Getting Started

This chapter gets orangu running against a local orangu-server, from nothing installed to a working prompt. After installing, it is five commands in this order:

orangu-server -i
orangu-server download unsloth/gemma-4-E2B-it-GGUF
orangu-server --all unsloth/gemma-4-E2B-it-GGUF
orangu -i
orangu

The order matters: steps 2 and 3 both read the models directory from the configuration step 1 writes, and step 4 asks step 3’s running server which model it is serving.

Install orangu

The quickest way to install the latest release is the one-liner installer. It installs the whole stack — orangu, orangu-coordinator, orangu-server, the benchmarking tool orangu-bench, and the model builder orangu-gguf.

Linux / macOS:

curl -fsSL https://mnemosyne-systems.github.io/orangu/install.sh | sh

Windows (Command Prompt):

curl -fsSL https://mnemosyne-systems.github.io/orangu/install.cmd -o install.cmd && install.cmd

Windows (PowerShell alternative):

Invoke-WebRequest -Uri https://mnemosyne-systems.github.io/orangu/install.cmd -OutFile install.cmd; .\install.cmd

The script installs to ~/.local/bin (Linux/macOS) or %USERPROFILE%\.local\bin (Windows) and warns if the directory is not in your PATH. See BUILDING.md for instructions on building from source.

Configure orangu-server

orangu-server -i (long form --init) writes ~/.orangu/orangu-server.conf. Run it before anything else: downloading and serving both read the models directory from that file, and with no configuration present they stop with Missing config file; pass --config or add ./orangu-server.conf or ~/.orangu/orangu-server.conf (see --init).

orangu-server -i

The wizard asks for every [orangu-server] key. Each prompt shows its default in brackets; Enter accepts it. The language-model keys (context to reasoning_effort, prefill_backend, mlp_unroll) are not asked for a picture model, which is asked its own keys instead.

The wizard prints the file before writing it and asks Write this configuration? [Y/n]; anything but Enter/y/yes aborts with nothing written. Only non-defaults are written, so a minimal run yields:

[orangu-server]
models = /home/you/.cache/huggingface/hub
host = all
port = 8100

[web]
host = all
port = 8101

-i always targets ~/.orangu/orangu-server.conf and ignores -c/--config. See the Inference server chapter for the full configuration reference.

Download a model

orangu-server download unsloth/gemma-4-E2B-it-GGUF

The argument is a Hugging Face repo, <user>/<model>[:quant]. With no :quant, Q4_K_M is preferred, then Q8_0, then the first GGUF file in the repo; append one to pin it, e.g. unsloth/gemma-4-E2B-it-GGUF:Q8_0. Every shard of a multi-part model is fetched, plus the matching mmproj-* sidecar when the repo has one.

Before anything is fetched, the repo is planned against this machine: each shard’s GGUF header is read over the network — a few hundred kilobytes rather than the model — and the result reports how much of the model has to stay resident and how much can stream from disk. A model that cannot run here at all stops to confirm first; -y skips that. orangu-server plan gives the same report for a model already downloaded.

Files land under the models directory from step 1 in Hugging Face hub-cache layout. Free space is checked up front, an interrupted transfer resumes where it stopped, and progress is redrawn in place — one line per file plus a total — ending with:

Downloaded to /home/you/.cache/huggingface/hub/models--unsloth--gemma-4-E2B-it-GGUF/snapshots/<commit>/gemma-4-E2B-it-Q4_K_M.gguf

Gated or private repos need HF_TOKEN set in the environment. orangu-server list shows what is installed, delete removes a model, and refresh re-fetches it at a newer revision; see the Inference server chapter.

Start orangu-server

orangu-server --all unsloth/gemma-4-E2B-it-GGUF

The model is the same spec that was downloaded, and now resolves to the local copy — nothing is fetched a second time. A local .gguf path or an NR/MODEL label from orangu-server list works just as well, and omitting the argument prints the installed models and asks which one to serve.

