Mesh is free. This is why we built it, who should use it, and why we are giving it away rather than selling it.
What: Mesh, an open source gateway that puts keys, budgets, model access and cost logging in front of every LLM call your company makes
Licence: Apache 2.0. Free to use, fork and build on
Cost: Zero to license. About ten dollars a month to host a single instance yourself
Who it is for: Teams of ten to a hundred people with more than one product or tool calling AI models, and the person who signs the bill
The problem it solves
Picture a team that has decided to take AI seriously.
In the first month, one person has a commercial API key and a good idea. By the third month there are twelve people using that key. It is in two products, a reporting script, an internal tool someone built on a Friday, and at least one laptop nobody has audited. Then the invoice arrives. It is larger than anyone expected, and the only question that matters cannot be answered: who spent this, on what, and was it worth it?
Now picture the team a year further on. AI is in production. Finance wants cost by product. Security wants to know which models are allowed and who can reach them. A client's procurement team wants to know how access is governed. The provider's dashboard shows one number, and the answer to all three questions is a spreadsheet someone maintains by hand.
Both teams have the same problem. The model providers give you very little control over how a key is used. For the most part it is on or off. There is no built-in way to say this person gets this much, this product may only call these models, and show me every request by who made it. A provider has no incentive to help you spend less or see more. That control has to be yours.
What Mesh does
Mesh is a lightweight gateway that sits between your products and the model providers. Your code calls Mesh. Mesh calls Anthropic, OpenAI, Gemini, Ollama, or anything that speaks a compatible shape.
On the way through, it does the things the providers do not. It issues its own keys, one per person or per product. Each key carries a budget, a rate limit, and a list of the models it is allowed to use. Every request is logged against the key that made it, with usage, cost and latency. One endpoint, one request shape, every provider you allow, and a single dashboard where the owner of the bill can finally answer the question.
It is fast. In our own testing it handled a thousand calls in under a second, because the request path reads from an in-memory snapshot and never touches the database per call. It is cheap to run: a single instance, one database, on the order of ten dollars a month. And it ships with a test server that impersonates every provider, so you can exercise the whole system, including failure modes, without spending a cent on real calls.
Why we built it rather than bought it
We looked at the alternatives, and the criteria were simple.
The paid gateways charge by volume. Volume is the thing you are trying to control, so paying for it per request puts the incentive in the wrong place from day one. The large open source gateways are excellent, and they are built for scale most teams do not have. Running them is a project in itself.
What we wanted was the smallest thing that gives a team of ten to a hundred people complete visibility and control, at negligible cost, that we understood down to the last line. That is what Aditya Raj, one of our engineers, built. He named it Mesh before he showed it to anyone, which tells you something about how he works.
Who this is for
Not everyone. If you run high-volume, multi-tenant traffic, use something built for that. The repo says so in plain language, and we would rather you read it there than find out in production.
Mesh is for the two teams in the opening scenario. If you are starting, put a gateway in on day one and the month-three invoice never becomes a mystery. If you are advancing, this is how you give finance cost by product, give security model-level control, and give a client's procurement team a straight answer about governance. In our experience that describes most agencies, most product teams, and most of the organisations we work with on AI enablement.
Why give it away
We are a research company. Everything we sell rests on showing our work: which query, which model, which run, where the number came from. A gateway that logs every model call is that principle applied to operations. Keeping it to ourselves would have been out of character.
Aditya built it and his name belongs on it. The people building in this space are his peers, and they should know who did the work.
And Bold gets something from it too. We would rather be known for putting useful things into the world than for calling ourselves an AI company. A tool people can fork, run and build on is proof of work. Stars on a repository are a better claim than a slide.
So Mesh is public, under Apache 2.0, at github.com/Bold-AI-Inc/meshcore. The Go gateway and admin API, a Next.js admin panel with public docs, SDKs in Python and TypeScript plus a raw curl reference, and the test server.
What we would ask of you
Run it. Break it. Tell us what you built on it. The contributing guide asks for one concern per pull request and a before-and-after comparison on anything that touches the request path. It is the same standard we hold ourselves to on client work.
If you lead an organisation that is moving from experimenting with AI to running it, start with one question: can you see every model call your company makes, by person, by product, by cost? If the answer is no, that is the first thing to fix. Now there is a free way to do it.
Alan Macdonald is co-founder and CEO of Bold AI. Aditya Raj built Mesh.
Coming soon from Aditya: the engineering decisions behind Mesh, why it runs as two planes, why the request path never touches the database, and what we learned building a test server that impersonates every provider.
