VX: an AI-agent-controllable desktop genome and transcriptome browser with a programmable analysis framework

A fast, cross-platform, GPU-accelerated desktop genome and transcriptome browser, built from the ground up for AI-agent driven exploration of sequencing data. Visualise FASTA / GTF / GFF / BAM / BED / BigBed / bedGraph / BigWig / VCF / BEDPE interaction tracks; run 59 built-in computational analyses against them; drive the whole thing from Claude, an LLM agent, or a plain curl command through an embedded HTTP API.

---

What VX is for

VX is a desktop genome browser in the family of IGV, UCSC Genome Browser, JBrowse, JBrowse 2, SeqMonk, Tablet, Pileup.js, and Pretzel, built around a different set of needs. It targets four audiences that the current generation of browsers serves less well:

1. Researchers working with very large sequencing datasets (multi-GB BAM, BigWig, BEDPE contact lists, multi-sample VCF) on a laptop, who need a viewer that stays responsive while the reads of a region of a 23 GB BAM load in the background.
2. Bioinformaticians and computational biologists who want to run analyses against the data while looking at it. Run peak calling, differential expression, splice junctions, transcript quantification, base modification sites, poly(A) tail statistics, copy number, alignment QC, variant density, contact matrices and TAD boundaries, cross-track correlation, directly from the viewer, in the background, with file exports.
3. AI / agentic-LLM users who want a genome browser their agent can see and control: navigate the viewport, load files, capture tracks as images, run analyses and read their result files, edit an annotation, export regions, all through a structured tool catalogue.
4. Educators and figure-makers who want a reproducible, scriptable viewer that produces publication-grade output (PNG, JPEG, TIFF and SVG images, and MP4 / WebM / GIF recordings of the viewport through ffmpeg).

If you do NGS, RNA-seq, ChIP-seq, ATAC-seq, Hi-C, long-read sequencing, direct RNA sequencing, variant calling, comparative genomics, transcript isoform analysis, ribosome profiling, CLIP-seq, methylation sequencing, MeRIP-seq, or any other sequencing-based assay and you've ever wished your genome browser was faster, smarter, or scriptable, VX is built for you.

