TokenSaver
FR

Platform

One control plane for every AI flow.

TokenSaver instruments every LLM call, tool call and decision without changing your agents’ code: you see the fleet, you optimise the spend, you enforce policy before the model answers.

Book a demoSee pricing

Architecture

Two planes. One governs, the other carries your data.

TokenSaver separates what decides from what executes. The control plane sets the rules without sitting on the path of your LLM calls. The data plane, an open-source gateway, applies them on every call, wherever you choose to run it.

Control plane

Commercial

Governs, off the critical path

  • Console, organisations and users
  • Virtual keys, teams and cost centers
  • Model catalogue and policy administration
  • Aggregated FinOps and alerts
  • Gateway enrolment, signed configuration, revocation

Signed configuration · revocations

Metadata · health signals

Data plane

Open source

Executes, on the critical path

  • OpenAI- and Anthropic-compatible gateway, with your LLM keys
  • Pipeline: cache, RAG, compression, PII, then the LLM
  • Hard limits: quotas, budgets, rate limits, model allowlists
  • Prompts, cache content and detailed replay
  • Agent egress capture

Secure

Your prompts stay on the data plane
In hybrid and self-host, the control plane receives metadata. Prompt content only goes up if you opt in.
Signed configuration
The gateway pulls its configuration and refuses any bundle with an invalid signature.
Enrolment and revocation
A one-time token enrols each gateway, and revoked keys are applied on the edge.
Controlled failure
If the control plane is unreachable, the gateway keeps serving on its last signed configuration.

Modular

One image, three roles
The same image runs as control plane, gateway or both. No fork of the LLM pipeline between modes.
Modules driven by policy
Cache, RAG, compression and PII are set per key and per organisation, not hard-coded in your apps.
Change mode, keep the product
Start in SaaS, move the gateway to your VPC, then bring the control plane home if you need to.

Built for scale

Control plane off the hot path
Limits apply on the gateway itself, with no round trip to the cloud on each call.
As many gateways as you need
Several gateways per organisation, by team, site or region, under one console.
Counters at the edge
Rate limits, monthly tokens and spend are counted where the traffic runs.
Ready to deploy
Helm profiles for hybrid, self-host and air-gap, plus Docker Compose.

Three deployment modes, one pipeline

SaaSHybridSelf-host / air-gap
Data plane (gateway)TokenSaver cloudYour infrastructureYour infrastructure
Control plane (console)TokenSaver cloudTokenSaver cloudYour infrastructure
Prompts, cache, detailed replayTokenSaverYour infrastructureYour infrastructure
Hard limits enforcedOn the TokenSaver gatewayOn your edgeOn your infrastructure
Quota limitAccording to your planNoneCapacity set by contract

And without the console?

The open-source gateway is 100% autonomous: with no control plane, you configure it by hand, with no quota limit. An enrolled gateway also keeps running after the 30-day trial, without the console. In both cases, the free VS Code extension for Claude Code plugs natively into it, locally, Keep my Claude mode included. The console and centralised governance are what the paid plans add. The public repository is coming soon.

What is open source, what is commercialPlans by deployment mode

Observability

See what your AI agents really do.

Agents, assistants, n8n workflows and API calls consolidated in one readable timeline, with the agentic graph to follow every chain of calls.

See the live agentic graph
01

Unified view of AI flows

Agents, assistants, workflows and API calls in a single timeline.

02

Real-time metrics

Latency, cost, tokens, cache rate and guardrail failures, per agent, team or model.

03

Agentic hierarchy

Multi-agent call chains and the tools invoked at each step.

04

Signed, immutable traces

Every call logged, signed and exportable as OpenTelemetry for your audits.

05

Root-cause analysis

Trace a behaviour back to its origin with execution context and decision lineage.

06

Alerts & anomalies

Drift, cost spikes, injections and policy breaches caught before they become incidents.

Observability

Optimisation · FinOps

Your agents read less, and understand better.

Model selection, semantic cache and compression by content type, applied inside your business flows. You keep control of cost, quality and latency.

LLM cost reduction

Semantic cache, smart compression and routing to the most efficient model, without touching your agents’ code.

Controlled latency

Automatic failover to the fastest endpoints, optimised streaming and configurable quality thresholds.

Enterprise AI FinOps

Budgets per project, overrun alerts, per-team cost allocation and traceability of every euro spent.

Typical savings by content type
JSON
−70–90%
Logs
−80–95%
RAG
−40–70%
Code
−40–70%
Diffs
−40–60%

Varies by workload, measured per run. Reversible compression (CCR) lets the agent fetch the original on demand.

Cyber & Zero Trust

Secure every AI flow, before it leaves your perimeter.

Nothing is allowed by default. Every action is identified, governed by policy and traced. Your teams move faster with AI; your organisation keeps control.

Ungoverned AI flows are the blind spot of the IS.

Sensitive data leaks

Personal data, secrets and confidential information injected into prompts or returned by models.

Agent hijacking

Injection attempts that push an agent outside its business scope or trigger unplanned actions.

Over-powered tools

Agents with too much autonomy calling sensitive tools without validation.

Shadow AI

Unapproved models, connectors or integrations used outside the catalogue.

Zero visibility

No clear view of who did what, with which agent, on which data, at what cost.

A trust posture applied to every flow.

  1. 01

    Never trust, always verify

    Every AI call is authenticated and tied to an identity: user, application key or agent.

  2. 02

    Default deny

    Unapproved models, tools and connectors are refused. Only your approved catalogue flows.

