Request early access.
Tinfirm is in development and not released yet. We're opening it to a few teams who can run it on their own infrastructure and will tell us what's missing.
- You host it
- On Docker Compose or Kubernetes, with your own runner hosts and your own LiteLLM gateway.
- Your models, your bill
- Model spend goes to your own provider accounts, metered per person by your gateway.
- It will change
- Expect changes between versions, and expect us to ask what your specialists got wrong.
- Licence and pricing
- An MIT core and a free Community edition; Pro is €15 a seat a month. Early-access teams get a discount off the list. See pricing
Questions teams ask first.
What do we need to run it?
Docker Compose or Kubernetes for the server and its PostgreSQL database, a LiteLLM gateway (bundled with Tinfirm, or your own), and at least one runner host with Docker or Kata Containers that can reach the server.
Which models and agent runtimes does it support?
Any model your LiteLLM gateway serves. The runtimes verified with the model proxy are Claude Code, Codex, pi and OpenCode. A member's own Claude or ChatGPT subscription can also serve their own work, on their own runner host.
Does our data leave our infrastructure?
Tinfirm has no vendor cloud: it runs where you run it. Model requests go to the providers you configure in LiteLLM, on your own accounts, and your systems are reached with the accounts you allow.
Can people use it from their own AI client?
Yes. Tinfirm is an MCP server for people's own AI clients, and signs them in itself over OAuth. A conversation started there acts with the person's own accounts and budget, like one started in Tinfirm.
Is it only for software teams?
No. To Tinfirm, code and non-code work are the same thing: a job with a workspace, tools, decisions and results. Business analysts, testers and finance teams use the same triggers, inbox and approvals as developers.