Core workflows
Bodega One Code (CLI)
Bodega One Code is the terminal surface of Bodega One, the same engine, agent, and QEL verification you use here, driven from a shell. It shares this app's brain: provider keys, settings, sessions, and MCP servers configured in either place are visible to both. Install it once and bodega works in any terminal with zero prerequisites, the bundle carries its own runtime.
What it is
A native terminal client that spawns the same local backend this app runs. All reasoning, model selection, and QEL verification happen in that shared engine, the CLI adds an interactive session (with a first-run wizard), a headless mode whose exit code follows QEL verification (scripts and CI can trust it), parallel fleet runs, and full MCP + local-model management from the shell.
Because the app and the CLI share one data dir, there is nothing to re-configure: a provider key you saved here just works in the terminal, and a session started there shows up in your history here.
Install
Windows: Scoop is the recommended path (no SmartScreen prompt):
scoop install https://github.com/BodegaoneAI/bodegaone-cli-releases/releases/latest/download/bodega.json
Or the one-liner installer:
irm https://github.com/BodegaoneAI/bodegaone-cli-releases/releases/latest/download/install.ps1 | iex
macOS / Linux:
curl -fsSL https://github.com/BodegaoneAI/bodegaone-cli-releases/releases/latest/download/install.sh | sh
macOS bundles are signed and notarized. Manual downloads (per-platform archives + checksums) live on the release mirror; unpack anywhere and run bin/bodega. Open a new terminal after installing so PATH picks it up.
First run
- Run
bodega. On a machine that already runs this app, it finds your existing providers and settings through the shared data dir, no setup. On a fresh machine, a short wizard walks you through picking a provider. bodega doctorverifies the install end to end (runtime, backend, data dir, providers) with PASS/WARN/FAIL rows and a fix for anything red.- No API keys?
bodega modelsrecommends local models that fit your hardware, installs the runtime, and pulls one, fully offline coding after that.
Startup problems always print a plain-English cause and the command that fixes it.
The command surface
| Command | What it does |
|---|---|
bodega |
interactive session (wizard on first run) |
bodega run "<task>" [--agent <name-or-id>] [--max-time 45m] [--max-steps 150] [--max-cost 3] |
headless one-shot; exit code derived from QEL; --bg detaches and returns a task_id; --max-time / --max-steps set the run's hard ceilings (the engine extends a progressing run up to them and never asks headlessly); --max-cost ends the run with exit 7 once spend meets it |
bodega run "<task>" --subagents |
opt THIS run into sub-agents (off by default unattended); the two settings it flips are put back when the run ends, however it ends |
bodega agents [--json] [--all] |
list custom agents: provider, model, isolation when spawned, and whether the main agent may spawn each |
bodega runs |
list recent background runs (status, QEL score); runs cleanup removes orphaned logs |
bodega logs <id> |
stream NDJSON events from a background run |
bodega attach <id> |
block until a background run finishes, print the result, exit with its code |
bodega models |
local runtimes + models: status, pull, rm, install-runtime |
bodega mcp |
add / edit / remove MCP servers, import from other MCP clients' configs |
bodega plugin install <path> |
install a plugin (skills + MCP servers bundle) from a folder or .zip |
bodega skills |
list / show SKILL.md files; skills trust, skills approve <name>, skills revoke <name> manage per-project skill approval |
bodega harness log / harness revert <editId> |
inspect and revert continual-harness edits (see below) |
bodega refine "<instruction>" [--apply] |
propose (and optionally apply) a continual-harness edit from plain language |
bodega review [--full] |
AI code review of your uncommitted changes |
bodega serve --acp |
start an ACP stdio relay for editor integration; --check diagnoses |
bodega doctor [--json] |
diagnose the local environment (PASS/WARN/FAIL) |
bodega init [--reset] |
re-run the first-run onboarding wizard |
bodega reset |
clear this CLI's own local state (--onboarding / --settings / --all) |
bodega fleet "<task>" --split N |
N parallel isolated-worktree runs; QEL recommends the winner |
bodega loops |
scheduled agentic tasks, QEL-gated, in isolated worktrees; shared with the desktop app |
bodega session |
portable session export / import (redaction-gated) |
bodega config |
typed settings against the shared store |
bodega self-update |
non-destructive update check against the official channel |
bodega help <topic> |
in-depth topic docs in the terminal |
Parallel runs, session trees, asides and the scheduler
Available now: bodega run --parallel "A" "B" "C" runs one session each in its own worktree, bodega sessions --tree shows sub-agents and asides under their session, and bodega attach <session> follows any of them live. Since CLI v0.2.17 (needs Bodega engine v1.0.0-beta.47 or newer; an older engine says so): bodega btw <session> "question" [--on same|small_local|cloud|<preset>/<model>] asks a side question and prints the answer when it is done, and bodega scheduler shows what holds a GPU slot and what is waiting, and why (--json for scripts).
Environments from the terminal
Available since CLI v0.2.17 (needs Bodega engine v1.0.0-beta.47 or newer). bodega env list and bodega env show <name> list and show your environments. bodega env add / edit / rm manage them (rm asks first; --yes in scripts). bodega env start <name> starts a session in one, and bodega run --env <name> "task" also runs a task once setup is done. When an environment has a setup script, the CLI prints the whole script and asks run, skip or stop; nothing answers for you, and in a script the approval stays waiting until you pass --run-setup or --skip-setup, or answer it later with bodega env setup <session> --run|--skip|--stop. Keep the desktop app or bodega serve running while setup runs: an engine the CLI starts for one command stops when that command ends. Only local environments exist; --target ssh or hq is refused.
Headless and CI
bodega run is built for scripts: the exit code is derived from the QEL verification result, not from "the model finished", so a pipeline can gate on verified work. Tool access can be restricted per run with --deny-tools, and air-gap mode is enforced by the shared engine exactly as it is in the app, a CLI run cannot reach the network when this machine is air-gapped.
bodega harness log and bodega harness revert
Want to undo what the agent changed to your persona, project rules, a skill, a remembered note, or a saved knowledge card? That's bodega harness revert. The backend keeps an audit trail of every write made through the continual-harness system, bodega harness log reads it back.
bodega harness log [--domain <d>] [--project <path>] [--limit N]
<d> is one of persona, project_rules, skill, memory, knowledge. Rows show newest first.
bodega harness revert <editId>
Replays that edit's before-value through the same write path the original edit used, restricted to your own account.
Revert works for all five domains: skill, project_rules, persona, memory, and knowledge. Knowledge revert covers saved knowledge cards; a plain URL/text reference you added with query_knowledge doesn't have a stable id to revert onto, so that one still refuses with a plain message rather than pretending it worked.
Revert also refuses if a later edit exists on the same target (revert that one first) or if the file was modified outside the log since the edit you're reverting, either way, nothing is silently overwritten.
Staying current
bodega self-update checks the official release channel and tells you when a newer version is out, it never replaces the binary behind your back. Upgrade by re-running the installer, or let Scoop / Homebrew handle it (scoop update bodega, brew upgrade bodega).
This page mirrors the in-app docs hub for app version 1.0.0-beta.54.1. Found something unclear or out of date? Tell us on Discord. New here? Download the free beta and follow along.