Agentic Science with Galaxy
Get Orbit
Orbit · Loom · Galaxy MCP

Agentic Science
with Galaxy

Plan, refine, and execute best-practice analyses, run on Galaxy's tools and workflows — every step reproducible and transparent.

Get Orbit See how it works →
Open source · MIT · macOS · Linux · Windows · npm i -g @galaxyproject/loom
Orbit, mid-analysis — chat, plan, and a live notebook in one window.

Orbit & Loom

A co-scientist desktop app and CLI. Loom is the agent brain; Orbit is the electron shell you observe it from. Same brain, terminal or desktop.

Explore Orbit →

Galaxy MCP

A Model Context Protocol server that connects any agent — Claude Desktop, Orbit, anything — to a Galaxy instance to run real, reproducible compute.

Connect your agent →

Reproducible by construction

Decisions stop happening in disposable chat. Plans, parameters, and provenance are written down as they happen — inspectable markdown, git-tracked.

See the method →
See it in action

Exactly what you can do

Real analyses, start to finish. Ask in plain language; Orbit drafts a plan, routes the heavy steps to Galaxy, and writes the notebook as it goes. Pick a scenario — or watch them play.

Orbit ~/chrM-variants
Chat▤ ⤓ ▥
1
Call chrM variants across these 4 samples on Galaxy.

Here's the plan — QC locally, then align and call variants on Galaxy. Approve to run.

chrM Variant Callinghybrid
  • ✓QC & adapter trimlocal
  • ✓Reference indexlocal
  • ✓Read alignmentgalaxy · bwa-mem2/2.2.1
  • ✓Variant callinggalaxy · bcftools_call
✓All 4 samples aligned (mean depth 812×). 27 variants pass filters — 3 are non-synonymous in MT-ND5.
NotebookActivityFile

Plan A: chrM Variant Calling hybrid

Question: how do mtDNA variants distribute across these tissues?

Steps
  • ✓1. QC & adapter trimfastp
  • ✓2. Reference indexbwa-index
  • ✓3. Read alignmentGalaxy · bwa-mem2/2.2.1
  • ✓4. Variant callingGalaxy · bcftools_call
loom-invocationcompleted
BWA-MEM alignment (chrM, 4 samples)
jobs 4/4usegalaxy.org
ready Galaxy CloudLocal Opus 4.8 notebook.md · git ✓

Representative transcripts. The plan format, [galaxy]/[hybrid] routing tags, and loom-invocation tracking are real; the exact tools and results a run produces vary by model and analysis.

Orbit & Loom

Your Galaxy co-scientist

Agentic bioinformatics that produces useful work — without the familiar failure mode where decisions happen in chat, parameters drift, and a plausible final answer hides weak intermediate validation.

The notebook is the project

Plans, decisions, parameter tables, and results accumulate as plain markdown in one durable notebook.md — no hidden state store to drift out of sync.

Plans you approve before they run

The agent drafts in chat, waits for your sign-off, shows you every parameter, waits again — then writes to the notebook and executes.

Galaxy-aware routing

With Galaxy connected, each step is tagged local, hybrid, or remote based on whether an IWC workflow matches or the heavy tool lives on Galaxy.

Live invocation tracking

Running Galaxy workflows are recorded as loom-invocation blocks and polled to completion or failure — in place, with no external tracker.

Git-tracked by default

Loom git-inits the directory and auto-commits every notebook change — full undo history, timestamped reproducibility evidence, shareable via GitHub.

Bring your own model

Anthropic, OpenAI, Google, Mistral, Groq, xAI, DeepSeek, or local Ollama — with live token, cost, and context-window meters in the footer.

…and more: a live activity stream of every tool call and subprocess, on-demand Galaxy skills, and one /orbit command to pop a terminal session into the desktop view.

How it works

Chat → approve → run → recorded

The working record is made explicit while the analysis is happening — not reconstructed after the fact.

  1. 01
    Chat

    Ask in plain language, drop file paths, request data lookups. None of it requires a “plan” — the conversation is just a conversation.

  2. 02
    Approve

    When you ask for a plan, the agent drafts it in chat and waits. It shows you the parameters and waits again. Nothing runs without your sign-off.

