Orbit — AI Research Harness for Galaxy

Orbit is an AI agent for biological (and not only) data analytics. It lets you have a conversation about your data, draft and approve analysis plans, route steps to Galaxy or run them locally, and watch everything accumulate in a notebook — the durable project record that persists across sessions.

What can it do? Here’s a few examples…

Installation

All installers are on the Releases page.

• macOS Apple Silicon (M1/M2/M3/M4) — Orbit-<version>-arm64.dmg
• Linux Debian / Ubuntu / Mint / Pop!_OS — orbit_<version>_amd64.deb
• Linux Fedora / RHEL / CentOS / openSUSE — orbit-<version>-1.x86_64.rpm
• Linux any distro — Orbit-linux-x64-<version>.zip (extract and run orbit)
• Windows — WSL2 + Ubuntu, then use the .deb (see below)

macOS

Download Orbit-<version>-arm64.dmg from the Releases page.

1. Double-click the DMG and drag Orbit to Applications.
2. Eject the DMG.

git clone --branch v0.1.1 --depth 1 https://github.com/galaxyproject/loom
cd loom/app && npm install && npm start

curl -L https://github.com/galaxyproject/loom/archive/refs/tags/v0.1.1.tar.gz | tar xz
cd loom-0.1.1/app && npm install && npm start

Linux

# Debian/Ubuntu
sudo dpkg -i orbit_<version>_amd64.deb
sudo apt-get install -f
orbit

# Fedora/RHEL
sudo rpm -i orbit-<version>-1.x86_64.rpm
orbit

Windows (via WSL2)

Native Windows builds are not yet available. Windows 11 users with WSL2 + WSLg can run the Linux .deb directly — WSLg provides native GUI support with no X server setup required.

From an elevated PowerShell, install WSL2 if needed:

wsl --install --web-download -d Ubuntu

Then inside the Ubuntu terminal, download the .deb from the Releases page and:

sudo dpkg -i orbit_<version>_amd64.deb
sudo apt-get install -f
orbit

Getting API keys

Orbit needs at least one LLM provider key, and optionally a Galaxy API key for routing analysis steps to Galaxy servers.

Anthropic (Claude)

1. Go to console.anthropic.com.
2. Sign in or create an account.
3. Navigate to API Keys in the left sidebar.
4. Click Create Key, give it a name, and copy the sk-ant-... string.

OpenAI

1. Go to platform.openai.com.
2. Sign in or create an account.
3. Open Settings → API keys.
4. Click Create new secret key, copy the sk-... string.

Google Gemini

1. Go to aistudio.google.com.
2. Sign in with a Google account.
3. Click Get API key → Create API key.
4. Copy the key string.

DeepSeek

1. Go to platform.deepseek.com.
2. Sign in or create an account.
3. Navigate to API Keys in the left sidebar.
4. Click Create API Key, give it a name, and copy the sk-... string.

DeepSeek V4 Pro and Flash are very cost-effective — roughly 10–20× cheaper than comparable Anthropic/OpenAI models.

Galaxy API key

1. Log in to your Galaxy server (e.g., usegalaxy.org).
2. Open User → Preferences → Manage API Key.
3. Click Create a new key and copy the key.

Entering credentials in Orbit

Open Preferences with Cmd/Ctrl+, or by clicking the Galaxy connection indicator in the footer.

In the Provider section, select your LLM provider (Anthropic, OpenAI, Google, DeepSeek, or other providers), paste the API key, and choose a default model.

In the Galaxy section, enter your Galaxy server URL and API key. The footer indicator turns green when the connection is confirmed.

Click Save. The agent restarts with the new configuration.

Selecting a model

Orbit uses a large language model (LLM) as its reasoning engine — the model is what reads your data descriptions, drafts analysis plans, decides how to route steps to Galaxy, and interprets results. Different models are suited to different parts of that workflow.

Two phases of work call for different models. Designing an analysis — figuring out what to do with a dataset, surveying the literature, or constructing a multi-step plan — benefits from the most capable model you can afford. Actually running the plan — invoking Galaxy tools, polling job status, writing results to the notebook — is mostly repetitive and a cheaper model handles it just as well. When you run /execute on an expensive model, Orbit surfaces a one-time reminder suggesting a switch to save cost.

