keys stored in the OS keychain.
- Works with a free Google Gemini key or a free local model through Ollama.
- One static binary for macOS, Linux and Windows.
Built to be checked
No third-party code
Built only on the Go standard library.
Fail-closed
Unsafe configuration stops startup. If the audit trail cannot be written, new requests are refused.
No sensitive content stored
Audit and logs hold metadata and hashes only: no prompts, completions or keys.
Data-location lock
On Vertex AI, requests for locations you have not allowed are refused before they leave MinTK.
Validation status
| Area | Status |
|---|---|
| OpenAI, Anthropic and Gemini APIs, buffered and streaming | Supported |
| Byte-exact forwarding, SHA-256 evidence, signed audit chain | Supported and tested |
| Azure OpenAI, Vertex AI and Bedrock paths | Tested end to end against local stand-ins; not yet against the live cloud services |
| macOS Keychain | Verified on macOS (Intel) |
| Windows Credential Manager | Built, not yet verified on Windows |
| Independent penetration test, code signing, load testing at customer scale, compliance certifications | Not yet done |
From MinTK's own support matrix and security model. We do not claim anything on this list until it is done.
Try MinTK with your own data
See the sample demo first, then tell us about your environment. We agree the scope with you before anything is installed.