Overview & setup

Lectern is one cockpit for the coding agents you already use — Claude Code, Antigravity (Gemini), and OpenCode (OpenRouter + free models). It adds what they don't have alone: a Conductor that routes each task to the best model, a persistent brain (memory, skills, your machine's profile, a code graph), and one session history shared by the desktop app, the terminal UI, and the CLI. Local-first, Linux-first, open source under Apache-2.0.

1 · Get the pieces

Everything builds from one public repo. From a fresh clone:

• Desktop app: cd apps/desktop && npm install && npm run app (Linux needs the WebKitGTK dev libs; unsigned Windows/macOS builds come off public CI — see Downloads).
• Terminal UI: cd apps/tui && bun install, then ./target/debug/lectern tui — or compile the single binary with bun build --compile.
• CLI + daemon: built by the cargo command above (lectern, lecternd).

2 · Connect the AI you already pay for

Lectern drives each provider's own CLI, so your keys and logins never touch Lectern. Sign in to Claude Code (claude) and/or Antigravity (agy) once, then:

No keys at all? OpenCode's built-in free models work with zero configuration, and the mock backend exercises every pipeline for nothing.

Want zero cloud? Point Lectern at Ollama and it runs entirely on your machine — no API keys, no per-token bill, and nothing leaves the box. Install Ollama, pull a code-tuned model, and Lectern auto-detects it (the model picker flags the ones that are strong for coding):

Local quality scales with your hardware — a code model like qwen3-coder shines on a 24-32GB GPU or Mac, while a smaller one still handles everyday edits on a 16GB laptop. It's the most direct answer to the two things developers most want from an agent: no surprise token bill, and code that never leaves their machine.

3 · First run

The first run indexes the repo into the brain; every later session starts already knowing your code. In the desktop app the same happens when you open a folder. Without --apply, edits are proposals held behind the Apply gate until you accept them.

Where to next

By feature — what each part does and how it works:

• Chat commands — every slash command, and how skill replay works.
• Conductor & routing — how Auto picks a model, and how /conduct plans, fans out, and cross-reviews.
• The brain — memory, recall without draining tokens, the code graph.
• Scheduling — queue a prompt; the daemon runs it when due.
• The Hub — create, import, and share skills.

By surface:

• Desktop app — every surface of the cockpit, from tiles to the preview rail.
• Terminal UI — the whole product in your terminal.
• CLI & daemon — scripting, scheduling, encrypted export.
• Engine internals — backends, the store, the security model.
• MCP & channels — tools and remote access.