  3. 03
    Run on Galaxy

    Steps route to Galaxy — IWC workflows, heavy tools — or run locally. Live invocation status streams back as jobs complete or fail.

  4. 04
    It’s recorded

    Every plan, parameter, and interpretation lands in a durable, git-tracked notebook.md. Come back tomorrow and the analysis is the project.

notebook.md 
## Plan A: chrM Variant Calling [hybrid]

Question: how do mtDNA variants distribute across tissues?

### Steps
- [x] 1. QC FASTQ {#plan-a-step-1} — fastp trim + per-base QC
  - Routing: local
- [ ] 3. Read alignment {#plan-a-step-3} — bwa mem, 4 samples
  - Routing: Galaxy (bwa-mem2/2.2.1)

```loom-invocation
invocation_id: f2db4a1c9e8b7654
galaxy_server_url: https://usegalaxy.org
notebook_anchor: plan-a-step-3
label: BWA-MEM alignment (chrM, 4 samples)
status: in_progress
completed_jobs: 1
total_jobs: 4
```

Galaxy histories remain the computational truth. notebook.md is the durable working record — Galaxy is the substrate, and Loom is one live authoring shell.

Grounded, not guessing

Built on Galaxy's tools, workflows, and contracts

Generic AI invents plausible-looking pipelines — fake tool names, wrong parameters, workflows that won't run. Orbit builds on Galaxy's contracts instead, so what the agent proposes is real, versioned, and reproducible.

Real, versioned tools

The agent works from Galaxy's actual catalog — thousands of containerized, version-pinned tools — through Galaxy MCP. What it runs is installed and reproducible, not a name it made up.

No tool for the job? It builds one.

When nothing in the catalog fits, the agent authors a User-Defined Tool right in Galaxy — guided by a curated skill — instead of dropping to ad-hoc local scripts. Even bespoke steps run through Galaxy's job system, so they stay reproducible and shareable.

Community workflows, not reinvented

When a curated IWC workflow fits the task, the agent finds and runs it instead of assembling a pipeline from scratch — decades of best-practice, tested methods as the default.

Reproducible by contract

Every step executes as a real Galaxy job or workflow invocation, with Galaxy's provenance, tool versions, and containers behind it — re-runnable and shareable in Galaxy, by construction.

Full agentic driving on top of the reproducible, transparent, web-accessible Galaxy platform — your analysis runs in the browser, on the Galaxy servers you and your collaborators already use. It's still Galaxy.

Galaxy MCP

Bring your own agent

Galaxy MCP is the bridge that lets a conversational agent do real compute on Galaxy instead of just talking about it. Point it at any Galaxy server and your agent can search and run tools, drive workflows, and manage histories through Galaxy’s API.

  • ✓Connect to any Galaxy instance — URL + API key, or browser OAuth for short-lived keys
  • ✓Search the tool catalog, view details, and execute tools
  • ✓Access and import workflows from the IWC
  • ✓Create, inspect, and manage histories & datasets
  • ✓Upload local files into Galaxy
  • ✓Ships stdio (local) and streamable-HTTP (remote) transports
Galaxy MCP on GitHub
claude_desktop_config.json 
{
  "mcpServers": {
    "galaxy-mcp": {
      "command": "uvx",
      "args": ["galaxy-mcp"],
      "env": {
        "GALAXY_URL": "https://usegalaxy.org",
        "GALAXY_API_KEY": "••••••••"
      }
    }
  }
}
Works with Claude Desktop, Orbit/Loom, and any MCP client.
Get started

Two minutes to your first agentic analysis

Desktop Orbit app

Native installers that bundle their own Node, uv, and Loom — macOS builds are signed & notarized. Just download and open.

Download for macOS · Linux · Windows From the GitHub Releases page.

Terminal Loom CLI

The same brain, in your shell. Needs Node 22.19+ and uv for Galaxy MCP.

npm install -g @galaxyproject/loom
loom
Or run it once with npx @galaxyproject/loom.

