All features & supported AI providers
The complete catalog of ThinkNEO. 19 AI providers with connectors — from OpenAI and Anthropic to self-hosted models — and 80+ features across governance, FinOps, observability, agents/MCP, security and compliance. No sales call to read the fine print.
Supported AI providers
One base URL, one BYOK key, every provider below. ThinkNEO auto-converts request format between the OpenAI schema (default) and each provider's native API (Anthropic /v1/messages, Google Gemini, Cohere) so client code doesn't change.
OpenAI
Chat Completions, Assistants API, embeddings, image, moderation
Anthropic Claude
Claude 4.x + Haiku via /v1/messages, format auto-converted from OpenAI
Mistral
Mistral Large/Small via OpenAI-compat /v1/chat/completions
DeepSeek
DeepSeek-V3 / R1 via OpenAI-compat endpoint
Google Gemini
Gemini 1.5/2.0 Pro & Flash via generativelanguage.googleapis.com
Azure OpenAI
Azure-hosted OpenAI deployments; same request shape, different endpoint
xAI (Grok)
Grok models via OpenAI-compat endpoint
Cohere
Command R+/R, embeddings, rerank via native API
Meta Llama
Llama 3.x served via Together / Fireworks / Groq / Bedrock
OpenRouter
Aggregator — 200+ models exposed through one BYOK key
Groq
Ultra-low-latency open models (Llama, Mixtral) via LPU
Together AI
Open-model inference (Llama, Qwen, Mistral) via OpenAI-compat
Fireworks AI
Serverless open-model inference, OpenAI-compat drop-in
Ollama
Self-hosted local models via Ollama's OpenAI-compat endpoint
Self-hosted
Bring any OpenAI-compat endpoint (vLLM, LocalAI, TGI, LiteLLM…)
AWS Bedrock
Native SigV4 signing + Bedrock model IDs — Q4 2026
Google Vertex AI
Native Vertex endpoint with service-account auth — Q4 2026
Replicate
Replicate model runs via their prediction API — 2026
Hugging Face
Serverless Inference API + private endpoints — 2026
Everything the system does
80+ features grouped by concern. Nothing is a mock — every bullet ships to production the moment its area does, and each one is measurable via the dashboard or the API.
Governance & Runtime Enforcement
Decisions made BEFORE the provider call, at gateway speed.
- Runtime guardrails evaluated inline at the gateway (before provider dispatch)
- Policy-as-code with allow/deny lists for models, providers and tools
- Per-key rate limiting (requests-per-minute + requests-per-day)
- Per-key daily token limits (input + output)
- Per-workspace budget hard-stops (402 on limit — CFO promise)
- Per-project model allowlist (enforcement + UI catalog)
- Per-project provider allowlist
- Global kill switch (emergency stop; degrades to observe or block)
- PII detection with regex + Aho-Corasick custom detectors
- Prompt injection detection (Goal 3 pipeline)
- Data-security custom rules per tenant (built-in + user-defined wordlists)
FinOps
Every dollar tracked, attributed and controllable.
- Real-time cost per request (tokens × pricing catalog)
- Monthly spend rollup + 12-month history (tz-aware bucketing)
- Budget alerts (threshold + hard stop, per-workspace)
- AI Cost Copilot — automated recommendations (model switch, waste, cache, routing)
- Cost attribution per user / project / team / model / provider
- Multi-currency support with per-tenant pricing
- Blocked-request savings estimation (“ThinkNEO saved you $X”)
- Wallet / top-up flow with Stripe integration
- Per-tier plan management (Start / Dev / SMB I / SMB II / Enterprise)
- Trial-with-auto-conversion via Stripe subscriptions
Observability
Every request logged, every trace explorable.
- Every request in a hash-chained append-only audit ledger (brain.gateway_audit_events)
- Real-time Request Timeline modal — full 14-event chain per trace
- Trace ID from the moment the request enters the gateway
- Latency breakdown (gateway ms / provider ms / total)
- Status codes + structured error taxonomy
- Filters: user, project, model, provider, status, reason, date range
- Public status page with live health probes + uptime history
- Configurable alert rules → email / SMS / WhatsApp / webhooks
- 30-day retention by default, extensible per tier
- Export to CSV / JSON for audit or BI import
Fallback & Reliability
Provider outage stays invisible to your users.