--all picks the role, not a set of models: it is the general-purpose one, and the default. The alternatives are --code, --review (suppresses reasoning), --explorer (broader, more varied output), and --embedding (embeddings only, with the chat and completion endpoints disabled). They are mutually exclusive. A sixth, --image, is not picked: a Qwen-Image (qwen_image or qwen_image_2_1) picture generator always comes up in it, and no other model can.

The server binds the host and port from step 1 and serves an OpenAI-compatible endpoint, which is what orangu connects to and what step 4 asks for:

http://localhost:8100/v1

-d/--daemon detaches it from the terminal; that needs [orangu-server].model set in the configuration, because there is no terminal left to prompt on — and log_type = file to keep its output, since detached it has no console to log to either (the file defaults to orangu-server.log in the directory it was started from). See the Inference server chapter for GPU backend selection and the rest of the operational detail.

To move a working server to another machine without repeating steps 1–3, bundle it and its model into one executable:

orangu-server bundle unsloth/gemma-4-E2B-it-GGUF:Q4_K_M --all -y

Copy the resulting orangu-server-bundle-<arch> over, chmod +x it, and run it: no models directory, no download, and no orangu-server.conf — the API comes up on 127.0.0.1:8100 and the web console on 127.0.0.1:8200. See Bundling in the Inference server chapter.

Configure orangu

With the server from step 3 running:

orangu -i

The first prompt is Endpoint: press Enter to accept the default http://localhost:8100 (http://localhost:8100/v1 is accepted too). The wizard queries that server’s /v1/models and offers the first model it advertises as the default for the Model prompt — if the server cannot be reached the prompt has no default and you type the name yourself, which is why it is worth starting the server first.

It then walks every remaining option showing its default, so a full run is a row of Enters, and omits from the file anything left at its default. It finishes by reporting which optional tools it detected (git lg, delta, bat, gh, glab), printing the configuration, and asking Write this configuration? (Yes/No) [Yes]. On confirmation it writes ~/.orangu/orangu.conf — ignoring -c/--config, like the server wizard — and installs any bundled skills into ~/.orangu/skills/ that are not already present:

[orangu]
server = main-server
model = unsloth/gemma-4-E2B-it-GGUF

[main-server]
endpoint = http://localhost:8100/v1

See Configuration for every key the wizard can set.

Or copy the sample instead:

cp doc/etc/orangu.conf ./orangu.conf

Default configuration lookup order is:

1. ./orangu.conf
2. ~/.orangu/orangu.conf

Run orangu

~/.orangu/orangu.conf is one of the lookup locations, so there is nothing left to pass:

orangu

Or point at a specific file:

orangu --config ./orangu.conf

Then start with:

/help
/skills
/server
/disconnect
/reload
/tools
/model
/session <UUID>
/list_files
/open_file README.md
/show_file README.md
/debugging reproduce the failing request path
/build
/shell ls -la
/create_file README.md
/auto_review
/create_patch
/amend "[#42] My feature"
/branch main
/branch -b feature/new
/branch -m new-name
/branch -d feature/old
/cherry_pick abc1234
/revert abc1234
/comment 51 "My comment"
/close -i 51
/issue reviewer 114 jesperpedersen
/get_comments -i 51
/commit "[#42] My feature"
/restore README.md
/diff
/add_repository Jubilee101 muse
/init_repo
/log
/log 5
/show
/show aafd1cb
/merge feature/foo
/move_file old.rs new.rs
/pull 42
/push
/push --force
/rebase
/delete_file README.md
/review
/squash
/status
/usage
/clear
/copy
/quit

Most of these are thin wrappers around the matching git/gh commands. Two open full-screen views instead: /review walks you through the branch’s diff file by file for a manual review, and /auto_review has the connected model review the branch’s changes by itself — per file and per category (Code, Security, Memory, Performance, Test Suite, Documentation) — lets you override its verdicts afterwards (approve a file, or reject it with your own categorized comments), and copies the resulting report to the clipboard on exit. /auto_review <file> (Tab-completes on the file name) reviews a single file — the whole file on main/master, or just its changes on a branch — /auto_review src/main/java/** reviews every file matching a glob pattern the same way, and /auto_review all reviews every Git-tracked file in the project. All three are described in detail in the Core tools chapter.

