jarv 0.19.0

A multi-provider AI-powered CLI agent

jarv

A multi-provider AI-powered CLI agent that can run shell commands, fan out work to parallel subagents, and keep track of conversation history across terminal sessions.

jarv                                    # start an interactive session
jarv whats the meaning of life?         # one-shot question
jarv commit all these files             # let it run commands to do the job
jarv refactor the auth module           # complex tasks get split across subagents

Install

Requires Python 3.10+ and an OpenAI API key.

pip install jarv
jarv /setup

The setup wizard will walk you through entering your API key and choosing a model. The key can also be set via the OPENAI_API_KEY environment variable or jarv /set api_key ....

To upgrade:

jarv /update

Usage

One-shot mode

Pass a prompt as arguments. Jarv answers (running commands if needed) and exits.

jarv what process is using port 8080?
jarv find all TODO comments in src/

Jarv also accepts piped stdin as one-shot input. If you pass both stdin and prompt arguments, the arguments are treated as the instruction and stdin is attached as context.

git diff | jarv review this patch
cat README.md | jarv summarize this
rg TODO . | jarv group these by subsystem

Heads-up mode

Run jarv with no arguments to enter an interactive prompt loop.

jarv> what files changed today?
jarv> now run the tests
jarv> /history
jarv> /new

• Type a prompt and press Enter.
• Slash commands start with / — type /help to list them.
• During a response, Ctrl+C stops further work, checkpoints the turn in history/context, and restores the prompt for editing. Use /undo to remove the turn.
• At the prompt, Ctrl+C clears existing text; press it again on an empty prompt to exit.
• You can also exit with exit, quit, or /exit.

Flags

Flags override config values for a single run and work in both one-shot and heads-up mode.

Flag	Short	Description
--provider PROVIDER		Override the provider (openai, anthropic, gemini, etc.)
--model MODEL	-m	Override the model (e.g. gpt-5.4-mini)
--effort EFFORT	-e	Override reasoning effort (low / medium / high)
--timeout SECONDS		Override command timeout in seconds
--system PROMPT	-s	Override the system prompt
--new		Start a fresh session (ignore prior history, but still save)
--incognito		Don't load or save session history
--version		Print the version and exit

jarv --provider anthropic -m claude-sonnet-4-6 "summarise this repo"
jarv -m gpt-5.4-mini "summarise this repo"
jarv --effort high "refactor the auth module"
jarv --new "start fresh without prior context"
jarv --incognito "one-off task, leave no trace"
jarv --timeout 120 --system "You are a poet" "write me a haiku"

How it works

Jarv wraps the OpenAI Responses API with a tool-calling agent loop. The model can call three tools:

Tool	Purpose
run_command	Execute a shell command and return stdout, stderr, and exit code
spawn	Fan out work to parallel subagents, each with their own tool access
read_artifact	Retrieve the full output of a completed subagent

On Windows, commands run through PowerShell. On other platforms, they run through the system shell.

Command safety

Before executing a shell command, jarv can prompt you for confirmation. The command_safety config key controls this:

Level	Behavior
risky (default)	Prompts for confirmation when a command matches dangerous patterns — recursive deletion, privilege escalation, network exfiltration, disk formatting, credential access, force pushes, and more.
all	Every command requires your explicit approval before running.
none	Commands run immediately with no confirmation prompt.

Set the level in the settings menu (jarv /settings) or at any time with:

jarv /set command_safety risky    # default — confirm dangerous commands
jarv /set command_safety all      # confirm everything
jarv /set command_safety none     # no prompts

Subagent orchestration

When the model calls spawn, Jarv runs N child agents in parallel. Each child operates independently — running commands, reasoning through subtasks — and terminates by calling finish with a detailed report and a short summary. The parent agent can then read any child's full output via read_artifact.

• Parallel by default — all children in a spawn call run concurrently in a thread pool.
• Artifacts — each child's output is stored as a named artifact. The parent (or siblings that declare a dependency) can fetch the full content.
• Recursive — children can themselves spawn further children, up to max_subagent_depth levels deep (default 4). Children are sterile by default; the parent must explicitly allow further spawning.
• Transcript scope — child-agent transcripts are discarded. Root history stores the parent spawn/read_artifact tool calls and returned outputs.
• Scoped per query — the artifact store resets with each new top-level prompt.

