Today we're shipping the @edgee_ai Claude Mod (and @MistralAI Large 4 is already available through it 🎉) You pay for a Claude plan, and Claude Code is where you work all day. But some tasks deserve another model, and until now that meant leaving your session. Not anymore. Launch Claude Code with the Edgee CLI and the mod comes along, right inside the terminal: → A live view of your session: requests, tokens, cache hits → A cost savings donut that shows what compression and routing save you → A model picker: search 200+ models, press Enter, and the next request goes there. Same conversation, same context. → One click to go back to your Claude plan Speaking of models: Mistral Large 4 just came out, and it's already available in Edgee. It's the one we pick in the demo below. 👇

@getqonto chose @edgee_ai to manage coding-agent usage across 500 engineers. You can easily imagine how proud the Edgee team and I are to make this official announcement ;) Migration began in late July. By 10 September, the rollout had reached 500 engineers, following an opt-in phase that grew from 10 to 350 in five weeks. Qonto reports 37.7% lower costs from a combination of compression and a first simple model rerouting rule. In our interview, Kevin Prettre, Director of Infrastructure and Engineer Support at Qonto, explains the decision to buy an Agent Gateway, the compliance review, and deployment through Jamf without changing engineers' tools. "In practice, Edgee has become our governance layer for AI usage." Read the conversation here edgee.ai/blog/posts/edgee-ha…

Yesterday at @AISummitBCN, I ran a workshop on controlling agentic coding costs. Two years ago, AI was a feature inside an app. Today, autonomous coding agents run in loops on developers' machines, and the budget broke at the agent layer. The room was packed with engineers living this exact shift: 1. The token budget explosion 18 months ago, "agent tokens" wasn't a line item in any engineering budget. Today, an active engineer can easily hit $1,000+/month, and engineering bills are doubling every six months. 2. Compression vs Routing: where the real leverage is While token compression (tool trimming, context optimization) realistically saves 15-20% when measured honestly, intelligent model routing across closed and open-weight models is what actually moves the needle by up to 10x. The questions went straight to the operational reality: cache arithmetic, context window degradation, and how to implement budget-aware routing policies without slowing down developers. Beyond the workshop, I'm leaving Barcelona deeply impressed by the tech scene here. The momentum, curiosity, and caliber of the engineers and founders I met were incredible. Huge thanks to everyone who joined the session and came to chat afterward.

Story of a pivot [1/5]: How we found the problem @edgee_ai needed to solve Today, Edgee helps companies reduce the cost of their AI coding assistants through an Agent Gateway that compresses, monitors, and routes LLM requests. For enterprise teams with more than a hundred developers, we've seen savings of 50-70%, depending on usage. We launched our per-developer offer in May. A first major customer trusted us, others followed, and our sales pipeline started growing. Getting there took a much less direct path than our homepage might suggest. A year earlier, we were running an edge computing platform, somewhere between a CDN and a tag manager. We handled 11 billion HTTP requests a month for media and e-commerce websites. We had reached 200K in ARR and figured out how to sell it. But every deal required buy-in from marketing, engineering, and often the data team. Sales cycles took six to eight months. I spent the summer of 2025 looking at that trajectory and asking myself how we could build a much faster-growing business. We already knew how to handle large volumes of traffic reliably. We had deep experience with HTTP proxies and routing. What we still needed was a market where that expertise solved a more urgent problem. Finding it involved several changes of direction, some uncomfortable conversations, and a lot of customer feedback. I'm going to tell that story over five posts. Because looking back, the most useful part is understanding how each decision brought us closer to a problem customers would actually pay us to solve.

Quand @DidierGirard himself parle d'@edgee_ai sur BFM Business… on se sent très reconnaissants, et on se dit qu’on est sur la bonne voie 😉 Merci Didier pour la mention 🙏

Yesterday, OpenAI, Claude, and Grok all suffered disruptions. If your engineering team was running Claude Code, Cursor, or Codex, you felt it immediately. When developers run agentic coding workflows, a provider outage breaks active multi-step execution loops, wipes working context, and pulls engineers straight out of their flow. Resilience requires routing across independent failure domains: different model providers and separate foundation models. That is why we built transparent multi-provider routing into @edgee_ai. When a frontier provider degrades, Edgee immediately fails over to an operational provider without dropping the agent session or losing context mid-task. Developers keep building, and platform leads do not have to spend their afternoon watching status pages. Don’t thank me, you’re welcome ;)