---

Why VX is different

Capability	What it means in practice
GPU-accelerated rendering	OpenGL drawing with levels of detail that follow the zoom: a whole chromosome (10⁸ bp) as feature density, then gene outlines, exons and, at the bottom, single bases (10⁰ bp).
Async, background-thread BAM loading	BAM regions are read on a worker thread. The view keeps panning and zooming over the reads already held while the next window loads, and a load-status strip at the bottom of the viewport tells you why the alignments aren't there yet.
80-tool MCP server + HTTP API	Built-in HTTP server on 127.0.0.1:9876 (91 commands through POST /command, plus GET endpoints for the state and for captures), and an MCP server of 80 tools over it. Claude Code, Claude Desktop or any other MCP client (through the MCP server), a Python script or curl can navigate, load files, query genes, capture images, run analyses and read their result files, edit the reference and its annotation, manage bookmarks and sessions, and record video.
59 built-in analyses	Peak calling, RPKM/TPM/CPM quantification, differential expression, splice junctions, transcript quantification, gene body coverage, ORF prediction, CNV segmentation, contact matrices, modification sites, poly(A) tail statistics, modification co-occurrence, GC content, motif search, signal correlation, variant annotation, coding impact, Ts/Tv, MAPQ distribution, insert sizes, strand bias, TAD boundaries, profile plots, gap detection and more, all in a background thread with file output (BED, BED12, bedGraph, TSV, bedMethyl, BEDPE) and a manifest of each run (its inputs with their SHA-256, the parameters as applied).
Reads in aggregate	Beside an alignment track, a modification density track (base modifications from the MM and ML tags, per site, window or bin, with a binomial or Fisher's test and Benjamini-Hochberg q), a poly(A) tail track (Dorado's tail lengths) and, new in 0.11.0, a variant summary (the bases of every read against the reference, with a binomial test or Fisher's test against a control library). Each can be drawn as a strip of the alignment track or as a track of its own, and its rows exported as TSV (a modification density's also as bedMethyl).
New views in 0.11.0	Sashimi tracks (depth and splice junction arcs) and overlay tracks (several signals in one lane); the Matrix display of an interaction track (contacts binned and balanced as cooler balances them); the calls matrix (a row a read, a column a modified site); a circular view of a mitochondrion, plasmid or bacterial chromosome with read chords; a dot plot of a region against itself, another region or a transcript.
Edit mode	Add, move, resize, rename and delete genes, transcripts, exons and features, set strands and biotypes, insert, delete or replace bases, with undo and redo; every coordinate of the tracks held in memory moves with a base edit. Write the edited reference (FASTA and .fai), annotation (GTF or GFF3) and features (BED).
Genomes and transcriptomes	Each reference has a group of its own, with a reference mode (auto, genome or transcriptome) that the overview, the analyses and the modifications' names follow. Every quantitative track has one value axis (linear, log2, log10, square root or asinh; fitted or a range you set). The Known Features pane shows what public services (Ensembl, UniProt, STRING, IntAct, PDB, NCBI Gene) say about a gene, transcript or protein you choose.
Built-in video / GIF recording	Capture the OpenGL viewport to MP4 (H.264), WebM (VP9) or GIF through ffmpeg (installed separately). Two modes: timer-paced (constant fps) or render-paced (every frame). Quality and frame rate user-controlled. For figure animations, talks, supplementary movies.
Subsampling + transparency	Alignment tracks honour a configurable read cap (default 500 reads in view) and a min_mapq filter, set per track in the track panel or through the API. The reads held are a sample of the whole window, the same at every visit, and the coverage strip counts every read of the window, not only the sample. When a view is subsampled, a badge on the track says so (kept/total). Captures and the SVG export render the reads the view shows.
Self-contained binaries	No Python runtime, no Conda environment, no web stack. One download per platform (9.9 MB for Linux, 24.7 MB for macOS, 61.7 MB for Windows): the macOS and Windows packages carry GTK and the libraries it uses; on Linux VX uses the distribution's GTK 3, FreeType and OpenGL driver. Unpack, or drag to Applications, and run.
Three platforms, pre-compiled	Linux x86-64 and Windows x86-64 (one build each, for x86-64-v3 processors), macOS on Apple silicon (VX.app in a .dmg). Built with LDC, the LLVM-based D compiler, with the D runtime and Phobos linked in statically; full LTO on Linux and Windows.
No account; usage statistics only if you opt in	VX is a desktop application with no account and no cloud service. Anonymous usage statistics are opt-in and off by default: VX asks at the first start (and, after a Not now, once more ten launches later), shows the exact fields, and sends nothing unless you agree. Otherwise VX goes on the network only for these: a demo dataset you open (this repository's demo-data-v1 release); a genome you retrieve (Retrieve Reference); UCSC's chromosome band table, when you load a reference read as a genome whose file or folder name names a known assembly (GRCh38, hg38, mm10 and others) and neither a cytoBand.txt beside it nor VX's cache holds the table (it is then kept in that cache); and a Known Features tab you open or refresh (the identifier you chose, and those the services' answers give, go to the services turned on in Settings; Use the cache only sends nothing). The HTTP API listens on 127.0.0.1 only and refuses requests from pages in a web browser. Your data stays on your machine.

If you've used IGV and wished it were faster, or JBrowse and wished it weren't a browser tab, or Pileup.js and wished it ran big files, or any genome browser and wished an AI agent could drive it, VX was written with those wishes in mind.

---

Screenshots

Chromosome scale: the whole of chromosome 17 (83 Mb), its genes drawn as a density histogram, with the sequence and an alignment track.	Gene level: 126 kb around BRCA1, each gene with its exons and introns, coloured by biotype.
Exon level: 8 kb of BRCA1, the translated part of each exon (the translon) at full height, the UTRs thinner.	Base level: 200 bp, each base a coloured bar.
Several tracks together: sequence, genes and alignments over 300 kb of chromosome 1.	Anatomy of a gene track: exons, UTRs, the translated part, introns, strand arrows.

The screenshots were taken with VX 0.9.0; the window of 0.11.0 differs in details, and the views new in 0.11.0 are not shown.

