Features
Everything NXBrain does,
and the plan that includes it.
62 features across routing, data protection, cost control, agents, change management, privacy, proof and compliance. Each plan includes everything in the plans before it.
Connect & route
One OpenAI-compatible endpoint in front of every model you use, in the cloud or on your own servers.
| Feature | What it does | Plan |
|---|---|---|
| OpenAI-compatible API | POST /v1/chat/completions with OpenAI’s request and response format, streaming included. Point your SDK at a new base URL. | Basic |
| Abstract models | Your code calls nxbrain/reasoning or nxbrain/fast. You decide which real model is behind each name. | Basic |
| Versioned routes | Each abstract model has a route: destinations with priority, weight, timeout and retries. A published version never changes. | Basic |
| Fallback and circuit breaker | Retryable errors move to the next destination. A gateway that fails three times in a row is skipped for 60 seconds. | Basic |
| Conditional routing | Project policies decide which models and gateways each project may use. The route skips any destination that doesn’t qualify and records why. | Pro |
| Your own provider keys | Use your OpenAI, Anthropic or Google keys. They are encrypted with your organization’s key, and we add no markup. | Basic |
| Self-hosted models | Register your own OpenAI-compatible servers, such as Ollama, vLLM or llama.cpp, over HTTPS with their own key. | Basic |
| Route simulator | See which destination a request would take, and why the others were left out, without making the call. | Basic |
Protect data
Inspect every request and response, and decide per project what happens when something sensitive shows up.
| Feature | What it does | Plan |
|---|---|---|
| Basic guardrails | Templates for personal data and secrets: email, phone, payment cards, IBAN, API keys and private keys. | Basic |
| All guardrails and your own rules | Prompt injection indicators, moderation by category, your own banned terms, a maximum input size and JSON output validation. Each detector can observe, warn, redact or block. | Starter |
| Input and output, streams included | Requests are checked before they reach the model, and responses before they reach your app. Streams are inspected as they arrive. | Basic |
| Incidents without content | Every guardrail hit is logged with its detector, stage and decision. The text itself is never stored. | Basic |
| Custom detectors | Detectors built for your data and your sector. | Enterprise |
| Keys per app, team or agent | Virtual keys with expiry, a rotation date and instant revocation. Archiving a project revokes all its keys at once. | Basic |
| Project policies | Allowed models and gateways, maximum input and output tokens, and whether prompts must come from the catalog. | Basic |
| Rate and concurrency limits | Requests and tokens per minute for each key, project and organization, answered with Retry-After when a limit is hit. | Starter |
| Roles and two-factor sign-in | Admin, AI operator, FinOps, auditor and viewer roles, checked on the server, with optional two-factor codes. | Basic |
| Single sign-on | Entra ID, Google, Okta or any OIDC provider on your verified domains, with SSO-only sign-in if you want it. | Pro |
Control costs
Every call is reserved against its budget before it runs and settled after it finishes, so a budget is a limit, not a report.
| Feature | What it does | Plan |
|---|---|---|
| Hard budgets | A budget per project and per agent, reserved before each call so parallel requests can’t overspend it. When it runs out, calls return 402 BUDGET_EXCEEDED. | Starter |
| Alerts at 50, 75, 90 and 100% | A metric, an alert and an email at each threshold. | Starter |
| Cost by project, model, gateway and day | Spend and tokens in real time, metered in millionths of a euro. | Basic |
| Blocked calls cost nothing | A request stopped by a guardrail never reaches the provider and never touches the budget. | Basic |
| Access requests | Teams ask for a model, more budget, a tool or a project. Someone else decides, and an approval is a decision, not an automatic change. | Starter |
Agents
Give each agent its own keys, models, tools and budget, and keep a person in the loop for what matters.
| Feature | What it does | Plan |
|---|---|---|
| Agent catalog | Purpose, authorized models, status, limits per run (steps, duration, tokens and cost) and a running budget. | Pro |
| Tool permissions | Allow or deny each tool for each agent, MCP tools included. Disabling a tool cuts it off for every agent at once. | Pro |
| Human approval | Sensitive tools wait for a person. Each approval works once and expires. | Pro |
| Kill switch | Pause or stop an agent and every key linked to it is rejected instantly. | Pro |
| Execution traces | Steps, tool calls and hand-offs between agents, recorded as metadata only. | Pro |
Ship changes safely
Routes, policies, guardrails and prompts change through the same reviewed process, and nothing is edited in place.
