Content libraryReference

Tool Landscape Guide

Curated guide to gateways, observability platforms, and FinOps tooling for LLMs.

MarkdownReference

TokenOps Tool Landscape Guide

A curated guide to the tooling ecosystem for managing LLM token costs, observability, and optimization.


1. LLM Gateways / Proxies

Central routing layer for all LLM API calls — the foundation of TokenOps.

Tool Type Pricing Best For
LiteLLM Open-source proxy Free (self-hosted) Teams wanting full control with multi-provider routing
Portkey Managed gateway Free tier + paid Production apps needing reliability, caching, and observability
Helicone Managed proxy Free tier + paid Developer-friendly logging and analytics
Kong / Envoy (+ plugins) API gateway + custom Free (self-hosted) Large orgs with existing API gateway infrastructure
Custom middleware Build your own Engineering time Teams with unique routing/tagging requirements

LiteLLM

Key Features: 100+ LLM providers supported, unified OpenAI-compatible API, load balancing, fallbacks, cost tracking, budget limits, Prometheus metrics
Integration: Drop-in proxy — point your OpenAI SDK to http://litellm-proxy:4000
When to use: Best starting point for most teams. Free, flexible, actively maintained.

Portkey

Key Features: AI gateway with automatic retries, caching, load balancing, guardrails, multi-provider routing, analytics dashboard
Integration: SDK wrappers for Python/JS, or as a proxy
When to use: Teams wanting a managed solution with built-in reliability features.

Helicone

Key Features: One-line integration, request logging, cost tracking, user analytics, prompt management, caching
Integration: Header-based proxy — add one header to existing API calls
When to use: Fastest setup; ideal for getting visibility quickly.


2. Observability Platforms

Monitoring, tracing, and evaluating LLM applications.

Tool Type Pricing Best For
Langfuse Open-source Free (self-hosted) / Cloud plans Teams wanting open-source observability with tracing
LangSmith Managed Free tier + paid LangChain users needing integrated monitoring
Arize Phoenix Open-source Free (self-hosted) ML teams wanting detailed trace analysis
Braintrust Managed Free tier + paid Evaluation-first teams doing prompt engineering
Weights & Biases Managed Free tier + paid ML teams already using W&B for experiment tracking

Langfuse

Key Features: Traces, scores, prompt management, dataset management, evaluations, cost tracking, usage analytics
Integration: Python/JS SDKs, OpenAI wrapper, LangChain/LlamaIndex integrations
When to use: Best open-source option. Self-host for data control or use cloud.

LangSmith

Key Features: Tracing, evaluation, prompt hub, monitoring, dataset management
Integration: Native LangChain integration, standalone SDK
When to use: If you're already in the LangChain ecosystem.


3. Cost Management

Dedicated tools for tracking, allocating, and optimizing AI spend.

Tool Type Pricing Best For
Custom dashboards (Grafana) Open-source Free Teams with existing Grafana/Prometheus stack
Finout Managed FinOps Paid Enterprise FinOps teams managing cloud + AI costs
CloudZero Managed Paid Connecting AI costs to business unit economics
Provider dashboards Built-in Free Quick visibility with no setup

Custom Grafana Dashboards

Key Features: Full control, connect to any data source, custom panels, alerts
Integration: Prometheus exporter from gateway, direct SQL from cost warehouse
When to use: Most flexible option. Recommended if you have Grafana expertise.

Provider Dashboards (OpenAI, Anthropic, Google)

Key Features: Usage, cost trends, API key breakdowns
Limitations: Single provider view, limited tagging, no cross-provider comparison
When to use: Quick start, but supplement with gateway-level tracking.


4. Prompt Management

Version control, testing, and deployment for prompts.

Tool Type Pricing Best For
Langfuse Prompts Open-source Free Teams already using Langfuse for observability
Humanloop Managed Free tier + paid Non-technical stakeholders collaborating on prompts
Promptfoo Open-source Free CLI-first prompt evaluation and testing
Git + YAML DIY Free Engineering teams wanting prompts-as-code

Promptfoo

Key Features: CLI-based prompt evaluation, comparison tables, multiple models, custom metrics, CI/CD integration
Integration: npm/CLI, YAML config, GitHub Actions
When to use: Best for systematic prompt testing in CI/CD pipelines.

