Open source · Built in Europe

The European AI gateway for your agents.

An open-source edge gateway that runs on your side: your prompts and your keys stay there. A control plane to govern, in our cloud or on your infrastructure.

Get notified at release See pricing

The public gateway repository is coming soon. Your prompts only go to the LLM provider you choose, or nowhere with on-premise models.

Architecture

Two planes. Your data stays on one of them.

The data plane sees your prompts and enforces your rules. The control plane defines those rules without sitting on the path of your data.

Your agents & appsOpenAI- or Anthropic-compatible

HTTPS →

Edge gatewayyour infrastructure · open source

→

LLM provideryour keys · BYOK

↓ signed config · ↑ metadata

Control planeconsole · policies · FinOps

Stays on your side

✓Prompts and responses
✓LLM provider keys
✓Cache and RAG content
✓Detailed run replay
✓Enforcement of quotas, budgets and allowlists

Goes through the control plane

↔Metadata for AI flows and aggregated FinOps
↔Gateway health signals
↔Signed configuration, pulled by the gateway
↔Prompt content only if you opt in

Why this architecture holds up

Enforced locally

Hard limits apply on the gateway itself, with no round trip to the cloud on the critical path.

Resilient

If the console is unreachable, the gateway keeps serving on its last signed configuration.

One pipeline, three modes

SaaS, hybrid or self-host: same family of images, same pipeline. Changing mode does not mean changing product.

No lock-in

A gateway running on your side is 100% autonomous: configured by hand, no quota limit, it keeps serving with your keys without the control plane or the console.

What is open source, what is commercial

The gateway is open source. Enterprise features are offered under a commercial license.

	Open-source gateway	Enterprise license
OpenAI- and Anthropic-compatible gateway, BYOK	✓	✓
Virtual keys and basic budgets	✓	✓
Execution graph and cost per Run	—	✓
Cache, compression and PII modules	✓	✓
Agent egress capture, base MCP gateway	✓	✓
AI agent-assisted development (VS Code extension for Claude Code)	✓	✓
SSO and SCIM	—	✓
Execution graph scoped by team and project	—	✓
SIEM audit export and retention	—	✓
Advanced chargeback and FinOps exports	—	✓
Secret managers (Vault, KMS)	—	✓
Air-gap and multi-region control	—	✓
SLA and dedicated support	—	✓

Compare plans The full architecture

Available today: free developer tools

01New · No API key

tokensaver-egress

Use your existing subscription, for example your Claude Code plan, through TokenSaver. No LLM API key needed, 100% free: a lightweight local proxy that makes every call visible in the console.

$ pip install -U tokensaver-egress

Discover egress GitHub PyPI ↗

02BYOK

tokensaver-cli

Bring your own LLM keys (BYOK) and access every TokenSaver feature from the CLI: governed pipelines, model catalogue, Agent Registry, MCP tools and Trust Gateway, for Claude Code, Cursor and Codex.

$ pip install -U tokensaver-cli

Discover the CLI GitHub PyPI ↗

03Python SDK

tokensaver-sdk

Build your own apps on the TokenSaver control plane: governed pipeline calls, RAG documents, chat sessions and governance policies from Python.

$ pip install tokensaver-sdk

Discover the SDK API reference PyPI ↗

Open source · Coming soon to the VS Code Marketplace

Coding in VS Code or Cursor? No terminal needed.

The free, open-source TokenSaver extension brings the same capture into Claude Code: keep your Claude plan, see cost and AI-written files in a side panel, and explain any file. It plugs natively into a local gateway, with no quota limit. Nothing to pip install.

Get notified at release See the extension

Which one should I use?

	tokensaver-egress	tokensaver-cli	tokensaver-sdk
LLM access	Your existing subscription (e.g. Claude Code Pro / Max), no API key	Your own LLM keys (BYOK): OpenAI, Anthropic, Mistral…	Your provider keys, per organisation or per run
Price	100% free	Free CLI · BYOK on every plan	Free on PyPI · usage per your plan
Best for	Observing what your agents do, with the plan you already pay for	Using every TokenSaver feature from the terminal	Building your own apps on the control plane
Traffic path	Local proxy, then straight to the provider	Through the TokenSaver governed pipeline	Your app, then the TokenSaver API
Clients	Claude Code, Cursor, n8n, any HTTPS client	Claude Code, Cursor, Codex, any OpenAI-compatible client	Your Python apps and services
Setup	tokensaver-egress setup	tokensaver login	TokenSaver(api_key="tsv_…")
Launch	tokensaver-egress claude	tokensaver route claude --launch	ts.ask("…", provider, model)
Undo	tokensaver-egress stop · unproxy	tokensaver unroute claude	remove the client

All three feed the same console: egress with your subscription, the CLI with your own keys, the SDK from your own apps.

New · 100% free · No API key

01 · tokensaver-egress

Your Claude Code plan, now observable.

Keep using the subscription you already pay for, for example your Claude Code plan: no LLM API key needed. tokensaver-egress is a lightweight local HTTPS proxy next to Claude Code, Cursor, n8n or any agent; it forwards traffic to the real provider and sends the audit to TokenSaver, where every run shows up in the console.

GitHub PyPI

100% free · no LLM API key · MIT license · Python ≥ 3.10 · free TokenSaver account.