---

Quick install

1. Download the right binary for your platform

Visit the Releases page and grab one asset for your system:

Platform	Asset	Size
Linux x86-64 (an x86-64-v3 processor, with AVX2: Intel Haswell, AMD Excavator or later; glibc 2.34 or later)	VX-0.11.0-linux-x86_64.tar.gz	9.9 MB
macOS arm64 (Apple silicon, macOS 26 or later)	VX-0.11.0-macos-arm64.dmg	24.7 MB
Windows x86-64 (Windows 10 or 11; an x86-64-v3 processor)	VX-0.11.0-windows-x86_64.zip	61.7 MB

Each asset has a matching .sha256 file beside it. Verify before running:

sha256sum -c VX-0.11.0-linux-x86_64.tar.gz.sha256          # Linux
shasum -a 256 -c VX-0.11.0-macos-arm64.dmg.sha256          # macOS
certutil -hashfile VX-0.11.0-windows-x86_64.zip SHA256     # Windows: compare with the .sha256 file

2. Unpack / install

Linux relocatable tarball:

tar xzf VX-0.11.0-linux-x86_64.tar.gz
./VX-0.11.0-linux-x86_64/bin/vx

The tarball holds VX, the font it draws with (DejaVu Sans Mono, with its licence) and its documents (README-INSTALL.txt says what is where). VX uses the distribution's GTK 3, FreeType and OpenGL driver (OpenGL 3.3 or later), which most desktops have; otherwise:

• Debian, Ubuntu: sudo apt install libgtk-3-0 libfreetype6 libgl1
• Fedora: sudo dnf install gtk3 freetype mesa-libGL
• Arch: sudo pacman -S gtk3 freetype2 libglvnd

The binary needs glibc 2.34 or later (Ubuntu 22.04, Debian 12, RHEL 9, Fedora 35 and later releases). fontconfig (the system's fonts), libcurl (the demo datasets, Retrieve Reference, the chromosome band tables, the Known Features pane and the opt-in usage statistics) and ffmpeg (video recording) are used when present.

macOS arm64 .dmg:

1. Open VX-0.11.0-macos-arm64.dmg and drag VX.app into the Applications folder.
2. First start: VX is signed ad hoc, not notarised, so macOS refuses it the first time. Click Done, open System Settings > Privacy & Security, click Open Anyway beside the line about VX and confirm. Later starts open at once.
3. Or, before the first start, from Terminal: xattr -dr com.apple.quarantine /Applications/VX.app

The disk image's README.txt says the same.

Windows x86-64 relocatable ZIP:

1. Extract VX-0.11.0-windows-x86_64.zip anywhere.
2. Run vx.exe in the extracted vx-windows folder. Nothing is installed system-wide. vx-console.exe beside it is the same program for a terminal: the command line, --headless and --analysis (whose answer it prints).
3. Windows SmartScreen may warn at the first start (the program is not signed): choose More info, then Run anyway.

3. Verify it works

While VX runs it serves an HTTP API on 127.0.0.1:9876 (on by default; Settings > MCP / API). From any shell:

curl -s http://127.0.0.1:9876/ping
# → {"status":"ok","app":"VX","api_version":2}

That's your installation test.

---

Drive VX from Claude or any LLM (MCP)

VX comes with an open-source MCP server (mcp/vx_mcp_server.py, MIT) of 80 tools that exposes the viewer to AI agents that speak the Model Context Protocol: Claude Code, Claude Desktop, and any MCP-compatible client. Its capture tools return the image itself, which a vision model sees.

# One-time setup
pip install -r mcp/requirements.txt      # fastmcp and httpx

# Then in your Claude Code config (or Claude Desktop config):
#   "vx": { "command": "python", "args": ["/path/to/vx_mcp_server.py"] }

See mcp/INSTALL.md for the full Claude Desktop / Claude Code JSON snippets and api/examples/claude_mcp_recipes.md for seven worked agent recipes (BRCA1 figure, coverage walk, GC analysis, subsampled-BAM inspection, bookmark tour, SVG export, video recording).

