Command line

Every command has --help. The reasoning model is chosen with --llm backend[:model] (for example openrouter:qwen/qwen3.7-flash) and its reasoning with --reasoning default|off|minimal|low|medium|high; see LLM backends.

Every command reads ./.env first. Variables already set in the environment win, and values are never printed.

thinkless doctor

Checks the environment: Python, torch and the accelerator, which optional backends are installed, which API keys are set, and the active settings. Include its output in bug reports.

thinkless demo

Runs the support agent on one ticket and prints the trace tree, the action and the reply.

Option	Default
--ticket	T-001	A scenario id from the bundled set (T-001 to T-053)
--message, --customer		Your own ticket text instead of a scenario
--mode	hybrid	hybrid, llm or models; repeat to compare
--llm	local	Reasoning model
--device	auto	Device for local models
--threshold	0.8	Engine default threshold
--view	off	Write and open the HTML viewer

thinkless bench support

Runs every scenario in every mode and writes results.json, report.md, per-ticket traces and viewer.html to the output directory.

Option	Default
--mode	llm, hybrid, models	Repeat to choose
--llm	local	Reasoning model, also the LLM decision provider
--limit	all	Only the first N tickets
--out	.thinkless/bench/support-<time>	Output directory
--reference	anthropic:claude-sonnet-5	Price list used to estimate cost
--threshold	0.8	Engine default threshold
--jev	off	Add TypeSafe Jev to the cascade (needs TYPESAFE_API_KEY)

thinkless bench intents

Accuracy, calibration and the simulated cascade on a public dataset.

Option	Default
--dataset	banking77	banking77, clinc150 or emotion
--provider	gliner, laya, llm	Repeat to choose; llm is the cascade fallback
--limit	500	Examples sampled from the test split
--seed	13	Sampling seed
--target	0.95	Accuracy target for threshold recommendations
--out	.thinkless/bench/intents-<dataset>-<time>	Output directory

thinkless calibrate DATA

Sweeps thresholds for one question and provider on labeled JSON Lines and recommends the lowest threshold that meets the target.

Option	Default
--provider	gliner	gliner, laya, llm or jev
--kind	choice	choice or yes_no
--question		The instructions to ask
--labels	from the data	Choice options, comma-separated
--demo-question		Use a question from the support demo by name
--text-field, --label-field	text, label	Field names in the data
--target	0.95	Accuracy the accepted answers must reach

thinkless trace

Command
trace ls [DIR]	Recent traces with their headline numbers
trace show FILE	One trace as a tree with its summary
trace view [PATHS...] [--out FILE] [--no-open]	Write the self-contained HTML viewer
trace export PATHS... --out FILE	Traced decisions as labeling rows for calibrate; filter with --question, --status, --plane, --limit
trace drift --baseline PATHS --current PATHS	Compare two periods question by question; exits with 1 when a question drifted

thinkless shadow

Reads the logs written by shadow mode.

Command
shadow report LOGS...	Agreement with a 95% interval, share settled without an LLM, cost and latency on both sides, fitted thresholds and a verdict per question
shadow export LOGS... --out FILE	Disagreements (or every call with --all) as labeling rows for calibrate

shadow report options: --target (agreement the lower bound must reach, default 0.95), --min-calls (default 100), --volume (decisions a month, for the monthly projection) and --json. shadow export options: --question, --all and --label-from primary|shadow|none.

thinkless serve ENGINE

Serves an engine over HTTP; see serving. ENGINE is module:attribute, an engine or a function that returns one.

Option	Default
--questions	QUESTIONS in the engine's module	Registered questions, as module:attribute
--host, --port	127.0.0.1, 8080	Where to listen
--api-key-env	THINKLESS_API_KEY	Variable holding the bearer token clients must send
--registered-only	off	Refuse question specs that are not registered

thinkless mcp ENGINE

Serves the registered questions as MCP tools, one decide_<question> tool each. --transport stdio (default) or streamable-http.

Environment variables

Variable	Default
THINKLESS_TRACE_DIR	.thinkless/traces	Where the CLI writes traces
THINKLESS_DEVICE	auto	Default device for local models
THINKLESS_CAPTURE_CONTENT	true	Record inputs and outputs in traces
THINKLESS_LOG_LEVEL	WARNING	Log level for the CLI
THINKLESS_PRICING	bundled table	Path to a pricing TOML file
TYPESAFE_API_KEY		For SystemOne.jev()
OPENROUTER_API_KEY		For openrouter: models
ANTHROPIC_API_KEY, OPENAI_API_KEY		For hosted reasoning models
THINKLESS_API_KEY		Bearer token for thinkless serve