Catalogue-backed calls
Every provider and model pair is checked against the same live catalogue the console uses — no shadow lists or one-off client configs.
Govern LLMs · OpenAI-compatible · Native API & SDK
LLM governance to cut cost, optimize usage, and secure every model call at scale
TokenSaver is the control plane for all your LLM traffic: enforce policies, access, and budgets; shrink token bills with semantic cache, RAG, and compression; and keep data safe with PII controls — in one predictable pipeline with full observability. Use the web console, point OpenAI-compatible clients at TokenSaver, or integrate with the native HTTP API and Python SDK— same pipeline, metering, and policies on every path.
Early Adopter Program — free plan to explore governed LLM at scale · One proxy · RAG, cache, and anonymization modules · Policies & spend controls · Works with OpenAI, Anthropic, Google, and more · Open-source Python SDK on PyPI (free for the community)
Early Adopter Program: request free access to discover TokenSaver — your feedback helps steer the product roadmap.
Unified control plane
Govern every LLM through one platform layer
TokenSaver sits in front of your vendor APIs: the same policies, catalogue, metering, and pipeline run whether traffic comes from the console, an OpenAI-compatible client, or the native API — so models stay under control as you add providers and scale usage.
- 6LLM providers routed
- 175+LLM models governed (chat & embeddings)
- 1Pipeline — cache, RAG, compression, PII, then model
- 3Ways in — console, OpenAI-compatible HTTP, native API & SDK
Org-scoped governance
Workspace keys, quotas, and spend views anchor on TokenSaver — vendor credentials stay on the platform while teams keep a single place to enforce access and budgets.
Same modules, any vendor
Semantic cache, retrieval, compression, and PII controls run in a fixed order before the upstream model — so optimization and safety don't depend on which provider answers the call.
Observable by design
Traces and metrics across pipeline steps give platform and security teams visibility into cost, latency, and policy outcomes — not just raw model output.
Explore the public model catalogue or call GET /api/v1/llm-reference/models for the authoritative list for your environment.
How you connect
One governed pipeline — three ways to integrate
Whether your teams live in chat UIs, IDEs, or custom backends, you land on the same TokenSaver path: cache, RAG, compression, PII, then the model — with org-scoped keys, quotas, and dashboards. Pick the surface that matches your stack; no duplicate policy layers.
- Same module order & observability
- One TokenSaver API key per workspace
- Provider keys stay on the platform
The platform reference catalog spans 6 providers (OpenAI, Anthropic, Google Gemini, Mistral, Grok via Groq, DeepSeek) and 175+ LLM models (chat and embeddings). Same model / provider pairs the native and OpenAI-compatible APIs validate against GET /api/v1/llm-reference/models.
OpenAI-compatible
https://api.tokensaver.fr/openai/v1Chat Completions, Responses API, models (all routed providers), embeddingsNative TokenSaver API
https://api.tokensaver.fr/api/v1Pipelines, RAG, chats — Python SDK & custom appsNo-code · Chat tools
OpenAI-compatible UIs
Connect LibreChat, Open WebUI, and other tools that accept a custom OpenAI base URL. Set base URL to …/openai/v1, use your TokenSaver API key as Bearer, and optional X-Tokensaver-* headers to tune cache, RAG, compression, and PII per request, or X-Tokensaver-Apply-Key-Pipeline-Defaults to reuse pipeline defaults saved on the API key — no fork of your client required.
Console · Teams
TokenSaver Platform
The hosted app for operators and builders: governed chat, dashboard and spend views, model pricing, API keys, pipeline defaults, and in-product docs. Ideal when you want clicks, not config files — still the same pipeline the APIs run.
Open the consoleDevelopers · Code
Native API & Python SDK
POST /api/v1/pipelines/run, typed tokensaver-sdk on PyPI, and the same reference your workspace shows after sign-in. Pick any catalog-backed provider/model pair — 175+ LLM models across 6 vendors in the reference catalog. Best when you need fine-grained requests, agents, or CI — governance and metering match the console and OpenAI-compatible traffic.
Provider-agnostic — use the models you already trust
Same governance, optimization, and security layer in front of OpenAI, Anthropic, Google Gemini, Mistral, Grok (via Groq), DeepSeek, and more — route and compare 175+ catalog models without duplicating policies or losing cost visibility.
Built for the move from pilot to production: one control plane as usage, providers, and internal stakeholders grow.
OpenAI
Anthropic
Google Gemini
Mistral AI
Grok
DeepSeek
Platform
What's powering TokenSaver
Four capabilities that work together — not a pile of disconnected APIs. Configure once, override per request when you need to.
Product teams get business value fast; platform and security teams keep governance, spend control, and reliability as adoption scales.
Pipeline console
A single workspace to tune cache thresholds, RAG depth, compression level, and PII policies — aligned with what the API actually runs.
Learn more
AI observability (OpenTelemetry)
Instrument LLM traffic with OpenTelemetry-friendly traces and metrics across every pipeline step. Expose the same signals to Prometheus and Grafana for alerting, SLOs, and cost or latency dashboards your platform team already runs.
Learn more
RAG & knowledge
Ingest PDFs and retrieve grounded chunks with multimodal-ready storage — context that stays scoped to the right user and workspace.
