CLI Commands

mentu is the CLI entry point for the engine. Commands are organized into groups: execution, CIR, auth and vault, infrastructure, management, audit, utility, and AI training.

mentu <command> [subcommand] [flags]
mentu --version
mentu --help

Global flags available on most commands:

Flag	Description
`--verbose`	Verbose output
`--quiet`	Suppress non-essential output
`--config <path>`	Override config file location

Execution

Commands that run recipes, sequences, and agent steps.

mentu do

Run an intent directly. This is the default command (invoked when no subcommand is given).

mentu do <intent...>

Passes your intent to an agent that selects the right tools and executes.

mentu step

Run a single agent step with a prompt.

mentu step -P <prompt-file> [flags]

Flag	Type	Description
`-P`	path	Path to prompt file
`-p`	string	Inline prompt text
`--max-iterations`	int	Maximum agent iterations
`--completion-keyword`	string	Keyword that signals completion (e.g. `STEP_COMPLETE`)
`-m`, `--model`	string	Model ID (e.g. `claude-opus-4-6`)
`--backend`	string	Backend to use (`anthropic`, `ollama`)
`--subagent-max-turns`	int	Max turns for subagents
`--pipeline-mode`	bool	Run in pipeline mode
`-n`, `--name`	string	Step name for logging

Flag	Type	Description
`-P`	path	Override prompt file
`--from`	string	Resume from step label
`--model`	string	Override recipe model
`--backend`	string	Override backend
`--auto`	bool	Non-interactive mode, approve all gates
`--var KEY=VALUE`	kv	Inject a variable (repeatable)
`--transfer`	string	Transfer mode: `none`, `worktree`, `mirror`, `rsync`
`--vm`	bool	Run tool calls inside a VM sandbox
`--worktree`	bool	Shorthand for `--transfer worktree`
`--no-cache`	bool	Skip cached results
`--max-budget`	USD	Maximum spend limit for the run
`--no-ane`	bool	Disable ANE inference
`--notify`	bool	Send notification on completion
`--workspace`	path	Override workspace directory

mentu pipeline

Run a chain of sequences defined by a pipeline JSON. Alias: chain.

mentu pipeline <pipeline-name> [flags]

Flag	Type	Description
`--from`	string	Resume from a specific sequence
`--model`	string	Override model for all sequences
`--backend`	string	Override backend
`--transfer`	string	Transfer mode
`--vm`	bool	VM isolation
`--workspace`	path	Override workspace directory

mentu compound

Multi-layer compound execution (layered sequences with shared substrate).

mentu compound <compound-name> [flags]

Flag	Type	Description
`-m`, `--model`	string	Override model
`--backend`	string	Override backend
`--transfer`	string	Transfer mode
`--vm`	bool	VM isolation
`--no-cache`	bool	Skip cached results
`--max-budget`	USD	Maximum spend limit
`--max-depth`	int	Maximum layer depth (default 10)
`--max-runtime`	seconds	Maximum runtime (default 14400)
`--from`	string	Resume from layer label
`--workspace`	path	Override workspace directory

mentu parallel

Run multiple recipes concurrently. Alias: fleet.

mentu parallel <names...> [flags]

Flag	Type	Description
`--parallel`	int	Max concurrent runners
`-m`, `--model`	string	Override model
`--backend`	string	Override backend
`--no-cache`	bool	Skip cached results
`--workspace`	path	Override workspace directory
`--from`	int	Resume from index

mentu adversarial

Red team / blue team adversarial pairing. Two agents argue, a judge picks.

mentu adversarial <recipe> [flags]

Flag	Type	Description
`-m`, `--model`	string	Override model
`--max-budget`	USD	Maximum spend limit

mentu convergent

Run N strategies in parallel and pick the best result.

mentu convergent <recipe> [flags]

Flag	Type	Description
`-m`, `--model`	string	Override model
`--max-budget`	USD	Maximum spend limit
`--parallel`	int	Number of concurrent strategies

CIR

Commands for the Contradiction-Invariance Record substrate.

