CLI Reference

CLI Reference

← Wiki index

Flags

Verified from the argument parser in src/agent8088/cli.py:

usage: agent8088 [-h] [--version] [--full-auto]
                 [--mode {readonly,full-auto}]
                 [--uninstall] [--workspace] [--all] [--yes]
                 [--non-interactive] [--dry-run]
                 [--update] [--force] [--setup] [--model-setup] [--sandbox-setup] [--memory-setup]
                 [--libreoffice-setup]
                 [--gateway] [--gateway-setup] [--mcp-serve] [--mcp-http]
                 [--mcp-port PORT] [--mcp-host HOST]
                 [--web] [--web-port PORT] [--web-host HOST] [--web-dev]
                 [--prompt-file PATH]
                 [--logs [MODE]] [-n LIMIT] [--level LEVEL] [--subsystem SUBSYSTEM] [--json]

Flag	Purpose
-h, --help	Show help and exit
-V, --version	Show version and exit
--mode {readonly,full-auto}	Permission mode at startup. Plan mode is not settable here — start it with /plan
--full-auto	Start in full-auto (no per-action prompts)
--setup	Interactive config wizard, then exit
--model-setup	Configure a model provider profile
--sandbox-setup	Install the native sandbox runtime
--memory-setup	Install the mem0 memory backend deps and set memory_engine=mem0; switch back with /memory engine native
--libreoffice-setup	Install LibreOffice (via winget/brew/apt-get/dnf/pacman, whichever is available) for document conversion, legacy Office formats, and formula recalculation
--gateway	Run the messaging gateway instead of the REPL
--gateway-setup	Configure gateway channels, then exit
--mcp-serve	Run as an MCP server
--mcp-http	Use HTTP transport (with --mcp-serve)
--mcp-port PORT	MCP HTTP port (default 8931)
--mcp-host HOST	MCP bind host (default 127.0.0.1)
--web	Launch the web UI instead of the terminal REPL
--web-port PORT	Web UI server port (default 8180)
--web-host HOST	Web UI bind host (default 127.0.0.1); loopback only
--web-dev	Run the web backend without serving built frontend files; use with Vite
--prompt-file PATH	Run one headless turn using the entire UTF-8 file as the task, then exit. Implies full-auto: use only in a trusted, isolated environment.
--logs [MODE]	Print the operational log; pass follow to tail it in real time. Bare --logs prints the last --limit lines and exits
-n, --limit LIMIT	With --logs: number of lines to print (default 50)
--level LEVEL	With --logs: filter by level (DEBUG|INFO|WARNING|ERROR)
--subsystem SUBSYSTEM	With --logs: substring filter on subsystem name
--json	With --logs: emit raw JSONL instead of the human-formatted view
--update	Pull latest code + reinstall, then exit
--force	With --update: discard local changes in the install dir first
--uninstall	Remove the install dir, shim, PATH/config lines, and crontab/scheduled-task entries, then exit. Trace logs and the WhatsApp session dir are kept unless --workspace/--all is also passed
--workspace	With --uninstall: also remove trace logs and the WhatsApp session directory
--all	With --uninstall: shorthand for --workspace
--yes	With --uninstall: skip the confirmation prompt
--non-interactive	With --uninstall: never prompt; requires --yes
--dry-run	With --uninstall: print what would be removed, remove nothing

Run with no flags for the interactive REPL.

Web UI

The web UI is the browser interface to the same Agent8088 engine, sessions, tools, permissions, and configuration used by the CLI.

Production

Launch the server from the repository root (it builds the frontend when needed):

agent8088 --web

Open http://127.0.0.1:8180.

Use the web flags to change the server settings:

agent8088 --web --web-port 3000

The web UI binds to loopback only, and a non-loopback --web-host is refused. Every endpoint is unauthenticated and POST /api/tool/{name} runs any tool in the registry, so exposing the port is equivalent to handing out a shell. To reach it from another machine, forward the port over SSH:

ssh -N -L 8180:127.0.0.1:8180 <the-machine-running-agent8088>

Development

Run the backend and Vite together:

uv run agent8088 --web

Open http://127.0.0.1:5180. Vite proxies /api and /ws to the backend.

