LUCID-TRAIN(1)LUCID-TRAIN(1)
Lucid Train, a coding agent that lives in your terminal
a coding agent that lives in your terminal · one ~6 MB Rust binary · 100% offline
$pip install lucid-train
$curl -fsSL https://lucidtrain.com/install.sh | sh
Lucid Train reads your repository, edits files as reviewable diffs, and runs and fixes tests, entirely from the command line. The whole agentic coding workflow in one portable Rust binary, with a Terminal-Bench-grade harness and your choice of model, local or remote.
the harness
- terminal-bench-grade
- Evolved AHE design. Contract-first prompting, execution-risk hints, a publish-state guard, and automatic context compaction.
- repo context engine
- Maps your codebase on launch and injects only the symbols each request needs. Grows .lucid/knowledge.md across sessions.
- plan, then build
- Plan mode proposes a reviewed plan with no side effects, then implements it step by step once you approve.
models
- local, zero-setup
- Opens with no key. /download installs Ollama and pulls the best model for your RAM: gpt-oss, Qwen3-Coder, DeepSeek.
- openrouter
- One key, 15+ models. Defaults to kimi-k2.7-code; also GPT, Gemini, Grok, DeepSeek V4, Qwen. Stored locally, sent only to the provider.
- self-hosted
- Point it at any OpenAI-compatible endpoint: vLLM, llama.cpp, SGLang, TGI. One flag.
tools
- edits as diffs
- write_file and edit_file apply changes and render reviewable green/red diffs. No opaque shell redirection.
- multi-format read
- read_file parses code, JSON, CSV, PDF, docx and images. Attach images for vision models.
- verify, don't guess
- web_search and read_url let the agent check facts, APIs and versions instead of hallucinating them.
- skills & mcp
- Install agent skills from skills.sh in-session and connect any MCP server. Standard SKILL.md packs work unmodified.
security
- risk tiers
- Every command is assessed: safe runs, caution and dangerous prompt, catastrophic (rm -rf /, fork bombs) refused in every mode.
- approval modes
- plan, auto, full-auto. Switching to full-auto always asks for confirmation first.
modes
- /agent
- Full tools. Edits files and runs commands.
- /plan
- Propose-only. A reviewed plan; nothing executes.
- /research
- web_search and read_url only. Cited answers, no shell.
install
$pip install lucid-train
$curl -fsSL https://lucidtrain.com/install.sh | sh
Also on pipx, uv, Homebrew and winget. Windows, macOS and Linux.