Any agent Galaxy MCP

Wire Galaxy into Claude Desktop or any MCP client with one command.

uvx galaxy-mcp
Reads GALAXY_URL / GALAXY_API_KEY from the environment.
FAQ

Questions, answered

Why two names — Orbit and Loom?
Loom is the brain; Orbit is the shell. Loom is the agent runtime — the system-prompt context, Galaxy routing and invocation tracking, the skills system, and the RPC contract — and you can run it straight from a terminal with loom. Orbit is the Electron desktop app that wraps that same brain in a chat + notebook + activity layout. Same brain, two ways to observe it; future shells (a Galaxy-embedded UI, a hosted mode) talk to it over the same RPC contract. The names trace the cosmology: the cosmic web weaves galaxies along filaments of dark matter — Loom weaves your research record the same way, and Orbit is the electron shell you observe it from.
Do I need a Galaxy account to try it?
No — Loom is useful for local exploration and planning before you connect anything. But Galaxy is the primary execution path: add a Galaxy URL and API key and the agent surveys the workflow registry and tool catalog while drafting plans, then routes the heavy steps to Galaxy and tracks each invocation in the notebook.
Which AI models can I use?
Bring your own. Anthropic, OpenAI, Google (Gemini), Mistral, Groq, xAI, DeepSeek, local Ollama, or any OpenAI-compatible endpoint — with live token, cost, and context-window meters in the footer. You supply the key; nothing is locked to one vendor.
How is this different from just chatting with an AI about my analysis?
The notebook is the project. Instead of decisions living in disposable chat, every plan, parameter table, and result is written to a durable, git-tracked notebook.md you can inspect, edit, and diff. Multi-step plans are drafted and approved — with the exact parameters shown — before anything runs.
Does this replace the Galaxy interface?
No — it's an additive way to work with Galaxy, not a replacement, and that's the point. Galaxy stays the substrate: every heavy step runs as a real Galaxy job or workflow invocation, so the analysis itself lands in your Galaxy history — fully reproducible, re-runnable, and shareable like any other Galaxy work. On top of that, Loom keeps a durable record of the reasoning — the plans, parameter choices, and interpretation — in notebook.md as it happens. So both halves are transparent and reproducible: the conversation and decisions in the notebook and the actual analysis in Galaxy. You can always open Galaxy and re-run exactly what the agent did — it's still Galaxy.
Won't the AI just hallucinate tools and parameters?
That's the failure mode Galaxy grounding is built to avoid. Instead of inventing a pipeline from text, the agent works through Galaxy MCP against your server's real tool catalog and the curated IWC workflow collection — so the tools it runs actually exist, are version-pinned, and are containerized. It drafts the plan with the exact tools and parameters for you to approve before anything runs, and the run itself is a real Galaxy job with full provenance. The output is grounded in Galaxy's contracts, not a plausible-looking guess.
What if the tool I need isn't in Galaxy?
It builds one. When no catalog tool fits the task, the agent authors a User-Defined Tool right in Galaxy — guided by a curated skill — rather than dropping to ad-hoc local scripts. The new tool runs through Galaxy's job system like any other, so even bespoke steps stay reproducible and shareable. Full agentic reach, still inside Galaxy.
Can I drive Galaxy from Claude Desktop or another agent — not just Orbit?
Yes. Galaxy MCP (uvx galaxy-mcp) is a standalone Model Context Protocol server — point any MCP-capable client (Claude Desktop, Orbit, others) at a Galaxy instance and it can search and run tools, drive workflows, and manage histories. Orbit and Loom use it under the hood.
Is it open source? What does it cost?
Loom, Orbit, and Galaxy MCP are open source (MIT). There's no license fee — you pay only for your own model usage (your API key) and whatever Galaxy compute your account uses.
What platforms does Orbit run on?
macOS (Apple Silicon; signed & notarized), Linux (.deb / .rpm / .zip), and Windows. The Windows build is remote-only (Galaxy execution, no local shell) and currently ships as a .zip while a signed installer is finalized. Prefer the terminal? The loom CLI runs anywhere Node 22.19+ does.
Is it ready for everyday use?
Orbit and Loom are in active development — think capable beta. The core loop (plan, route to Galaxy, track invocations, record the notebook) works today; end-to-end validation against live Galaxy servers is ongoing. Galaxy credentials are stored in your OS keychain in Orbit, and local execution is confined to your working directory by default.
Documentation

Go deeper

Getting started → Install, connect Galaxy, run your first plan. Concepts → The notebook model, plans, routing, invocations. Galaxy MCP → Connect any agent to a Galaxy instance. Architecture → Brain, shells, and the RPC contract.