If you are using Anthropic API keys, we recommend starting with Claude Sonnet 4.6 for most sessions: capable enough for demanding work, fast, and roughly 5× cheaper than Opus.

Anthropic (Claude)

OpenAI

Google Gemini

DeepSeek

Both DeepSeek V4 models support a thinking mode for extended reasoning and a standard mode for direct responses.

Other providers — Mistral, Groq, xAI/Grok, and local Ollama models (Qwen3) are available in Preferences → Provider. Ollama models run entirely on your machine at no API cost, but need to be installed and configured first. Orbit can help you do that, however you need a computer with a substantial GPU to take advantage of local inference (see Nekrutenko 2026).

Switching models

You can switch models at any point during a session in three ways:

1. Footer model button — click the model name shown in the footer bar at the bottom of the window. This opens Preferences directly on the provider/model selector.

2. Preferences — open with Cmd/Ctrl+, and choose a different provider or model from the dropdowns, then click Save.

3. Slash command in chat — type /model <name> and press Enter. Orbit restarts the agent with the new model immediately. Use the exact model ID (e.g. /model claude-sonnet-4-6, not /model sonnet).

Switching mid-session is intentional — start on a capable model to design the analysis, then drop to a cheaper one.

API keys vs. subscription

For most providers — Anthropic, Google, DeepSeek, Mistral, and others — Orbit connects through the provider’s API. You pay per token, directly to the provider. Rates are shown in the tables above.

For OpenAI, Orbit also supports signing in with your existing ChatGPT subscription. Instead of an API key, you authenticate through a browser sign-in flow (OAuth), and the models run against your subscription rather than the pay-per-token API. This is generally cheaper if you already pay for ChatGPT Plus or Pro. The OpenAI subscription provider is listed as OpenAI (Codex) in Preferences; click Sign in with OpenAI to authenticate.

Direct subscription auth currently works only for OpenAI. We are exploring whether similar flows can be added for Anthropic (Claude.ai) and Google (Gemini Advanced) in a future release.

Interface overview

Orbit presents a three-pane layout:

Left — Files panel. A file tree for the current working directory. Click any file to open it in the File tab. At narrow window widths (below 900 px) the tree auto-collapses; restore it with Cmd/Ctrl+B or the toolbar button.

Center — Chat. The main conversation area. Type a message or a slash command and press Enter. Responses stream in with a thinking indicator; markdown tables and code blocks render inline. Prompt turns are numbered — /summarize 3 5 references those numbers. Press ↑/↓ to recall previous prompts.

Right — Artifact pane. Three tabs:

• Notebook — live render of notebook.md, auto-refreshed on every file change. This is the accumulating project log.
• Activity — split horizontally: agent shell stream on top (live tool calls, stdout from run_command), proc monitor on the bottom (CPU / memory / runtime for every subprocess the agent spawns).
• File — appears when you open a file from the left tree. Previews text (Markdown, code, JSON/YAML, FASTA/FASTQ/VCF/BED/GFF/SAM/Newick and more), images, and PDFs. Dismiss with ×.

Toggle the artifact pane with Cmd/Ctrl+\. The footer shows the Galaxy connection indicator (green = connected, red = no key) and live session cost/token totals.

Starting an analysis

Open Orbit in your analysis directory (Cmd/Ctrl+O to switch directories). Then just talk:

You: I have RNA-seq data from a drug treatment experiment — 6 samples,
     3 treated and 3 control HeLa cells. The data is at GEO accession GSE164073.

The agent responds conversationally. It can look up data, answer questions, check Galaxy’s workflow registry, and browse the tool catalog — none of this requires a formal plan.

When you ask for a plan, the agent drafts it in chat as a markdown section, then pauses. Four stages follow before anything lands in notebook.md:

1. Draft in chat. The agent writes a candidate plan. Not written to the notebook yet.
2. Plan approval. You review the steps and routing tags ([local], [hybrid], [remote]). Reply “looks good”, “go”, or request changes. The agent revises and asks again.
3. Parameter table in chat. The agent shows every configurable parameter per tool with defaults. You can adjust inline (“set min_qual to 30, leave others”).
4. Parameter approval. Once you approve, the agent writes the plan section into notebook.md and begins execution.

