Reference
Configuration
Nothing is required. The first-run wizard writes what it needs; everything else has a sensible default.
Where§
~/lampson/.lampson/config.json— provider, model, keys (sealed), preferences. Written by/setup, the web ⚙ and the provider pill. Shared by the terminal and the web UI.~/lampson/.env— optional overrides for the same settings, plus the knobs below..envwins overconfig.json.LAMPSON_HOME— move~/lampsonsomewhere else.
Provider§
LAMPSON_PROVIDER=deepseek # openai | deepseek | kimi | groq | grok | openrouter | ollama | anthropic | minimax
LAMPSON_API_KEY=sk-... # or LAMPSON_API_KEY_DEEPSEEK=… one per provider
LAMPSON_MODEL=deepseek-chat
LAMPSON_BASE_URL=… # any compatible endpoint (no trailing slash)
LAMPSON_WIRE=openai # openai | anthropic, when the preset cannot guess
Behaviour§
LAMPSON_PERMISSION=ask # ask | yolo | strict (the default mode)
LAMPSON_MAX_STEPS=40 # tool steps per turn
LAMPSON_BUDGET_TOKENS=400000 # cost budget per turn (cached tokens count 10%)
LAMPSON_COMPACT_AT=80000 # compact the context above this estimate
LAMPSON_MAX_TOKENS=8192 # output cap per response (auto-doubles on an empty answer)
LAMPSON_TIMEOUT=180 # seconds per model call
LAMPSON_EXPLORE_CAP=12 # read-only calls in a row before a nudge to act
Shell (Windows)§
Commands run through Git's bash by default. To use something else:
LAMPSON_SHELL=C:\Program Files\Git\bin\bash.exe
LAMPSON_SHELL_FLAG=-c
Unattended runs and approvals§
LAMPSON_TZ=-03:00 # for "daily 09:00" (default: the system's)
LAMPSON_PUBLIC_URL=https://… # how this Lampson is reached from outside
LAMPSON_WEBHOOK_URL=https://… # POST for each pending approval
LAMPSON_WEBHOOK_SECRET=… # HMAC-SHA256 signature of the body
LAMPSON_WEB_TOKEN=… # let a non-local script call the web API (Bearer)
See Scheduled tasks.
Per project§
Things you may want to commit, so your team gets them: .lampson/plugins/ (plugins), .lampson/mcp.json, .lampson/lsp.json, skills/, AGENTS.md. Lampson writes nothing else into your project.