| Feature | What it does | Plan |
|---|---|---|
| Change center | Draft, validate, approve and publish policies, routes and guardrails. The author can’t approve their own change. | Basic |
| Versions and restore | Published configuration never changes. Restoring an older version creates a new one. | Basic |
| Prompt catalog | System prompts used by name, with versions, declared variables, two-person publishing and rollback. The text is encrypted. | Starter |
| Evaluations | Versioned datasets; deterministic, model-based or external evaluators; weighted scorecards. Run them by hand, from CI or on a schedule. | Pro |
| Promotion gates | A protected route or prompt only goes live with a recent evaluation that meets your minimum score and maximum regression. | Pro |
| Playground | Try models and prompt versions with a real key’s identity, under the same policy, guardrails and budget as production. | Basic |
Privacy
We record what happened, not what was said.
| Feature | What it does | Plan |
|---|---|---|
| No prompts stored by default | Only metadata, counts, hashes and IDs are kept. | Basic |
| Opt-in encrypted capture | Keep the full text if you need it: encrypted with AES-256-GCM, deleted automatically after the days you choose, every read logged. | Pro |
| Your own encryption key | Provider keys, captured content, evidence and prompts are encrypted with a key that belongs to your organization. | Basic |
| Isolation between organizations | Row-level security on every table that holds customer data. | Basic |
| Export everything | Your full sealed history in an open format, verifiable without us. | Basic |
Proof
Every call, change and approval becomes evidence that anyone can check.
| Feature | What it does | Plan |
|---|---|---|
| Sealed audit log | Every five minutes, new events are sealed into a Merkle chain (RFC 6962). Altered or deleted events are detected. | Basic |
| Public verification | Check a seal, an event or a report at /verify in the browser, or offline with the nxbrain-verify tool. No account needed. | Basic |
| Trusted timestamps | Each sealing pass can be timestamped by an independent authority (RFC 3161), with a daily locked copy of the seals. | Pro |
| Inclusion proofs and export | Prove that one event is in a sealed batch, and export any period as NDJSON. | Basic |
| Systems outside the gateway | Send events from other AI systems with POST /v1/events, OpenTelemetry traces or the TypeScript and Python SDKs. | Basic |
Compliance
Organize AI compliance around what actually runs. A person always makes the final call.
| Feature | What it does | Plan |
|---|---|---|
| AI inventory | Every AI system, connected or not: purpose, role, data, providers and DPAs, linked to projects and agents, with 30 days of real activity. CSV import and export. | Basic |
| Risk classification | An EU AI Act questionnaire with an explainable result that cites its articles. It is indicative, and two people approve it. | Basic |
| Obligations | Each with an owner, a start date and a due date, and reminders by email and webhook. | Starter |
| Impact assessments and AI policy | DPIA, FRIA, transparency notice and AI policy, drafted from your inventory and approved by two people. | Starter |
| Reports without watermark | PDF reports per system and per organization. On Basic they carry a watermark. | Starter |
| Technical documentation | EU AI Act Annex IV documentation drafted from your inventory and classification. | Pro |
| ISO/IEC 42001 | Statement of applicability for the 38 Annex A controls, with evidence collected from real activity. | Pro |
| Incidents | AI Act incident categories with their notification deadlines (2, 10 or 15 days), followed up to closure. | Pro |
| AI literacy | Training records per person, with expiry (AI Act, article 4). | Pro |
| Evidence packs | A ZIP with a manifest. Every file is sealed and anyone can verify it. | Pro |
Developer tools
Wire NXBrain into the tools you already run.
| Feature | What it does | Plan |
|---|---|---|
| Webhooks | Signed with HMAC-SHA256 for access requests, compliance, prompt and team events, with retries and delivery history. | Pro |
| Automation API | Tokens with scopes, expiry and revocation to read configuration and manage access requests. Documented in OpenAPI. | Pro |
| SDKs | TypeScript and Python, to send events from systems outside the gateway. | Basic |
| OpenAPI contract | The full contract at /openapi.json. | Basic |
Enterprise
For regulated industries, built to fit.
| Feature | What it does | Plan |
|---|---|---|
| Dedicated deployment | NXBrain deployed for your organization only. | Enterprise |
| Custom limits and retention | Limits and retention set in your contract. | Enterprise |
| SLA and dedicated support | A service-level agreement and a named support contact. | Enterprise |
| Assisted review | We review your setup with your team. | Enterprise |
Plan limits
Plans differ in how much you can govern and how long the record is kept. Priced by AI systems and usage, not per seat: each plan includes its users. No markup on your model keys.
| Basic€10/mo · €0 during the beta | Starter€79/mo · €790/yr | Pro€399/mo · €3,990/yr | EnterpriseCustom | |
|---|---|---|---|---|
| AI systems | 3 | 10 | 50 | Custom |
| Users | 2 | 5 | 20 | Custom |
| Events a month | 10,000 | 250,000, then €20 per 100k | 2M, then €15 per 100k | Custom |
| Retention | 7 days | 30 days | 12 months | Custom |
High-risk AI systems keep their activity for at least 183 days, as the EU AI Act requires, within your plan’s retention.
Start with one URL.