Sema

Your own Claude Code-grade coding agent — in the terminal.

Quick start

npm install -g @sema-agent/cli

# Point sema at any OpenAI-compatible endpoint you control or pay for:
export OPENAI_BASE_URL=https://your-llm-endpoint/v1
export OPENAI_API_KEY=your-key
export OPENAI_MODEL=your-model-id

sema "fix the failing test in this repo"

Exported this way, the three OPENAI_* variables are equivalent to MODEL_GATEWAY_BASEURL, MODEL_API_KEY and MODEL_ID: either spelling works for sema, sema -p and sema doctor alike, and when both are set the MODEL_* value wins. The OPENAI_* names are read only from your shell: in the env block of a settings file, use the MODEL_* names.

No account, no sign-up, no default cloud — the agent runs locally and talks only to the model endpoint you configured. The official plugin marketplace is not added for you, and sema does not fetch its catalog until you add it (/plugin offers to).

Sema infers the context window and output-token budget for known model families automatically. If your model id is not recognized, set MODEL_MAX_TOKENS yourself — for reasoning-heavy models use at least 32000, or the whole budget can be consumed by reasoning and the run fails with a max_tokens truncation error.

• sema — interactive REPL (bare prompt argument starts the REPL with that task)
• sema -p "…" — non-interactive print mode for scripts and pipes
• sema update — self-update from the npm registry

Security notes

<system-reminder> tags embedded in user-provided message text are treated as plain data — they carry no system-instruction effect. Genuine system reminders are minted by the engine itself and carry a session-scoped integrity mark.

License

BUSL-1.1 — free for personal, educational, research, and other non-commercial use; commercial production use requires a commercial license. Converts to Apache-2.0 on 2030-07-13.

Claude Code is a product of Anthropic. Sema is an independent project, not affiliated with or endorsed by Anthropic; the name is used only to identify what we compare against.