  3. 03

    Policies at the centre

    Rules defined in the platform, per organisation, workspace or key, not in each client app.

  4. 04

    Least privilege for agents

    Allowed tools, bounded autonomy and human approval when needed.

  5. 05

    Full visibility

    Every decision, allowed, blocked or pending, leaves a trace for the SOC and the CISO.

Govern by policies.

Configure once, apply everywhere: chat, API, agents, MCP.

Sensitive data protectionAnti-injection guardrailsTool allowlistOutput handlingContent moderationSafe model routingCatalogue & approvalOpposable traceability
TokenSaver is not
  • A simple model router
  • A cost-only tool
  • A chatbot
TokenSaver is
  • ✓The layer that decides what may pass
  • ✓That protects data inside the flow
  • ✓That bounds agent autonomy
  • ✓That proves what happened

Integrations

Keep your tools. Add the trust layer.

OpenAI and Anthropic compatible APIs, Python SDK, MCP tools or the open-source CLI: every entry point leads to the same control plane.

  • Enterprise assistants & chat→
  • LLM-compatible APIs→
  • Agents & tool calls→
  • MCP connectors→
  • AI traffic outside instrumented apps→
Claude CodeCursorCodexLangChainLangGraphCrewAIn8nMakeZapierOpen WebUILibreChatLiteLLMAgentgateway

Vibe coding at scale · Open-source VS Code & Cursor extension

Your agents say “all green”. TokenSaver shows the truth.

Every developer can launch agents in one sentence. The TokenSaver extension for VS Code and Cursor augments Claude Code with tracking and control of the code it generates: what agents really did, which files they wrote, what it cost and what was left broken. That is how vibe coding scales across the company.

Other extensions read logs. TokenSaver sits in the flow.

Monitors built on local Claude logs tell you what a session cost, after the fact. TokenSaver is on the path of the request, so it can act on it and leave a trace your team can open.

Observe

Prompts, file writes, tool calls and cost, captured session by session and joined to the same runs as the console.

Intervene

Serve from cache, mask PII and compress tool output before the model. In routed mode, also refuse a model your Agent Registry has not approved.

Prove

“Explain this file” traces the chain from prompt to tool call to file version. Share the session link with your lead or security team.

Two modes, your choice

Keep my Claude

Default

Keep your Claude plan. No extra LLM bill.

  • Anthropic keeps billing your plan, no LLM API key needed
  • Capture, session graph and IDE dashboard included
  • PII, cache and compression policies: 30 days on Free, no limit on paid plans
  • After 30 days on Free, it keeps working: plug it into a local gateway, the open-source data plane
  • Runs through a local proxy and a CA generated on your machine; Stop restores your Claude settings

Route with TokenSaver

Governed

Route through TokenSaver to govern models and policies.

  • Plugs natively into your local gateway (data plane): free, no quota limit
  • Claude Code goes through the TokenSaver pipeline, with your own LLM keys
  • Model picker limited to what your Agent Registry approves
  • Policies and quotas applied per key and team
  • Every intervention visible in the IDE and in AI Flows

Free and open source · coming soon to the VS Code Marketplace · no CLI or pip install required · macOS and Linux · prompt content capture is opt-in. The ✓ badge means a file was captured and policy-checked; it is not a compliance certificate.

Claude Code, augmented in VS Code or Cursor · 3 steps

  1. 1

    Install the open-source TokenSaver extension, soon on the VS Code Marketplace: it adds a TokenSaver panel to the activity bar. No CLI to set up.

  2. 2

    Connect your account. Keys stay in VS Code secret storage.

  3. 3

    Start a mode “Keep my Claude” keeps your Claude plan; “Route with TokenSaver” adds virtual keys, model choice and quotas.

  • Sidebar dashboard
  • Status bar · cost & alerts
  • Explain this file
  • File badges ○ ✓ ⚠
  • Review queue
  • Copy PR recap
  • Policy toggles
01 · In the IDE

One prompt, agents at work

Developers keep coding in VS Code with Claude Code, nothing changes in their workflow. The extension captures every prompt, file write, tool call and dollar in the background, even on their own Claude plan, with no extra LLM bill.

02 · The VS Code panel

Live cost, files and alerts

The TokenSaver panel shows session cost, $ per task, savings and the files the AI wrote, with loops, failures and retries in the status bar. “Explain this file” gives the provenance of any file. One click on a KPI opens the agentic graph.

03 · Under the hood

The proof behind every “done”

In our demo, the report was all green. The graph showed 12 failed tools and 2 red tests hidden behind an exit code 0. Errors, retries and their cost, traced call by call.

From one developer to every team

  • Cost per task
  • Idle-loop alerts
  • Untracked AI
  • Quotas & policies
  • Review queue & PR recap
  • Keep my Claude
See it on your codebaseGet notified at release

Frequently asked questions

Does TokenSaver replace my SIEM?+

No. It brings the missing AI layer (identity, policies, agent control and traceability) and exports its signals to your SIEM.

Do we need to change our AI tools?+

No. TokenSaver plugs into existing flows: chat, LLM-compatible APIs, agents, MCP connectors.

Who defines the rules?+

Your administrators and security teams, through centralised policies at organisation, workspace or key level.

How are autonomous agents handled?+

Least privilege: allowed tools, bounded autonomy, human approval steps when needed, and full traces of every decision.

From opaque AI to auditable AI.

Let us connect your existing flows in a few minutes.

Book a demoTalk to a security expert