In the last 90 days, European engineering team onboarding on @edgee_ai often brought the exact same constraint: Zero data retention, hosted 100% in Europe. When you have hundreds of autonomous coding agents reading proprietary codebases and running execution loops around the clock, that requirement makes complete sense. Data sovereignty is non-negotiable for European enterprise legal teams. The problem has always been execution. In AI infrastructure, compliance is never an acceptable excuse for poor latency. If a European provider adds 300ms to Time-to-First-Token or runs at half the throughput of US clusters, developers will simply find a way to bypass it. Legal writes the policy, but engineers pick the tools that do not break their flow. European inference does not just need to be sovereign. It has to be fast enough that engineers cannot tell the difference. That is why we are partnering with @sferenceai The Sference team built something rare: strict European hosting with zero data retention, without sacrificing raw speed or TTFT. Starting today, Sference endpoints are natively integrated into Edgee. Engineering teams can now route high-volume coding agent workloads straight to European infrastructure while keeping Edgee's observability, compression, and routing layers in place. We spent the last two weeks stress-testing their clusters under heavy agentic traffic, and the endpoints are live in the console now. Thank you to @jernejstrasner and Sference team!

The first time a recruiter asked where I got my CS degree, I was playing guitar in bars across Paris. I didn't have one. In fact, I don't have a single university diploma to my name. I taught myself to code out of sheer curiosity because I wanted to build things. That habit turned into my first junior developer job, which eventually led to becoming a CTO, and today building Edgee as a founder. For a long time early on, you feel like you're supposed to hide that background. You worry people will think you're missing some secret foundational knowledge taught in lecture halls. Now that my track record is established, I bring it up openly. Every week, I meet incredible entrepreneurs, developers, and builders. The ones shipping the most ambitious work rarely talk about where they studied. They pull up their terminal, walk you through a tricky production failure from last Tuesday, or show you an architecture decision they had to invent from scratch. A diploma might get you past an initial resume filter. But execution, curiosity, and taste don't come with a stamp from a university. My old guitar is still sitting in the corner of my office. When people ask what prepared me most for running a tech company, I usually tell them the truth: years of trying to get people to like my music and playing gigs to empty rooms.

Anthropic announced reducing Claude usage limits by 17%. It is another clear signal that the era of providers handing out heavily subsidized tokens is winding down. For teams building with agentic coding, the operational reality is setting in fast: provider quotas will keep tightening, pricing models will keep evolving, and relying 100% on a single closed API is becoming an expensive single point of failure. Open-weight models are the logical counterbalance. They give teams control over cost and availability. The bottleneck has always been operational friction: getting an entire engineering team to plug their coding harnesses (Cursor, Claude Code, Codex, Github Copilot...) into open-weight models without rewriting configs or losing visibility. That is why we built @edgee_ai's gateway architecture. Any team, from 5 to 500 engineers, can enroll their developer harnesses in under 60 seconds. The gateway handles three core layers automatically: 1. Token compression to reduce context bloat before payload dispatch 2. Team-wide observability across all seats and agent harnesses 3. Smart routing that directs traffic dynamically between frontier APIs and open-weight models The result is cutting token spend by up to 70% without interrupting how developers build every day.

Edgee went from 352M tokens in February to 662B in August. That’s a 1,880x jump in 6 months, alongside 529 new users. We're still miles behind OpenRouter in total volume. But OpenRouter is built for routing LLM calls, while agentic coding workflows chew through context and tokens in a completely different way. Agentic coding needs a dedicated gateway. More on our first major reference in a few days. We can hardly wait.

Three weeks ago, a prospect (a bank) told me: "This looks great, Edgee, but if we can't run your gateway on our own infrastructure, we won't be able to do anything with you!" A week later, we provided them a Helm chart to deploy our technology on their own servers. I love those little bursts of pressure! They guide us, help us find our positioning, and also let us show what we're made of. Since then, the POC has kicked off with this prospect. Wish us luck! 🤞 So, just so you know, @edgee_ai is now available on-premise for the most critical environments ;) @KawaKiwi938 wrote a blog post about that, feel free to go to our website a read it.

My Claude Code session cost $56. The same one on open models? Almost free with @edgee_ai

Token Compressor V2 is available today at @edgee_ai . A few months ago, we released our first version of a token compressor, built on RTK technology. The response was immediate, and it gave us a clear picture of how effective it could be. In real-world conditions, this compression saves up to 10% on token costs. We wanted to go further. Today, we are adding two complementary compression algorithms that deliver significant gains without compromising model performance. Token Compressor V2 features 3 compression algorithms, each of which can be enabled independently: - Tool Result Trimming: still based on RTK, reducing tool result pollution (10% cost savings) - TSR (Tool Surface Reduction): a kind of super MCP that reduces the declaration surface of MCPs that are almost never used (10% savings) - Output Brevity: radically reduces model verbosity in its responses. Since output tokens are the most expensive, this strategy is extremely effective and delivers 30% savings In total, Edgee Token Compression saves you up to 50% on your agentic coding bill... and it takes just 1 minute to install.