jarv 0.37.4

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

Jarv can be installed as a standalone binary or as a Python package. After installing, run jarv /setup to choose a provider, enter an API key, and pick a model.

# WinGet package in progress:
# winget install JamesWHomer.Jarv
irm https://github.com/JamesWHomer/jarv/releases/latest/download/install.ps1 | iex

# Homebrew tap in progress:
# brew install JamesWHomer/tap/jarv
curl -fsSL https://github.com/JamesWHomer/jarv/releases/latest/download/install.sh | sh
uv tool install jarv
pipx install jarv
pip install jarv

Python package installs require Python 3.10+. Keys can also be set via provider-specific environment variables or jarv /set.

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, Esc or 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, Esc or Ctrl+C clears existing text; press either 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 with a value supported by the selected model
--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 uses a multi-provider tool-calling agent loop (OpenAI Responses API, Anthropic Messages, Gemini, and OpenAI-compatible endpoints). The root model can call five tools:

Tool	Purpose
run_command	Execute a shell command and return stdout, stderr, and exit code
web_search	Search the web through DuckDuckGo's public HTML endpoint
read	Page through command output, artifacts, URLs, or local files
spawn	Fan out work to parallel subagents, each with their own tool access
ask_user	Ask you a question and wait for a reply

Each tool can be enabled or disabled from jarv /settings. Disabled tools are not sent to the model and are also unavailable to spawned subagents. The subagent-only finish tool remains enabled so child agents can always return their result.

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

run_command accepts optional head_chars and tail_chars parameters that control how much of the beginning and end of command output is returned to the model. Omitted values split max_tool_output_chars between the two sides. Explicit values override that configured limit for the command. When output is longer than the requested head and tail, Jarv retains the full result under a session-scoped cmd_<id> and reports the exact omitted offset and size so the model can retrieve it with read.

In the terminal, command output uses at most one-third of the screen height. Truncated output is biased roughly 2:1 toward the first lines, followed by the omitted-middle count and the final lines. Jarv also shows the resolved head_chars and tail_chars for each command.

read(input, offset, size) uses Unicode character offsets for text. offset defaults to 0 and size defaults to max_tool_output_chars; an explicit size is returned without generic tool-output truncation. Inputs resolve as retained command IDs, visible artifact labels, HTTP(S) URLs, or local file paths. Relative paths use the current working directory. Local and HTTP(S) PDFs with embedded text are extracted with page markers; scanned/image-only PDFs are not OCR'd. Consecutive reads requested in one model response run concurrently, with results returned in call order.

When read sees a direct local or HTTP(S) image (png, jpeg, webp, and provider-supported gif) and the active model advertises image input support in Jarv's cached provider/OpenRouter catalog, Jarv returns it as native multimodal tool output for that provider. Image reads ignore offset and size, are capped at 10 MiB, and are not resized or transcoded. If the active model route does not advertise image capability, Jarv returns a text result telling the model that image reads are unavailable instead of sending base64 text.

Web search and reads require no extra API key. web_search accepts any positive max_results plus a non-negative offset, following DuckDuckGo result pages as needed. URL reads preserve HTTP(S) hyperlinks as absolute URLs in extracted text, do not execute JavaScript, cap text and PDF responses at 2 MiB, and mark every returned page as untrusted content. Public, private, and localhost HTTP(S) addresses are supported, including custom ports.

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.

• 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 tool calls and returned outputs.
• Session-scoped — artifacts persist for the active session and are available on later prompts until you start a new session or archive.

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 sidecars
/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 for the active install channel

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.

Agent tool calls have a separate tool_call_display setting. auto uses print for one-shot runs and fullscreen in heads-up mode. print is resize-safe and left-aligned; fullscreen uses bordered cards with right-aligned status.

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
provider	"openai"	API provider (openai, anthropic, gemini, openrouter, etc.).
api_key	""	Legacy single API key field (migrated to api_keys).
api_keys	{}	Per-provider API keys. Falls back to provider env vars when empty.
base_url	""	Custom API base URL. Overrides the provider default.
model	"gpt-5.4-mini"	Model name passed to the API.
service_tiers	{}	Per-provider processing tier: standard, flex, or priority. Missing providers use standard; unsupported tiers are not offered.
reasoning_effort	""	Model-supported reasoning effort. Empty uses the provider/model default; none explicitly disables reasoning only where supported.
max_history	40	Max stored history items sent as model context (item cap before token trimming). Does not delete saved history.
context_budget_ratio	0.75	Share of the context window used for input.
context_compaction_threshold	0.85	Fill ratio that triggers history compaction.
context_output_reserve_ratio	0.15	Context window share reserved for model output.
context_window_fallback	128000	Context window when model metadata is unknown.
max_stdin_chars	200000	Maximum piped stdin characters attached to a one-shot prompt.
max_tool_output_chars	20000	Maximum generic tool output characters returned to the model. It also supplies the default head/tail budget for run_command.
disabled_tools	[]	Tool names omitted from root agents and subagents. Configure these from the Tools section in /settings.
command_timeout	60	Seconds before a shell command is killed.
web_timeout	15	Seconds before a web search or URL read 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; PyPI for Python installs, GitHub Releases for standalone installs).
read_only_command_display	"fullscreen"	Display mode for /help, /about, /usage, and /config: temporary fullscreen view or permanent print output.
tool_call_display	"auto"	Tool-call layout: auto selects print for one-shot runs and fullscreen in heads-up mode; explicit modes override it.
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.

Processing tier choices depend on the active provider. OpenAI, OpenRouter, and Gemini offer all three choices where the selected model supports them. Anthropic offers Standard and Priority; its Priority mode uses committed Priority capacity when available and otherwise falls back to Standard. Other providers remain on Standard.

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
│   ├── reads-<hash>.json            # retained command outputs
│   ├── 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. Cost tracking is request-based and grouped by provider and processing tier. Provider-reported cost is used when available; otherwise Jarv matches the selected model to pricing from OpenRouter's public model catalog and labels the calculation as estimated. The OpenRouter catalog is refreshed when provider model choices are refreshed, and its input, cached-input, and output rates are shown with model choices. Unknown and contract-priced requests are shown separately instead of being priced as Standard.

Dependencies

Package	Role
httpx	Direct provider API transports
pypdf	Lazy-loaded embedded-text extraction for PDF reads
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.