Install once, then launch Claude Code through it

$ pip install -U tokensaver-egress

$ tokensaver-egress setup

$ tokensaver-egress claude

# open the console: every call shows up within seconds

Built to stay out of the way.

Your agentClaude Code · Cursor · n8n · Any HTTPS agent

HTTPS →

tokensaver-egresslocalhost · MIT

→

Anthropic, OpenAI…TLS end to end

↓ Audits

TokenSaver consoleapp.tokensaver.cloud

Runs next to the agent

On a laptop, in CI or in a pod. The SaaS never terminates your LLM TLS for you.

Traffic still goes to the provider

Anthropic or OpenAI remain the destination. Unset HTTPS_PROXY and tools talk to providers directly again.

Metadata first

Request bodies are off by default; when you opt in, auth headers are masked.

Streaming stays interactive

No “buffer everything then reply” for normal chat.

Fail soft

A brief platform outage does not block your agents.

Your key is the trust boundary

Policies (cache, RAG, compression, PII, routing) come from the settings of your TokenSaver key.

Everyday commands

$ tokensaver-egress claude: Auto-start the proxy, then launch Claude Code
$ tokensaver-egress serve: Run the proxy in the foreground, logs visible
$ tokensaver-egress run -- CMD: Run any command through the proxy
$ tokensaver-egress install-skills: Install the TokenSaver skills for Claude Code
$ tokensaver-egress stop: Stop an auto-started proxy

02 · tokensaver-cli

Your LLM keys. Every TokenSaver feature. One CLI.

Bring your own LLM keys (BYOK) and access every TokenSaver feature from the CLI: governed pipelines (cache, compression, RAG, PII, quotas), the model catalogue, the Agent Registry, MCP tools and the Trust Gateway, for Claude Code, Cursor or Codex. The CLI configures your machine; the platform enforces and traces every call.

GitHub PyPI

BYOK with your OpenAI, Anthropic, Mistral… keys on every plan · standard library only · Python ≥ 3.10.

Get started in 60 seconds

$ pip install -U tokensaver-cli

$ tokensaver login

$ tokensaver route claude --launch

$ tokensaver flows --open

What you get

One command, everything wired

tokensaver route claude --launch sets up Claude settings, MCP tools, the Trust Gateway and the tokensaver-router plugin in one shot.

Governed pipelines

Cache, compression, PII, RAG and quotas follow your API key policies: runtime governance, not hard-coded in the CLI.

Your keys, the full catalogue

BYOK on every plan, Free included: the model catalogue resolves through your own provider keys, with no markup.

Runs in the console

tokensaver flows lists your recent runs; tokensaver flows --open opens them in the browser.

Zero-trust model catalogue

tokensaver approve and tokensaver use manage approved models in the Agent Registry.

Clean undo

tokensaver unroute claude restores your backed-up config files. Zero runtime dependencies.

Agent compatibility

Claude Code	tokensaver route claude --launch	Settings, MCP, plugin
Cursor	tokensaver route cursor	Writes .cursor/mcp.json
OpenAI Codex	tokensaver route codex --launch	Exports OPENAI_BASE_URL + key
Any OpenAI-compatible client	tokensaver route proxy	Prints the env to use
MCP only	tokensaver route mcp	Tools + gateway, no agent launch

Command map

Auth: login · logout · whoami · keys
Models: models · use · profiles · approve · catalog
Observability: flows · loop · status
Route: route <claude|cursor|codex|proxy|mcp> · unroute
Health: doctor [--claude]

03 · tokensaver-sdk

Build your own apps on the control plane.

The Python SDK for the TokenSaver API: governed pipeline calls, RAG documents, chat sessions, governance policies and cost estimates, from your own code. Only HTTP transport and response normalisation run in the package; all pipeline logic stays on the platform.

API reference PyPI

pip install tokensaver-sdk · Python ≥ 3.10 · providers on the hosted API: OpenAI, Anthropic, Google, Mistral, Grok, DeepSeek.

Ask a governed question over your own document

$ pip install tokensaver-sdk

from tokensaver_sdk import TokenSaver ts = TokenSaver(api_key="tsv_...")doc = ts.rag_ensure_document("handbook.pdf")result = ts.ask(    "What are the key points?",    provider="openai", model="gpt-4o",    rag_options={"document_ids": [doc["document_id"]]},)print(result.text, result.metrics)

What the SDK gives you

Governed pipeline calls

ts.ask() returns a RunResult with .text, .metrics and .trace; run_pipeline() returns the raw API JSON.

Governance from code

Create, update or disable cache, RAG, compression and PII policies scoped to your API key.

RAG documents

Upload PDF, DOCX, Markdown, CSV or JSON, wait for ingestion, then query them with rag_options.

Chat sessions

Stateless, local or server-side history, with knowledge attached once to the whole session.

Cost estimates

estimate_cost() prices a call before you make it, without calling the LLM.

Typed and explicit

Typed configs and options, and validation errors that fail fast on an unsupported provider or model.

Install, audit, contribute.

Stars, issues and PRs are welcome on the open-source repositories.

tokensaver-egress tokensaver-cli tokensaver-sdk Open the console

Talk to the team

Included from the Free plan

✓Your own LLM keys (BYOK)
✓Run observability in the console
✓Cache, RAG, compression and PII
✓MCP commands inside Claude Code

Pro SaaS and Pro Hybrid · console with no time limit, on our cloud or with your gateway.

Compare plans