Lemma Platform

Why it's ranked here

The engineering discipline is written down and mostly enforced by CI gates rather than review. State changes and the events announcing them commit in one transaction, consumers assume at-least-once delivery, and the generated client libraries fail the build if they drift from the committed API spec. The licensing split is deliberate: copyleft for the server, permissive for client tools. It runs on a laptop, on one self-hosted machine, or hosted, and can drive agents through a Claude Code or Codex subscription a team already pays for. A serious option for teams that want shared agents with real permissions.

What's good

Approvals are concrete. When an agent asks to use a tool, the request is parked until a person decides, and an unanswered request is denied when its timeout elapses. An "always allow" is remembered for that conversation and releases sibling requests in the same parallel batch, so nobody approves the same thing five times. The self-hosting guide is candid: agent sandboxes sit on a network that cannot reach the database, sandbox images are refused unless pinned by digest, and pod apps live on subdomains that never receive operator session cookies. The guide states plainly that the Docker socket equals root on the host.

Tradeoffs

The server core is AGPLv3, so modifying it and offering it over a network obliges you to publish those changes. Self-hosting asks for a dedicated Linux machine with 4 vCPU, 8 GB of RAM and 40 GB of disk, and the default sandbox mode gives the API and worker containers root-equivalent access to the host. Nothing works without your own model provider key, and app connectors want a separate Composio key. Email ships writing to disk, so invitations and password resets do not arrive until you configure SMTP or Resend. Installing the command-line tool with pip on an older Python quietly gets a stale release.

How to use it well

Fit: a small team that wants one agent handling shared work, such as triaging feedback or routing approvals, across the chat tools it already uses, with per-person and per-row access. Start on the hosted service or the desktop app before self-hosting. Install the command-line tool with uv rather than pip, run its doctor command when the app and the terminal disagree, and let your coding assistant generate the starter pod. On a server, either dedicate the machine or switch sandboxes to E2B, back up the encryption key with the database, and set a real email transport before inviting anyone. It is not a single-user coding agent.

Observed

License

AGPLv3 for backend, frontend and desktop; Apache-2.0 for CLI, SDKs, skills, stack tool and bundle format

Languages

Python (FastAPI backend, CLI, stack tool), Rust (desktop agent host and local daemon), TypeScript (web app and SDK)

Interfaces

Web app, command-line tool, Python and TypeScript SDKs generated from a committed OpenAPI spec, pod MCP server mount

Deployment

Hosted cloud, desktop app with a private local VM, or self-hosted via Docker Compose

Install surface

CLI published on PyPI as lemma-terminal, installed with uv tool install

Model providers

Claude Code, Codex, Cursor or OpenCode login via the desktop agent host, or any OpenAI- or Anthropic-compatible endpoint

Chat surfaces

Slack, Microsoft Teams, Telegram, WhatsApp and email

Sandboxing

Agent code runs in Docker, E2B or the desktop VM, never in the backend process