Subcommand	Description
`query`	Filter signals by recipe, step, type, trust, confidence
`capture`	Inject a signal into the substrate
`search`	Semantic search over signals
`stats`	Summary counts by type, recipe, trust
`contradictions`	List unresolved contradictions
`patterns`	List detected patterns
`relate`	Create a relationship between two signals
`reinforce`	Reinforce a signal's confidence
`decay`	Apply time-based confidence decay
`embed`	Regenerate embeddings
`compact`	Compact the JSONL ledger
`export`	Export signals to a file
`validate`	Validate ledger integrity
`evaluate`	Evaluate signal quality
`cost`	Show CIR storage cost
`sync`	Sync signals across stores
`chain`	Walk the evidence chain for a signal
`train`	Extract training data from signals
`surprisal`	Compute surprisal scores
`perceive`	Classify input via ANE perception
`judge`	Run trust judgment on a signal
`promote`	Promote a pattern to a rule
`capabilities`	List available CIR capabilities

mentu health

Health metrics across the CIR store.

mentu health [flags]

Flag	Type	Description
`--days`	int	Days of history to include
`--format`	string	Output format

mentu verify-ledger

Ledger integrity verification. Walks the hash chain and reports mismatches.

mentu verify-ledger [flags]

Flag	Type	Description
`--format`	string	Output format

Auth & Vault

Authentication and secret storage.

mentu auth

API authentication (OAuth tokens).

mentu auth <subcommand>

Subcommands: login, status, logout

Subcommand	Description
`login`	Start OAuth flow, store token
`status`	Show current auth identity
`logout`	Clear stored token

mentu vault

Secrets management backed by the macOS Keychain.

mentu vault <subcommand> [flags]

Subcommands: set, get, list, delete, migrate

Flag	Type	Description
`--scope`	string	Vault scope (default or custom namespace)
`--tier`	string	Secret tier (`user`, `workspace`, `recipe`)

mentu vault set OPENAI_API_KEY
mentu vault list --scope workspace

Infrastructure

Commands for scheduled, monitored, and substrate-level operations.

mentu temporal

Scheduled execution of recipes at defined times or intervals.

mentu temporal <subcommand>

Subcommands: run, daemon, list, status, history, health, add, remove, enable, disable

See Scheduling guide.

mentu cron

Cron-style scheduling interface (delegates to temporal).

mentu cron <subcommand>

Subcommands: add, list, remove, run, status

mentu sentinel

Continuous monitoring. Runs a sequence when a condition triggers.

mentu sentinel <subcommand>

Subcommands: run, daemon, list, status, history, add, remove, enable, disable, resolve, check

mentu substrate

Meta-operations on infrastructure state (temporals, sentinels, workers, vault).

mentu substrate <subcommand>

Subcommands: run, apply, rollback, diff, list, status, history, show, reset

Management

Commands for recipes, scripts, webhooks, and registered tools.

mentu recipe

Recipe management (create, validate, export, commit).

mentu recipe <subcommand>

Subcommands: create, list, export, import, clone, validate, commit, insights

Subcommand	Description
`create`	Scaffold a new recipe from a template
`list`	List recipes in `~/.mentu/recipes/`
`export`	Export a recipe and its prompts to a tarball
`import`	Import a recipe tarball
`clone`	Copy a recipe to a new name
`validate`	Validate the JSON schema
`commit`	Register as a reusable recipe (constitutional gate)
`insights`	Print execution insights from past runs
`discuss`	Interactive discussion about a recipe

mentu script

Script execution (TypeScript SDK scripts).

mentu script <subcommand> [flags]

Subcommands: run, list

Flag	Type	Description
`--var`	kv	Inject a script variable
`--workspace`	path	Override workspace directory
`--timeout`	seconds	Script timeout

Flag	Type	Description
`--event`	string	Event name to listen for
`--url`	string	Callback URL
`--secret`	string	HMAC shared secret

mentu worker

Edge worker lifecycle — deploy, secrets, logs.

mentu worker <subcommand>

Subcommands: setup, create, deploy, list, status, delete, secret, logs

secret sub-subcommands: set, delete, list

The CLI auto-detects the deployment path. If Cloudflare credentials exist in the vault (via mentu worker setup), all commands use your CF account directly (BYOK). Otherwise, commands route through Mentu's managed edge at api.mentu.ai, deploying to *.edge.mentu.ai. Same commands, invisible backend switch.

Subcommand	BYOK (own CF account)	Managed (no CF setup)
`setup`	Stores CF credentials in vault	Not needed
`create`	Local scaffold	Local scaffold (same)
`deploy`	Direct CF upload	POST to api.mentu.ai
`list`	CF /workers/scripts	api.mentu.ai/v1/edge/workers
`status`	CF /workers/services	api.mentu.ai/v1/edge/workers/{id}
`delete`	CF delete	api.mentu.ai delete
`secret *`	CF secrets API	api.mentu.ai secrets proxy
`logs`	WebSocket tail	Not yet available (managed)

Managed workers have a free tier limit of 3 active workers per user.