- Multi-provider fallback chains configured per-tenant per-workspace
- Format-preserving fallback (OpenAI → Anthropic → OpenAI-compat)
- Automatic retry on 5xx / 408 / 429
- Circuit-breaker per provider endpoint (auto-recover)
- Streaming support preserved across fallback
- Health checks per component (readyz / livez / status)
- SLA tracking with breach reports
- Degraded-mode enforcement (fail-neutral for reads, fail-closed for security gates)
- Global env chain as safety-net when tenant has none configured
- Fallback events emitted in the audit ledger (tenant_fallback_chain_{started,succeeded,exhausted})
Data Security
Your keys, your data, your rules — cryptographically.
- BYOK (bring your own keys) end-to-end — ThinkNEO never re-issues provider keys
- Provider keys encrypted at rest with Fernet (canonical secret file, no DSN-derived nonsense)
- Automatic key rotation with new-first / invalidate-old-after pattern
- Optional prompt / response retention per tenant (default OFF)
- No training on customer data — contractual guarantee
- Data residency configuration per region (compliance packs enforce it)
- Custom detectors at Aho-Corasick speed (thousands of patterns, sub-ms)
- Immutable audit trail (append-only trigger blocks UPDATE / DELETE)
- SHA-256 hash chain over every audit event (tamper-evident)
- Tenant isolation at schema + row-level (RBAC + workspace filters everywhere)
Compliance
Regulator-ready evidence, not screenshots.
- LGPD compliance pack (CPF/CNPJ/email/phone redaction + DLP + lineage + monitor)
- GDPR compliance pack (email/phone/SSN redaction + DLP + Art. 30 lineage; EU residency as manual toggle)
- HIPAA compliance pack (SSN/email/phone redaction + DLP + §164.312(b) audit; MRN/NPI/DEA/ICD-10 in roadmap)
- SOC 2 Type II alignment in progress
- Compliance Agent — automated periodic re-evaluation of Projects
- Ledger-anchored decisions (immutable receipts per approval)
- Preset installations per tenant (Layer 1 of the policies epic)
- Compliance scorecards with drill-down and delta history
- Regulatory-ready audit trail (hash chain + append-only)
- Per-project scope evolution (resubmission workflow, F5)
Agents, MCP & A2A
First-class protocol support, not adapters.
- MCP server exposing 62 tools (Spec 2024-11-05) — AAA-rated on Glama.ai
- A2A (Agent-to-Agent) protocol native support — AIRGP + A2ASTC DOIs registered
- Agent lifecycle: create / pause / revoke / archive with immutable audit
- Institutional agents (agent-owned API keys, actor attribution)
- Agent governance evaluation pipeline (per-request)
- Connector governance (per-tool allow/deny)
- 24+ A2A skills exposed via bridge
- Bidirectional bridge MCP ↔ A2A
- Agent-scoped rate limits and budgets
- Endpoint module (downloadable, per-tenant signed)
Multi-Tenancy & Identity
Enterprise organization model out of the box.
- Per-tenant schema-level isolation + row-level filters everywhere
- Multiple workspaces per tenant (Start=1 → SMB II=10)
- RBAC with 4 roles (tenant_admin, admin, manager, user)
- Groups + Departments hierarchy (per-user membership)
- Per-project access control (F4 project-scope launched 2026-06-17)
- Invite-based user provisioning via email verification code
- Google / GitHub SSO
- Email + password with 6-digit MFA (SMS + authenticator app)
- Session management (multi-device, revocable)
- User attribution on every request (user_id, name, email)
Providers & Interoperability
One integration point for the whole model stack.
- 19 providers via one gateway (see the provider list above)
- OpenAI drop-in replacement (base URL swap — no code change)
- Native Anthropic /v1/messages endpoint
- Format auto-conversion between OpenAI ↔ Anthropic ↔ Gemini
- Provider-agnostic routing rules (per-tenant)
- Per-model pricing catalog (input + output token pricing)
- Custom endpoint override (self-hosted URL per workspace)
- Streaming support with chunk-passthrough
- Round-robin + weighted routing across compatible models
- Idempotency keys (same request twice = single billed operation)
Developer Experience
Ship in 5 minutes; audit forever.
- 5-minute quickstart (base URL swap + BYOK, no SDK change)
- REST API versioned under /v1
- OpenAI SDK compatibility (Python / Node.js / any language)
- LangChain compatibility (OpenAI-compat provider)
- Python + Node.js quickstart examples
- Interactive dashboard for policies (no YAML wall)
- Postman collection + Swagger docs
- CLI tool (ThinkNEO connector, per-tenant signed download)
- Docker image on Docker Hub (thinkneoai/bridge)
- Open protocols published (AIRGP + A2ASTC — DOI on Zenodo)
Every feature listed here is live in production.
No demoware, no roadmap-only bullets except where explicitly marked. Start free in Monitor Mode, BYOK — 5-minute setup, no credit card.