What --uninstall does and doesn't remove

By default --uninstall removes everything the installer created: the install directory (venv, bundled uv/Node, sandbox runtime), the agent8088 command shim, the PATH and AGENT8088_CONFIG lines it added to shell rc files (or the Windows user-environment PATH entries for the bundled Git/Node on Windows), any crontab entries or Windows Task Scheduler entries a cron_mode schedule registered, and the SearXNG Docker container from /search setup if one exists — it runs with --restart unless-stopped, so Docker itself would otherwise keep it running (and restart it on reboot) indefinitely, since deleting the install directory doesn't touch a running container.

Two things are deliberately not touched:

• Trace logs and the WhatsApp session directory — these are your data, not installer residue, so they're kept by default. Pass --workspace (or --all) to remove them too.
• A pre-existing shared Playwright browser cache (~/.cache/ms-playwright on Linux, ~/Library/Caches/ms-playwright on macOS, %LOCALAPPDATA%\ms-playwright on Windows) — other Playwright-based tools on the same machine can share that cache, so it's never auto-deleted. A fresh install now downloads Chromium into $AGENT8088_HOME/playwright-browsers instead, so this only applies to installs from before that change; uninstall prints the exact manual command to remove the shared cache yourself if one is found.

agent8088 --uninstall --dry-run              # preview only, nothing removed
agent8088 --uninstall --all --yes --non-interactive   # fully unattended, full removal

Slash commands

45 registered commands (COMMANDS in src/agent8088/cli.py). Prefix-matched, so /mo offers /mode, /model, /models.

Pasting a bare file path into the prompt — nothing else on the line — reads it immediately: images go to a vision model, documents are extracted to text. Unlike a tool call the model makes on its own, this works outside allowed_paths, because it's a path the user personally typed. The sensitive-file floor still applies unconditionally.

Session

Command	Does
/help	List all commands
/new [name]	Start a fresh session, optionally named
/sessions	List saved named sessions
/resume <name>	Reload a named session (restores skill state too)
/reset	Clear the current conversation
/history	Show conversation history
/compact [n]	Summarise older turns, keep the last n verbatim
/save <file>	Export conversation + trace to JSON (mode 0600)
/status	Model, mode, tools, skills, token usage
/usage	Token usage for this session
/exit	Quit

Model

Command	Does
/model <provider:model>	Switch provider + model
/model setup	Add or update a provider profile
/models [provider]	Fuzzy model picker, fetched live
/temp <float>	Sampling temperature
/maxturns <int>	Max agent turns per prompt (same as /limits max_turns)
/reasoning	Toggle reasoning display
/tool-selection [hybrid|full|auto]	Show or set how native tool schemas are narrowed per request — see Tool-calling compatibility
/raw <text>	One raw model call — content, reasoning, tool_calls
/fusion setup	Interactively pick a default panel and judge (checkbox pickers), saved to config.txt
/fusion <query>	Ask the panel in parallel; a blind judge picks the best answer. --panel p:m,.../--judge p:m flags override the saved config for one call

Tools and execution

Command	Does
/tools	List tools with mode, args, description
/capabilities	Full self-report: tools, MCP servers, skills, subagents, limits, active guardrails
/tool <name> <json>	Invoke one tool directly
/plan [task]	Enter plan mode: propose a plan, approve it, then it runs
/audit [on|off]	Show or change step verification; no argument reports the current setting and the last turn's cost
/image <path> [question]	Analyze an image with a vision-capable model
/agents	List sub-agent profiles, with source, model, and the active provider's models
/agents models	List every model the active provider offers
/agents new [name]	Create a custom sub-agent profile interactively
/agents edit <name>	Open a custom profile in $EDITOR
/agents delete <name>	Delete a custom profile (bundled ones are refused)
/agent <type> <task>	Run a sub-agent directly; the prompt stays on it until /quit
/skills	List skills; disable/enable <name>
/cli-anything [task]	Show integration status, or route a task through the CLI-Anything skill
/browser [status|close|reset|stop]	Show the headless browser session state, or close it and clear its temp files
/local [list|available <query>|pull <name>|remove <name>]	Bare form probes hardware and recommends local Ollama models; the CLI counterpart to the local_models-mode tools
/task [list|start <goal>|resume <id>|end <id>|output <id>]	Start, resume, or inspect a restart-safe model task backed by task_runtime.TaskStore. Bare or list shows stored tasks