Git + YAML (Prompts as Code)

Key Features: Full version history, PR reviews, branch testing, CI integration
Integration: Any CI/CD system
When to use: Engineering-led teams who want prompts treated as production code.


5. Caching Solutions

Reduce redundant LLM calls with response caching.

Tool Type Pricing Best For
Provider prompt caching Built-in Discounted tokens Repeated long system prompts or context
Redis + custom layer Open-source Free Exact-match caching with full control
GPTCache Open-source Free Semantic caching with vector similarity
Portkey cache Managed Included Easy semantic caching without infrastructure

Provider Prompt Caching (OpenAI, Anthropic, Google)

How it works: Provider caches the prefix of your prompt. Subsequent calls with the same prefix get discounted input tokens (up to 90% off).
When to use: Always enable for workloads with consistent system prompts.

Redis + Custom Logic

How it works: Hash the request (model + messages + params) → check Redis → serve cached response or call API
When to use: High-volume deterministic tasks (classification, extraction).


6. Vector Databases (for RAG)

Store and retrieve embeddings for retrieval-augmented generation.

Tool Type Pricing Best For
Pinecone Managed Free tier + paid Serverless, zero-ops vector search
Weaviate Open-source / Cloud Free (self-hosted) Teams wanting hybrid search (dense + sparse)
Qdrant Open-source / Cloud Free (self-hosted) High-performance, Rust-based, quantization support
pgvector PostgreSQL extension Free Teams already on PostgreSQL wanting to avoid new infra
Chroma Open-source Free Prototyping and small-scale applications

pgvector

Key Features: PostgreSQL extension, HNSW/IVFFlat indexes, no new infrastructure
When to use: If you're already on PostgreSQL and vector data fits in memory. Best for < 10M vectors.

Pinecone

Key Features: Serverless, zero-ops, metadata filtering, namespaces, hybrid search
When to use: Teams wanting managed infrastructure with scale-to-zero pricing.


7. Evaluation Frameworks

Automated quality testing for LLM outputs.

Tool Type Pricing Best For
RAGAS Open-source Free RAG-specific evaluation (faithfulness, relevance, recall)
DeepEval Open-source Free General LLM evaluation with built-in metrics
Promptfoo Open-source Free Comparative prompt evaluation across models
Custom (pytest + LLM-as-judge) DIY Free Teams with specific quality criteria

RAGAS

Key Features: Faithfulness, answer relevance, context precision, context recall metrics
When to use: Primary evaluation tool for RAG pipelines.

DeepEval

Key Features: 14+ built-in metrics, LLM-as-judge, Pytest integration, CI/CD ready
When to use: Comprehensive evaluation beyond RAG.


Tool Selection Matrix

By Team Size

Team Size Gateway Observability Caching Evaluation
1–5 engineers Helicone (easy setup) Langfuse Cloud Provider caching Promptfoo
5–20 engineers LiteLLM (self-hosted) Langfuse (self-hosted) Redis + custom RAGAS + Promptfoo
20+ engineers LiteLLM + custom plugins Langfuse + Grafana Redis + semantic Full evaluation pipeline

By Budget

Budget Recommended Stack
$0 (all free) LiteLLM + Langfuse + Grafana + pgvector + Promptfoo
< $500/month Helicone + Langfuse Cloud + Pinecone free + Promptfoo
$500–$5K/month Portkey + Langfuse Cloud + Pinecone + DeepEval
Enterprise Custom gateway + full observability + dedicated vector DB

By Maturity Level

Maturity Focus Area Key Tools
Level 1 (Ad Hoc) Get visibility Helicone or provider dashboards
Level 2 (Aware) Add tagging LiteLLM + Grafana
Level 3 (Instrumented) Full observability LiteLLM + Langfuse + custom dashboards
Level 4 (Optimized) Caching + routing + Redis cache + Promptfoo + evaluation pipeline
Level 5 (Governed) Automation + Custom policy engine + automated guardrails

Reference from the TokenOps Atlas — tokenops-atlas