A plan section in the notebook looks like:

## Plan A: HeLa Drug Treatment RNA-seq DE [hybrid]

### Steps
- [ ] 1. **Quality + trimming** — fastp paired collection
  - Routing: Galaxy (fastp/0.23.4)
- [ ] 2. **Alignment** — HISAT2 to hg38
  - Routing: Galaxy (hisat2/2.2.1)
- [x] 3. **featureCounts**
  - Routing: Galaxy (featurecounts/2.0.3)
- [ ] 4. **DESeq2 differential expression**
  - Routing: Galaxy (deseq2/1.40.2)

Step status: - [ ] pending, - [x] verified complete, - [!] failed.

Come back the next day in the same directory and the session resumes automatically:

Orbit: Loaded notebook: HeLa Drug Treatment RNA-seq DE
       Plan A is in progress (1 of 4 steps complete).
       The fastp invocation finished successfully. HISAT2 alignment is queued.
       Should I check_all to advance, or do you want to review the QC report first?

Slash commands

Type / to open the autocomplete popup. Tab to complete; Enter submits.

The notebook

notebook.md in your working directory is the project record. Every plan section, executed step, parameter table, interpretation, and follow-up plan appends below the previous one. Multiple plans coexist over a project’s lifetime.

When Loom starts in a directory that is not already a git repo, it runs git init, drops a bioinformatics-friendly .gitignore (excluding FASTQ, BAM, VCF, and other large files), and sets git config loom.managed true. From that point on every notebook write triggers an auto-commit, giving you:

• Full undo history via git log.
• Timestamped, immutable record of every decision.
• Branch-based exploration: try an alternative analysis on a branch, compare with git diff.
• Collaboration: push to GitHub, collaborators pull and continue.

If you start Loom in an existing git repo, auto-commit is off by default. Opt in with git config loom.managed true.

Activity tab

The Activity tab provides two live views:

Shell stream (top half). Every tool call the agent makes — run_command stdout, Galaxy API calls, file reads and writes — streams here in real time. Use it to follow what the agent is doing and catch errors early.

Proc monitor (bottom half). A live table of every subprocess the agent has spawned: process name, PID, CPU%, memory, and elapsed time. Long-running jobs (alignment, assembly) appear here so you can watch progress without leaving Orbit.

Galaxy connection and routing

When Galaxy credentials are present, Loom registers the Galaxy MCP server automatically. The agent then:

• Surveys the Intergalactic Workflow Commission (IWC) workflow registry for a full-pipeline match before drafting any plan.
• Queries Galaxy’s tool catalog for available tool versions.
• Tags each plan step with a routing decision: [local], [hybrid], or [remote].

A remote plan maps cleanly to a single Galaxy workflow invocation. A hybrid plan mixes local light steps with remote heavy compute. A local plan runs entirely on your machine. These are outcomes of the plan, not settings you configure.

In-flight Galaxy invocations are tracked directly in notebook.md as fenced YAML blocks:

invocation_id: abc123
galaxy_server_url: https://usegalaxy.org
status: in_progress

The /execute command (also /run) tells the agent to advance the next pending step. The polling tool scans the notebook for in-flight blocks, queries Galaxy, and updates status in place.

Skills

Loom can fetch operational know-how from curated GitHub repos. The default repo is galaxyproject/galaxy-skills, seeded automatically on first use. It covers collection manipulation, Galaxy MCP usage, workflow report templates, Nextflow-to-Galaxy conversion, and tool development.

Add additional skill repos in Preferences → Skills. Each entry is a URL under github.com/galaxyproject/ (arbitrary repos are restricted in alpha to prevent prompt-injection). Fetched skills cache locally for 24 hours to cover offline use.

Getting help

Questions, bug reports, and feature requests are welcome on the Galaxy Help forum:

help.galaxyproject.org/c/orbit-support/16

For bugs and crash reports, use the GitHub issue tracker.

Related resources

• Loom/Orbit GitHub — source code, releases, issue tracker
• galaxy-mcp — MCP server for the Galaxy API
• galaxy-skills — curated operational skills
• Galaxy Training Network (GTN) — tutorials for Galaxy analyses
• usegalaxy.org — main Galaxy server for obtaining an API key