The terminal shows a live progress panel as children run, with a green checkmark or red cross as each finishes.

Commands

Command	Description
/help	Show all commands
/about	Detailed info and examples
/set <key> <value>	Set a config value
/unset <key>	Reset a config key to default
/settings	Open the interactive settings menu
/config	Show raw config values
/setup	Run the setup wizard
/new	Start a fresh session on the next prompt
/archive	Archive session history and artifacts
/sessions	Browse sessions (interactive when in a TTY)
/sessions <id>	Load a specific session by ID prefix
/history	Show recent conversation history
/undo [n]	Remove last n exchanges (default 1)
/redo [n]	Restore last n undone exchanges (default 1)
/usage	Show token usage, cost, and context breakdown for the current session
/usage day / /usage week / /usage month	Show system-wide usage for the last 24h, 7d, or 30d
/usage --all [--since 24h]	Show system-wide usage across Jarv sessions
/update	Update Jarv to the latest version

All commands work both as jarv /command (one-shot) and inside heads-up mode. Read-only commands (/help, /about, /usage, and /config) use a temporary display by default in interactive terminals; change read_only_command_display in /settings to print them permanently instead.

Sessions

Each terminal is automatically bound to its own session. Jarv identifies terminals using environment variables (WT_SESSION, TERM_SESSION_ID, TMUX, STY) with a parent-process fallback, so history persists across runs in the same terminal.

• /new starts a fresh session on the next prompt without archiving the current session.
• /sessions opens an interactive browser (arrow keys to navigate, Enter to load, a to archive, d to delete, p to preview, Tab to switch views, Ctrl+F to search).
• /history opens an interactive transcript where Up/Down scroll and Left/Right jump to the previous or next chat/reply.
• /undo and /redo let you step through recent exchanges.

Config

Settings live in ~/.jarv/config.json (created on first run). Use /settings for the common controls, or edit the file directly with /set and /unset.

Key	Default	Description
api_key	""	OpenAI API key. Falls back to OPENAI_API_KEY env var if empty.
model	"gpt-5.4-mini"	Model name passed to the API.
reasoning_effort	""	Reasoning effort level. Leave empty to disable.
max_history	40	Number of recent stored history items included as model context. Does not delete saved history.
max_stdin_chars	200000	Maximum piped stdin characters attached to a one-shot prompt.
max_tool_output_chars	20000	Maximum tool output characters returned to the model. Longer output is returned with the middle omitted.
command_timeout	60	Seconds before a shell command is killed.
command_safety	"risky"	Command confirmation level: all (confirm every command), risky (confirm dangerous commands only), none (no confirmation).
audit	true	LLM auditor for flagged commands.
auditor_auto_approve	true	Let the auditor auto-approve commands it deems safe.
auditor_model	""	Auditor model. Empty uses the active model.
max_subagent_depth	4	Maximum nesting depth for spawned subagents.
subagent_thread_pool_max_workers	8	Max parallel subagents per spawn call.
check_updates	true	Background update check on startup (non-blocking, throttled to once per 24h).
read_only_command_display	"auto"	Display mode for /help, /about, /usage, and /config: auto, print, inline, or fullscreen.
print_usage_after_agent	false	Print a compact token usage line after each completed agent run.
system_prompt	"You are Jarv..."	System instructions sent with each request.

Local files

All state is stored in ~/.jarv/ (on Windows, %USERPROFILE%\.jarv\):

~/.jarv/
├── config.json                      # settings and optional API key
├── sessions.json                    # terminal → session mappings
├── sessions/
│   ├── history-<hash>.json          # conversation history
│   ├── artifacts-<hash>.json        # subagent artifacts
│   ├── usage-<hash>.json            # session token usage totals
│   └── redo-<hash>.json             # undo/redo stack
├── usage.json                       # future system-wide token usage ledger
└── archive/                         # archived sessions

max_history counts stored items, not exchanges or tokens. User messages, assistant messages, reasoning items, function calls, and function call outputs each count as one item.

System-wide usage tracking starts from the version that records ~/.jarv/usage.json; older session totals are not backfilled into time-window reports.

Dependencies

Package	Role
httpx	Direct provider API transports
rich	Terminal styling, live rendering, markdown

License

Elastic License 2.0 (ELv2) — free to use, modify, and redistribute. You may not offer jarv as a hosted/managed service.