The right plan for every team.
Start free. Add security and scale as you grow. Every plan includes full LLM observability.
Free
Free forever
Start freeUnlimited traces
1,000 evaluations / month
14-day trace retention
1 seat
- fluiq.instrument(): full observability
- Trace explorer & live dashboard
- Streaming traces
- Community support
- fluiq.optimize(): response caching
- CI/CD eval gates
- Slack anomaly alerts
- fluiq.secure(): security scanning🔒 Growth+
Team
billed monthly
Start 5-day free trialUnlimited traces
10,000 evaluations / month
Unlimited trace retention
Up to 5 seats
- Everything in Free
- fluiq.optimize(): response caching
- CI/CD eval gates
- Slack anomaly alerts
- Email support (48h SLA)
- fluiq.secure(): security scanning🔒 Growth+
Growth
billed monthly
Start 5-day free trialUnlimited traces
100,000 evaluations / month
Unlimited trace retention
Up to 20 seats
- Everything in Team
- fluiq.secure(): security scanning
- Prompt injection, PII, jailbreak & secret-leak protection
- Custom eval thresholds
- SSO (single sign-on)
- Priority support (24h SLA)
Enterprise
Unlimited traces
Unlimited evaluations
Unlimited trace retention
Unlimited seats
- Everything in Growth
- VPC / on-prem deployment
- Custom SLA & dedicated support
- Audit logs & compliance exports
- SAML / SCIM provisioning
- Dedicated onboarding
Compare plans
Every feature, side by side.
| Features | Free | Team | Growth | Enterprise |
|---|---|---|---|---|
| Observability | ||||
| Traces / month | Unlimited | Unlimited | Unlimited | Unlimited |
| Trace retention | 14 days | Unlimited | Unlimited | Unlimited |
| Live dashboard | ||||
| Trace explorer | ||||
| Multi-agent DAG rendering (LangGraph, CrewAI, ADK) | ||||
| Streaming traces | ||||
| Evaluation | ||||
| Evals / month | 1,000 | 10,000 | 100,000 | Unlimited |
| LLM-as-judge metrics | ||||
| Agentic evaluation (tool selection, trajectory, coordination) | ||||
| Multi-model judge jury with audit trail | ||||
| CI/CD eval gates | ||||
| Custom eval thresholds | ||||
| Datasets | ||||
| Golden datasets from traces | ||||
| Whole-trajectory capture (steps, tools, MCP, media) | ||||
| Connect Agents auto-sync | ||||
| Batch agentic-eval & security runs | ||||
| Security: fluiq.secure() | ||||
| Prompt injection blocking | ||||
| PII detection & redaction | ||||
| Jailbreak detection | ||||
| Secret leak prevention | ||||
| Indirect injection detection | ||||
| RAG poisoning detection | ||||
| Tool-input exfiltration & allowlist | ||||
| Cross-agent injection & trust-boundary escalation | ||||
| Optimization: fluiq.optimize() | ||||
| Response caching | ||||
| Cache hit dashboard | ||||
| Team & Access | ||||
| Seats | 1 | 5 | 20 | Unlimited |
| SSO | ||||
| SAML / SCIM | ||||
| Audit logs | ||||
| Support | ||||
| Community support | ||||
| Email support | 48h SLA | 24h SLA | Dedicated | |
| Slack alerts | ||||
| Dedicated onboarding | ||||
| Deployment | ||||
| Cloud (managed) | ||||
| VPC / on-prem | ||||
What you unlock
Two calls. Security and speed, handled.
fluiq.optimize() ships with Team, and fluiq.secure() unlocks on Growth.
fluiq.secure()
Included in Growth & EnterpriseOne call. Full pipeline protection.
Wrap your pipeline with server-side security scanning before any data is stored. Fluiq checks every prompt and response, so attack patterns are never shipped in the public SDK.
- PII Detection & Redaction: Names, emails, phone numbers, SSNs, and credit cards, detected and redacted before persistence.
- Prompt Injection Blocking: Catches injection patterns, jailbreak attempts, and skeleton key attacks in real time.
- Jailbreak & Semantic Attack Scoring: Semantic similarity scoring against known attack vectors, even when phrasing varies.
- Secret Leak Prevention: Scans LLM outputs for leaked API keys, tokens, and high-entropy credential strings.
- Indirect Injection & RAG Poisoning: Inspects tool outputs and retrieved documents for second-order injection, and flags chunks that semantically resemble an attack.
- Tool Abuse Defense: Catches sensitive data exfiltrated through tool-call arguments and tools invoked outside your configured allowlist.
- Multi-Agent Trust: Detects cross-agent injection and risk escalating across agent handoffs in the trace DAG.
- Warn or Block mode: warn (default) flags risks and attaches security metadata to the trace. block intercepts before the LLM call and raises FluiqSecurityError.
fluiq.instrument(api_key="fl_...")
fluiq.secure() # warn mode flags risks on the trace
fluiq.secure(mode="block") # block modefluiq.optimize()
Included in Team & aboveServe repeated prompts from cache.
Fluiq analyses your historical traces to find which LLM calls repeat most often and provisions a dedicated Redis cache for your account. Repeated prompts are served instantly, saving both latency and cost.
- Trace-Driven Cache Profiling: The backend mines your trace history to build a cache profile, no manual configuration needed.
- Automatic Cache Population: Real LLM responses are stored automatically on the first call; subsequent matches are served from Redis.
- Cache mode: Full interception: matching prompts never reach the LLM API.