# No setup needed for managed edge — just deploy:
mentu worker deploy my-api
# → https://my-api-a1b2c3d4.edge.mentu.ai
 
# Or use your own CF account:
mentu worker setup
mentu worker deploy my-api
# → https://my-api.your-subdomain.workers.dev

mentu dispatch

Async command dispatch (plugin tools and built-in tools).

mentu dispatch <subcommand> [flags]

Subcommands: send, watch, status, list

Flag	Type	Description
`--params`	json	JSON params to pass to the tool
`--device`	string	Target device ID

mentu plugin

Plugin management.

mentu plugin <subcommand> [flags]

Subcommands: list

Flag	Type	Description
`--json`	bool	JSON output

See Plugins guide.

mentu ratchet

Code quality ratchets (monotonic regex baselines).

mentu ratchet <subcommand> [flags]

Subcommands: check, baseline

Flag	Type	Description
`--source`	path	Source directory to scan
`--config`	path	Path to `ratchets.json`

See Ratchets guide.

Audit

Web and binary audit commands.

mentu audit web

Audit a web target (inspects network, scripts, and headers).

mentu audit web <url> [flags]

Flag	Type	Description
`--intercept`	bool	Intercept network traffic
`--format`	string	Output format (`json`, `markdown`)

mentu audit binary

Audit a binary (symbols, strings, imports, linked libraries).

mentu audit binary <path> [flags]

Flag	Type	Description
`--full`	bool	Full audit including decompilation hints
`--deep`	bool	Deep semantic analysis
`--format`	string	Output format

mentu audit full

Combined web and binary audit.

mentu audit full <url> [flags]

Flag	Type	Description
`--binary`	path	Local binary to audit alongside the URL
`--intercept`	bool	Intercept network traffic

Utility

General-purpose utility commands.

mentu init

Initialize a new mentu workspace in the current directory.

mentu init [name] [flags]

Flag	Type	Description
`--tier`	int	Governance tier (1, 2, or 3)
`--governance`	bool	Enable governance features

mentu list

List all recipes in the workspace.

mentu list [flags]

Flag	Type	Description
`--validate`	bool	Validate each recipe on load

Flag	Type	Description
`-l`, `--lines`	int	Number of lines to show (default 50)

mentu trace

Perfetto trace management (execution timeline files).

mentu trace <subcommand>

Subcommands: list, open, clean, test

mentu rewind

Restore files to the state before a step ran.

mentu rewind <target> [flags]

Flag	Type	Description
`--session`	string	Session ID to rewind from
`--dry-run`	bool	Show what would change without applying
`--force`	bool	Skip confirmation
`--backfill-cir`	bool	Backfill CIR signals after rewind

mentu watch

Watch recipe directories for file changes and trigger runs.

mentu watch

mentu setup

Local LLM setup (Ollama model installation).

mentu setup [flags]

Flag	Type	Description
`--model`	string	Ollama model to install
`--skip-install`	bool	Skip the install step

mentu preset

Agent preset management (prompt + model + tools templates).

mentu preset <subcommand>

Subcommands: list, get, set, create, delete

mentu device

Device presence tracking for multi-device workflows.

mentu device <subcommand>

Subcommands: id, status, list

mentu agent start

Start an agent session from a preset.

mentu agent start <presetId> [flags]

Flag	Type	Description
`--prompt`	string	Initial prompt
`--workspace`	path	Workspace directory
`--model`	string	Override model
`--background`	bool	Run in the background

mentu workspace

Workspace management.

mentu workspace <subcommand>

Subcommands: create, list

mentu profile

Workspace profile management (per-workspace settings).

mentu profile <subcommand>

Subcommands: show, list, add, refresh

mentu reasoning

Reasoning tier presets (effort levels for models).

mentu reasoning <subcommand>

Subcommands: list, show

AI Training

Commands for LoRA fine-tuning and hyperparameter search.

mentu lora

LoRA fine-tuning lifecycle.

mentu lora <subcommand>

Subcommands: list, register, status, train, eval, extract-sft

Subcommand	Description
`list`	List registered LoRA adapters
`register`	Register a new adapter
`status`	Show training status
`train`	Kick off a training run
`eval`	Evaluate a trained adapter
`extract-sft`	Extract SFT training data from CIR

mentu autoresearch

Autonomous hyperparameter search.

mentu autoresearch <subcommand>

Subcommands: run, list, status, history, add

Exit codes

Code	Meaning
`0`	Success
`1`	General error
`2`	Invalid arguments
`3`	Recipe validation failed
`4`	Step failed (agent error or non-zero exit)
`5`	Canary check failed
`6`	Auth required
`7`	Resource locked (workspace in use)