Learn more
Trust by design
Data sensitivity by design: automatic PII detection and configurable anonymization or redaction on every exchange, so regulated or personal content is stripped before it reaches model APIs. Harden data flows against accidental leakage to LLMs and downstream logs while you scale usage with confidence.
Learn more
Intelligent routing
One governed path from prompt to model response
Every call flows through the same modules — so you optimize tokens, enforce guardrails, and keep audits simple. Swap providers without losing cost signals or safety posture.
Get free access
Lifecycle tools
Tools to manage the full request lifecycle
Test overrides, inspect per-step metrics, and correlate runs with traces — low-friction operations for platform teams and app owners alike.
See use casesArchitecture
Trusted pipeline architecture
Embed anywhere your apps run — console, OpenAI-compatible clients, or native API / SDK — supervise usage centrally, enforce org boundaries, and keep provider keys off end-user devices.
Custom apps and agents use the native API; LibreChat, Open WebUI, and similar tools use the OpenAI-compatible base path — one product behind both.
Developers · Open source
Python SDK on PyPI — LangChain & LangGraph, extended with TokenSaver governance
tokensaver-sdk is open source (MIT) and free for the open-source community, published on PyPI. TokenSaver builds on LangChain and LangGraph to run governed LLM pipelines and adds a governance overlay on top — access control, budgets, semantic cache, RAG, compression, PII policies, and observability. The SDK is the HTTP client to that stack: same behaviour and limits as the console, whether you trigger runs from code or the UI.
Orchestration inside the platform leans on LangChain / LangGraph patterns; TokenSaver enforces policy and cost before traffic reaches commercial model APIs.
Use cases
TokenSaver powers critical LLM workflows
Same pipeline, different jobs — customer-facing or internal.
Customer support
Grounded answers with RAG, redacted transcripts with PII policies, and cache for repeat intents.
Benefit: Lower cost per ticket, faster replies.
Internal copilots
Central keys, workspace isolation, and spend visibility across teams — without shadow IT keys.
Benefit: Governance without blocking builders.
Research & docs
Long-context compression before the model call, retrieval over your knowledge base, full run history.
Benefit: Deeper context, controlled spend.
Analytics & data apps
Trace IDs, token and cost breakdowns, export-friendly metrics for FinOps and security reviews.
Benefit: One source of truth for LLM spend.
Trust
Trusted operations, built in
Guardrails and visibility that stay aligned as you scale traffic.
Pipeline guardrails
Fixed module order, plan-aware feature flags, and configurable PII strategies — reduce drift between environments.
Data boundaries
Org-scoped provider keys, user-scoped cache and retrieval, and masking before payloads leave your policy boundary.
Observability
Per-run steps, cache hit types, RAG attribution hooks, and OTel — so incidents are short and audits are boring.
Getting started
From API key to governed rollout
1
FoundationConnect providers once, control access centrally
Register OpenAI, Anthropic, Google, or Mistral keys at org level. Applications call TokenSaver only, so provider credentials stay off app servers and client devices.
Outcome: one secure gateway for all teams and environments.
2
OptimizationSet smart defaults to reduce spend and drift
Define governed defaults for semantic cache, RAG, compression, and PII controls. Keep safe per-request overrides for advanced cases, without losing consistency.
Outcome: lower token cost, predictable behavior, faster delivery.
3
ScaleMonitor cost, quality, and security in real time
Track usage, spend, cache hit types, and step-by-step traces in one place. Detect anomalies quickly and provide audit-ready visibility for platform, security, and finance teams.
Outcome: confident rollout across products, teams, and providers.
Explore more
Integrate — OpenAI-compatible, console, or native APISee the three pathsDocs — OpenAI SDK, cURL, LangChain & pipeline headersCode examplesPython SDK — LangChain, LangGraph & governanceInstall from PyPIAPI SDK — full referenceRead the docsHow the pipeline worksLearn moreWhy TokenSaverLearn moreUse cases by teamLearn morePricing & plansLearn more
Frequently asked questions
What is an LLM gateway?
A single API and policy layer in front of multiple model providers. Your applications call TokenSaver; TokenSaver applies cache, retrieval, compression, and safety, then calls the right model with your org's keys.
Can I use LibreChat, Open WebUI, or other OpenAI-compatible clients?
Yes. Configure the client's OpenAI base URL to point at TokenSaver's OpenAI-compatible path (…/openai/v1), and authenticate with your TokenSaver API key. The same governed pipeline runs as for the console and the native API; optional X-Tokensaver-* headers adjust cache, RAG, compression, and PII per request. Code examples: OpenAI-compatible docs.
Can I self-host?
The platform is designed for both cloud SaaS and self-managed deployments — keep data and keys inside your perimeter when required.
How does billing visibility work?
Runs record tokens, estimated cost, and cache outcomes so finance and engineering can reconcile usage without exporting raw logs.
Is the pipeline order fixed?
Yes — cache, RAG, compression, PII, then LLM, then metrics. That predictability is what makes debugging and compliance reviews tractable.
Is there a Python SDK? Does it work with LangChain or LangGraph?
Yes. tokensaver-sdk is open source (MIT), free to install from PyPI. It calls the TokenSaver HTTP API: the product uses LangChain and LangGraph as part of its pipeline implementation and layers governance (cache, RAG, compression, PII, quotas, observability) on those foundations — you integrate from Python without re-implementing that stack yourself.