Welcome to Cognocient AI Docs
Popular: Attribution Headers, Budget Enforcement, Python Wrapper, CSV Import
Cognocient AI is an AI spend intelligence platform that proxies every API call your application makes to AI providers — OpenAI, Anthropic, Google Gemini, Mistral, Groq, Together AI, Azure OpenAI, AWS Bedrock, Google Vertex AI, and your own OpenAI-compatible servers. Every request is observed, attributed to a feature or team, analysed for waste, and surfaced in real-time dashboards your CFO can actually use.
The integration is a single configuration change. No new SDKs, no new dependencies, no application code to rewrite.
Three ways to get started
The proxy above is the flagship path — live attribution AND pre-call budget enforcement. If you'd rather evaluate without a proxy hop or a production code change, see the Python async wrapper (live attribution, no pre-call enforcement) or the CSV/OTel importer (historical data, no code change at all).
Explore the docs by category
Getting started
The problem
58% of organizations describe their AI costs as a "black box" (Capgemini Research Institute, 2025). Engineering teams see token counts in Datadog. Finance teams get a consolidated line item on the credit card bill — weeks later. Nobody can answer:
- Which product feature is driving the cost spike?
- Is this spend waste, or is it generating revenue?
- Are we about to breach our AI budget before the month ends?
Cognocient AI closes all three gaps — automatically, in real time, without requiring your developers to instrument anything manually.
How Cognocient AI works
Your application code points at https://api.cognocient.com/v1 instead of the AI provider directly. Cognocient AI:
- Forwards the request to the original provider — same API surface, same models, same response format. Your application code sees zero difference.
- Records metadata — model, token counts, cost, latency, and your optional attribution headers. Prompt and response text are not logged (opt-in response caching is the one exception).
- Runs analytics in-path — waste detection, budget enforcement, anomaly scoring, and cache hit analysis — before the response returns to your app.
- Surfaces insights — into the Cognocient AI dashboard, your BI tools via the FOCUS-aligned export, or your alerting pipeline via webhooks.
The proxy adds about 2–8ms at the median at low concurrency in our published benchmark, which also shows how that grows under heavy load on one instance. Budget checks run in Redis (sub-millisecond) before every call. Cache hits (Growth plan) skip the provider entirely.
One change unlocks everything
Change one line. That's it.
2-minute setup
No SDKs to install. No infrastructure to manage. If you already use the OpenAI SDK — or any SDK that accepts a base_url / baseURL — you are already compatible. The same URL works for every supported provider; the model name picks which one serves the call.
Key capabilities
Cognocient AI is built around nine capability pillars. Each unlocks a different part of the AI spend visibility problem.
Tag every AI call by feature, team, session, and GL account using HTTP headers. Slice and dice costs any way finance needs.
Tag calls with X-Cost-Session: JIRA-1234 or pr-456 to see AI cost per story, PR, or agent task run — down to the cent.
Retry waste, model mismatch, context bloat and context starvation detected automatically by feature, with token maxing and context tax analysis on Growth.
Live dashboard panel showing investment vs. waste split, the productive share of spend, and a one-sentence board summary generated from your actual spend data.
Hard budget limits enforced at the proxy — before the charge reaches your provider bill. Hierarchical: run → feature → department → org. The tightest limit wins.
Budget reservations are enforced physically, not just logically. Forwarded max_tokens is clamped to what the remaining budget can actually afford.
Semantic caching via pgvector HNSW, prompt-cache routing, and AI-generated one-click recommendations that right-size your spend.
Full cost tree across multi-agent workflows: which agent called which tool, how much each step cost, and which team was responsible.
Board-ready PDF executive reports with 3 narrative tones. FOCUS-aligned CSV export. Scheduled monthly delivery to finance teams.
One OpenAI-compatible endpoint for every provider, with the Auto Router, load-balanced model groups, provider failover and content guardrails for PII and prompt injection.
A one-click kill switch for proxied spend, dollar-denominated velocity limits for image/video generation floods, and daily reconciliation against what OpenAI, Anthropic, Google Cloud, AWS and Azure actually billed.
Run Claude Code through the proxy with ANTHROPIC_BASE_URL for per-developer cost and budgets, and see Cursor and GitHub Copilot spend alongside it.
Tag calls with X-Cost-Customer and see each of your customers' AI cost next to what they pay you, from Stripe or your own figures.