orangu also supports Agent Skills: reusable directories containing a SKILL.md. Skills are discovered from ~/.orangu/skills/, ~/.agents/skills/, <workspace>/.orangu/skills/, and <workspace>/.agents/skills/. Use /skills to list them. Invoke one directly with /skill-name, for example:

/debugging reproduce the failing request path and identify the root cause

Review your first branch

The review workflow is orangu’s standout feature, so it is worth trying right away. From a feature branch with some changes (committed or just edited in the working tree):

review

This opens the interactive reviewer: a two-pane view with your changed files on the right and the selected file’s diff on the left. Use Alt+j/Alt+k to move between files, Alt+a/Alt+r to approve or reject one, and Alt+c to leave a categorized comment on a line. Type a question such as is this thread-safe? and press Enter to ask the model about the selected file. Press Alt+x to leave; the report is copied to your clipboard. (On a Mac these are Option-key chords, and the terminal must send Option as Meta — see “Alt on macOS” in the Terminal section.)

To have the model do the work, run:

auto review

orangu reviews the whole change and each file across the Overall, Code, Security, Memory, Performance, Test Suite, and Documentation categories, marks every file with a green or red dot, and ends with an orangu approves/rejects this patch verdict. When the run finishes you can override any verdict and remove findings before the report lands on the clipboard.

The branch must be rebased up to date first — if it is behind, orangu points you at /rebase. If you review with a thinking model and the answers look truncated, raise review_max_tokens in [orangu] (e.g. 2048); see the Configuration chapter.

Neither reviewer touches your code. To act on what it found, run:

create patch

The model gets the findings of whichever review ran last, checks each one against the current code, fixes the valid ones, and reports anything it deliberately left alone. It edits the working tree and never commits, so /diff shows you exactly what it did. The same command resolves the conflicts of an in-progress merge, rebase, or cherry-pick, with or without a review to go with them.

Share the result without leaving the terminal:

export review              # write the report to a PDF in the workspace root
comment on 42 with auto review  # post it on GitHub/GitLab issue/PR #42

By default the tools operate on the current directory. Use --workspace /path/to/project to point orangu at another tree.

Most startup flags have a short form: -c for --config, -t for --theme, -w for --workspace, -r for --resume (a UUID, or with none a list of stored sessions to pick from), -a for --all (reopen the tabs from the previous run), -l for --list (print every stored session as a table and exit), -i for --init, -p for --prompt (run one prompt or command and exit), -q for --quiet (print nothing on success), and -s for --shell-completions. --developer (the default) and --committer pick the prompt mode the interface opens in, as /developer and /committer do. -h/--help and -V/--version are available on all four binaries. orangu --help lists them all.

orangu automatically resumes an existing session when you return to the same workspace and Git branch. When a previous session is found, the status bar shows:

Resuming session 550e8400-e29b-41d4-a716-446655440000

for five seconds or until the first command is run.

On exit, the resume command is printed so you can return to the session from a different branch or machine:

orangu --resume 550e8400-e29b-41d4-a716-446655440000

Sessions that had no LLM interaction on main, master, or outside a Git repository are deleted automatically on exit. Feature branch sessions are always kept.

Use /session to list all sessions and their branches. Use /session <uuid> (Tab completion cycles UUIDs and workspace paths) to switch to a specific session; passing a workspace switches straight to it when it matches exactly one session, otherwise it lists the matches. Passing a directory path that no session uses yet opens it as a new workspace — Tab falls back to filesystem completion (with ~ expansion) so you can navigate there, e.g. /session ~/Po<Tab>/pga<Tab>/of<Tab>.

Lines whose first non-whitespace character is # stay local and are not sent to the model. Lines whose first non-whitespace character is \ are ignored.