Permissions and isolation

Command	Does
/mode [readonly|full-auto]	Show or switch permission mode. Use /plan to enter plan mode
/reset	Discard the conversation — asks first unless destructive_slash_confirm=0
/sandbox [auto|native|docker|setup]	Show, select or install isolation
/search [status|setup|stop|doctor|use <backend>]	Show, provision, or pin a web search backend

MCP

Command	Does
/mcp	Server status + discovered tools
/mcp reload	Reconnect after editing mcp.json — asks first unless mcp_reload_confirm=0
/mcp add <name> stdio <cmd> [args...] [--project]	Add a stdio server
/mcp add <name> http <url> [--project]	Add an HTTP server
/mcp remove <name> [--project]	Remove a server

Diagnostics

Command	Does
/limits	Show every limit: turns, budgets, write caps, sub-agent turns, tool timeouts
/limits <key> <value>	Change one — persists to config.txt
/limits subagent <name> <turns>	Per-profile sub-agent round cap
/limits tool <name> <seconds>	Per-tool timeout
/limits provider <name> <key> <value>	Per-provider token limit (context_window or max_completion_tokens)
/schedule list	Show all scheduled recurring tasks (displays: #, cron schedule, task text)
/schedule add <cron> <task>	Add a recurring task; cron is 5 fields (minute hour day month weekday), e.g. /schedule add "0 9 * * *" check disk space
/schedule remove <index-or-task>	Remove by the # from /schedule list, or by exact task text
/config	Active config + file path
/capabilities	What the agent can do and which guardrails are in force
/cost [on|off|<task_id>]	Local telemetry summary of model cost/usage; on/off toggles recording, persisted to config.txt
/review	Bare: list stored reviews. With arguments: launch review_code from the CLI — same tool the model calls, see review_code
/doctor [--fix]	Environment health check; --fix repairs a broken web-search install
/dump	Write a redacted diagnostic bundle to disk, for sharing in a bug report
/trace [on|off]	Toggle JSON trace capture
/verbose	Toggle verbose output

Memory

Command	Does
/memory	Status: count, embedder, store size, last call's cost
/memory engine [native|mem0]	Show the active memory engine, or switch it live — persists to config.txt, never migrates or deletes either store
/memory search <query>	Run the search and show each leg's rank (words vs meaning)
/memory add <text>	Store a fact by hand
/memory forget <id>	Delete one (the short id from /memory search works)
/memory notify off|on|verbose	Change what memory prints when it learns something
/memory test	Run one extraction call and show the raw reply, parsed facts, and elapsed time
/memory clear	Delete all memories, with confirmation
/memory off	Stop recalling and learning; keeps what is stored

See Memory for what each does and when to reach for them.

Gateway commands

Inside Slack / WhatsApp / Discord / Telegram / Email:

Command	Does
/new	Clear the current session
/stop	Cancel queued messages for this chat
/help	List available commands
/capabilities	Tools, MCP servers, skills, limits, active guardrails
/approve	Approve the pending action (add session to hold for the session)
/deny	Refuse it

Discord also offers ✅ / ❌ buttons with a fail-closed timeout.

Gateway commands count against gateway_rate_limit_per_min like any other message — otherwise /help would be a free channel for flooding the gateway.

Keyboard

Key	Does
ESC	Interrupt the current turn
Ctrl+C	Cancel input / exit
Tab	Complete a slash command

Environment variables

Var	Purpose
AGENT8088_CONFIG	Config file path (/nonexistent forces packaged defaults)
AGENT8088_HOME	Data directory
AGENT8088_CLI_ANYTHING_HOME	Optional override for the isolated CLI-Anything runtime directory
AGENT8088_PROVIDER	Active provider
AGENT8088_PERMISSION	Starting permission mode (readonly or full-auto; plan-only falls back to readonly)
AGENT8088_SANDBOX	Sandbox backend

Exit codes

Code	Meaning
0	Success
1	Error (or a failing verification script)
2	Invalid CLI arguments

---

Source of truth: docs/wiki/ in the main repository. Edits here are overwritten by the next sync.