- Observe mode: Records what would have been a cache hit without intercepting: review your savings before opting in.
- Zero code changes: One fluiq.optimize() call after instrument(). The SDK handles connection, profiling, and cache lookup.
- Cache hit dashboard: See hit rates, latency savings, and estimated cost savings in your Fluiq dashboard.
fluiq.instrument(api_key="fl_...")
fluiq.optimize() # cache mode
fluiq.optimize(mode="observe") # observe modeFAQ
Frequently asked questions
What counts as a trace?
One traced span, typically one LLM call, one retriever call, or one decorated function invocation. A single end-to-end agent run usually emits 5-20 traces depending on how many tools and LLM calls it makes. Trace ingestion is unlimited and free on every plan — the Free plan keeps a rolling 14-day history, while Team and above retain your traces indefinitely.
Which frameworks does Fluiq support?
Fluiq instruments at the function-call level and ships integrations for OpenAI, Anthropic, Gemini, LangChain, LangGraph, CrewAI, Google ADK, and raw HTTP calls via the @trace decorator. Streaming, tool calls, thinking tokens, and MCP servers are all captured automatically.
What counts as an evaluation?
One LLM-as-judge scoring call: e.g. a hallucination check on an answer or a relevance score over a retrieved chunk set. Metrics include hallucination, faithfulness, relevance, toxicity, coherence, and completeness. Free includes 1,000 evals/month, Team 10,000, Growth 100,000, and Enterprise is unlimited.
When do I need fluiq.secure()?
fluiq.secure() runs server-side security scanning: PII detection and redaction, prompt-injection and jailbreak blocking, secret-leak prevention, indirect-injection and RAG-poisoning detection, tool-input exfiltration and allowlist enforcement, and multi-agent trust checks (cross-agent injection and trust-boundary escalation). It is included on the Growth and Enterprise plans. In warn mode it flags risks on the trace without blocking; in block mode it raises FluiqSecurityError before a HIGH-risk prompt reaches the LLM.
How does fluiq.optimize() work?
Available on Team and above. After you call fluiq.optimize(), the SDK fetches your trace-derived cache profile from the Fluiq backend, connects to a dedicated Redis instance provisioned for your account, and begins serving repeated prompts from cache. In observe mode it records what would have been a cache hit so you can review projected savings before enabling full interception.
Do you support self-hosting?
Yes. VPC and on-prem deployments are available on the Enterprise plan. The SDK is a thin instrumentation layer and can be pointed at your own backend endpoint if you prefer full self-hosting.
Can I switch frameworks later?
Yes. Because Fluiq instruments at the call level, the same SDK works across all supported frameworks simultaneously. Switching from LangChain to LangGraph, or adding a new provider, requires no changes to your instrumentation.
How do I know if my LLM is actually working in production?
Fluiq traces every LLM call, tool call, and decorated function automatically, so you watch live token usage, latency (p50/p95/p99), USD cost per agent node, and pass/fail status stream onto the dashboard as runs complete. Automated LLM-as-judge evals score hallucination, faithfulness, relevance, and more on real production traffic, and Slack alerts fire the moment quality regresses or failure rates climb — so "is it working?" becomes a number you watch, not a guess.
What's the cheapest way to monitor my LLM application?
Start on Fluiq's Free plan: unlimited traces and 1,000 evaluations per month at no cost, with full tracing, cost attribution, and latency analytics included. Instrumentation is one line — fluiq.instrument() — so there's no agent to run and no infrastructure to host. Free keeps a rolling 14-day history; when you outgrow it, Team and above retain traces indefinitely, and fluiq.optimize() caches repeated prompts to cut model spend, so monitoring can actually lower your bill instead of adding to it.
Will adding observability slow down my LLM?
No. The Fluiq SDK is a thin instrumentation layer that records spans and ships them to the backend in the background, so it adds negligible overhead to the call itself. It is also fail-open by design: if the Fluiq backend is ever slow or unreachable, your application keeps running and never blocks waiting on a trace. You get full visibility without paying for it in latency.
How do I know if my AI is actually secure?
fluiq.secure() scans both sides of every call. Before the model runs, pre-call scanning catches prompt injection, jailbreaks, and skeleton-key attacks; after it runs, post-call scanning redacts PII and secrets and inspects the whole trace tree for agentic threats — RAG poisoning, tool-input exfiltration, tools used outside their allowlist, and multi-agent trust attacks. Every risk is flagged on the trace with a severity and category, so security is something you see per request instead of assume.
What happens when an AI system gets hacked?
Common attacks — a jailbreak prompt, a poisoned retrieved document, a tool tricked into leaking data — try to make your model ignore its instructions or exfiltrate sensitive information. With fluiq.secure() in block mode, Fluiq raises a FluiqSecurityError before a high-risk prompt ever reaches the model; in warn mode it records the attempt on the trace without interrupting traffic. Because scanning fails open, a scanner error degrades to observe-only instead of taking your app down, and every blocked or flagged event can alert your team in Slack.
Can someone trick my LLM into giving away secrets?
That is exactly the class of attack Fluiq is built to stop. The response gate scans model output for PII and high-entropy secrets and redacts them before they are stored or returned, while tool-input exfiltration and allowlist checks catch sensitive data being smuggled out through tool calls. Pre-call injection detection blocks prompts engineered to override your system instructions, and post-call scanning reads retrieved docs and sibling spans to catch indirect injection planted in your RAG sources.
Ship safer AI, faster.
Start free, then add security and scale when you need it.
No credit card required on Free and Team plans.