Finding features in the dashboard
The sidebar groups pages into Observe, Monitor, Control, Intelligence and Executive. The most-used ones:
| To... | Open | Plan |
|---|---|---|
| See every call and check your tags | Monitor → Live Calls | every plan |
| See today's cost, errors and retry waste | Observe → Overview | every plan |
| Set or check a limit | Control → Budgets | every plan |
| Stop all spend now | Control → Safety & Limits (Emergency Freeze) | every plan |
| Check an agent run's cost | Monitor → Run Budgets | every plan |
| Find recoverable waste | Monitor → Waste | Base |
| Charge spend back to teams | Control → Chargeback | Base |
| Apply one-click savings | Intelligence → Recommendations | Growth |
| Brief the CFO | Executive → Executive View | Growth |
The status of every budget is also available to orchestration code at GET /api/budgets/status.
Read the full Dashboard Walkthrough →
Who Cognocient AI is for
Engineering teams use Cognocient AI to understand which features drive AI costs, catch runaway usage before it hits the bill, right-size models with one-click recommendations, and debug anomalies with per-request attribution trails.
FinOps and finance teams use Cognocient AI to allocate AI spend to the right cost centres, enforce monthly budgets in real time, produce chargeback reports for internal GL accounts, and export FOCUS-aligned data for the cloud cost management platforms they already use.
CTOs and CFOs use Cognocient AI's Executive View to answer the board question every quarter: "We spent $X on AI last month — what did we actually get for it?" Unit economics (cost per ticket resolved, cost per report generated, cost per sale influenced) make the answer concrete.
Free forever on one provider — no credit card
Real-time proxy and attribution on day one. Your first attribution dashboard in under 5 minutes. Find my waste — start free
Supported AI providers
Cognocient AI proxies all major providers through a single base_url configuration. No separate keys or SDK changes per provider.
OpenAI
GPT, o-series, embeddings, gpt-image-1
Chat, Responses, embeddings, images, audio, batches
Anthropic
Claude Opus, Sonnet, Haiku
OpenAI SDK (translated) or native SDK (passthrough)
Google Gemini
Gemini Pro, Flash, Imagen
OpenAI-compatible endpoint; Imagen translated
Azure OpenAI
Your deployments
Deployment name mapped to the real model for cost
AWS Bedrock
Claude, Nova, Llama, Mistral
SigV4 + Converse handled; streaming and tools
Google Vertex AI
vertex/gemini-*
Service-account auth, billed to your GCP project
Mistral
Large, Codestral, Magistral
All Mistral model families
Groq
gpt-oss, Llama, Qwen, Whisper
Fast inference
Together AI
Any org/model name
Llama, DeepSeek, Qwen and more
Self-hosted
custom/<model>
Ollama, vLLM, LM Studio, llama.cpp
See Supported Providers for SDK-specific configuration examples for each provider.
Security and privacy
Cognocient AI is built on a metadata-only logging principle:
- Prompt and response text is not logged. We log model, token counts, cost, latency, and your attribution headers. The one exception is opt-in response caching, which keeps a cached response for up to 24 hours.
- Provider API keys are encrypted at rest using Fernet symmetric encryption. Keys are never logged or exposed in API responses.
- All traffic is encrypted in transit (TLS 1.2+).
- No SOC 2 report yet. See Security & Privacy for the full, current state of what's built.
Popular how-to guides
Step-by-step walkthroughs for the most common tasks. Each is self-contained and takes under 30 minutes. See all 16 how-to guides ↓
All how-to guides
| Guide | Time |
|---|---|
| Tag your first AI call | 5 min — add 2 headers and see per-feature spend immediately |
| Set a monthly spending limit | 5 min — hard budget enforced at the proxy |
| Cut your AI bill with one-click recommendations | 30 min — AI Advisor finds savings, one click applies them |
| Get Slack alerts on spend spikes | 10 min — anomaly and budget alerts in your Slack channel |
| Track cost per agent run | 15 min — per-execution cost breakdown and per-run budget |
| Map spend to GL accounts | 10 min — finance-ready chargeback reports with GL codes |
| Set up hierarchical budgets | 15 min — enforce limits at every level simultaneously |
| Debug runaway agent loops | — find, stop, and prevent looping agents |
| Prepare an AI ROI board report | 30 min — cost per outcome, efficiency trend, board-ready PDF |
| Enable semantic caching | 5 min — eliminate duplicate calls, 25–35% bill reduction |
| Track cost per business outcome | 15 min — cost per ticket, contract, or conversion with one header |
| Set up MCP agent attribution | 15 min — full cost tree for Claude + MCP tool call workflows |
| Apply routing rules to auto-switch models | 10 min — no code changes, savings apply immediately |
| Schedule monthly AI spend reports | 5 min — automated PDF delivery to your finance team |
| Export AI spend to your FinOps platform | 10 min — FOCUS-aligned CSV for Apptio, CloudZero, or Spot |
| Migrate from LiteLLM, Langfuse, or Helicone | 15 min — side-by-side comparison and first-5-minutes checklist |
Frequently asked questions
Ready to start? Find my waste — start free