Lemma (GitHub Repo)
GitHub
Lemma turns a team’s apps and agents into one shared, permissioned system. It keeps learning from the work, so the harness gets better as people use it.
Based on reporting by GitHub — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Lemma is pitching a different way to build agentic software: not one chat session at a time, but a shared system a whole team can use. The project calls that system a pod — a self-contained environment that holds state, agents, workflows, permissions, and apps. The same records can be read and written by people and agents, with access shaped by roles, table grants, resource visibility, and approval gates.
The big idea is persistence. Lemma says the harness does not stop when a session ends. It can keep running on schedules, webhooks, and table events, and it can turn repeated corrections into standing instructions, repeated sequences into workflows, and recurring judgment into an agent role. In other words, the system is supposed to improve because it is being used, not despite it.
The build flow is aimed at existing coding agents like Claude Code, Codex, Cursor, OpenCode, and Antigravity. You describe the job, and the agent writes the whole system as files: the app, the tables underneath it, the agents, the workflows, and the permissions. Then the same CLI imports and verifies the result. Lemma also wants that system reachable in a lot of places: a URL, Slack, Teams, Telegram, WhatsApp, or email.
There’s a lot of emphasis on control and portability. Lemma is open source, can run on a laptop or a server, and offers a hosted option called Lemma Cloud. It says the stack uses your own subscriptions, Lemma-managed models, or OpenAI- and Anthropic-compatible providers. The docs also get unusually specific about setup: use uv tool install rather than pip, because the CLI and server version skew can otherwise quietly drift apart. That kind of detail tells you the project is trying to be infrastructure, not demoware.
The repo also pushes the “pod” idea pretty far. It treats tables as typed business data, files as markdown memory, workflows as graphs with waits and human approval steps, and surfaces as the chat and email fronts that feed the same core. Lemma is not shy about the pitch: one system, many users, many agents, one shared state layer.
My take — AI-written commentary, not fact-checked reporting
Lemma gets the important part right: agents are boring until they have memory, permissions, and a place to live. The web is already full of clever demos that forget everything the moment the tab closes; this is the opposite, and that’s healthier. Also refreshing: it talks less like a magic wand and more like plumbing, which is usually where the real value hides.
Read more about this at: GitHub