---

Drive VX from a script (HTTP API)

Every MCP tool calls VX's HTTP API, which any HTTP client can call directly: 91 commands through POST /command, beside GET /ping, GET /state and the capture endpoints. An error answer carries an HTTP status (400, 403, 404, 405, 409, 501, 503, 504 or 500) and a JSON body with error and code. Since 0.11.0 the API refuses a request sent by a page in a web browser (403 FORBIDDEN_ORIGIN). The schema is api/openapi.yaml (OpenAPI 3.1). Three runnable examples:

• bash + curl: api/examples/curl_navigate.sh
• Python (stdlib only): api/examples/python_client.py
• Claude agent recipes: api/examples/claude_mcp_recipes.md

# Navigate to BRCA1
curl -X POST -H 'Content-Type: application/json' \
     -d '{"command":"navigate","params":{"chromosome":"17","start":43044295,"end":43125483}}' \
     http://127.0.0.1:9876/command

# Screenshot the viewport at 2x scale
curl -s 'http://127.0.0.1:9876/screenshot/viewport?scale=2.0' \
     -o brca1.png

# Run a peak-calling analysis on the visible region
curl -X POST -H 'Content-Type: application/json' \
     -d '{"command":"run_analysis","params":{"type":"SimplePeakCalling","tracks":["ChIP_signal"]}}' \
     http://127.0.0.1:9876/command

---

What you get out of the box

Built-in computational analyses (59, each in a background thread)

The names are the API's. The window shows each in words ("GC Content"); four go by other names there, given in brackets.

Signal & quantification: SignalDifference, SignalRatio, ZScoreNormalization, FeatureCounts, RPKM, TPM, CPM, PerExonCoverage, BinStatistics, RegionDifferential

RNA-seq: DifferentialExpression, SpliceJunctions, TranscriptQuantification, GeneBodyCoverage

Peak calling & enrichment: SimplePeakCalling, EnrichmentOverInput, FeatureProminence, CoverageAnomalies, VariantDensityHotspots, GapDetection, CnvSegmentation

Sequence: GCContent, CpGDensity, MotifSearch, RepeatDensity, NucleotideFrequency, OrfPrediction

Intervals: IntervalIntersection, IntervalSubtraction, IntervalUnion, ClosestFeature

Variants: VariantAnnotation, TsTvRatio, VariantDensity, AlleleFrequencySpectrum, VariantIntersection, CodingImpact

Alignment QC: MappingQualityDistribution, InsertSizeDistribution, StrandBias, DuplicateRate, SoftClipFrequency, ChimericReadDensity, ReadLengthDistribution, PerBaseMismatchRate (Edit Rate), MismatchRate

Modifications and poly(A): ModificationSites, PolyAStatistics, ModificationCoOccurrence

Hi-C / interaction: ContactFrequencyVsDistance, InteractionEnrichment (Interactions at Features), InteractionEnrichmentOverExpected (Interaction Enrichment), TADBoundaryDetection, ContactMatrix

Cross-track: SignalCorrelation, SignalAtFeatures, HeatmapMatrix, ProfilePlotData

Export: TrackExport (Export Track Data)

New in 0.11.0: DifferentialExpression, SpliceJunctions, TranscriptQuantification, GeneBodyCoverage, OrfPrediction, CnvSegmentation, ContactMatrix, ModificationSites, PolyAStatistics and ModificationCoOccurrence.

Each analysis writes its result to a configurable results directory: a BED, bedGraph or TSV file, and for some analyses also a BED12 (ORF Prediction), bedMethyl (Modification Sites) or BEDPE (Modification Co-occurrence) file (Export Track Data writes a variant track as VCF), with a manifest beside it (<result>.manifest.json: the VX version, each input file with its size and, up to a size set in Settings, its SHA-256, the parameters as applied, the region), and can add the result as a new track, right below its input track. The 49 analyses of 0.10.0 are checked from outside by a validation package (MIT): Python reference implementations of the 49 analyses, fixtures with expected outputs, and a runner that drives a VX binary through its HTTP API and compares every file it writes. The ten analyses added in 0.11.0 were tested on real data.

File formats supported

Sequence: FASTA (.fa, .fasta, .fna, .fas), plain or gzip. Annotation: GTF, GFF, GFF3, plain or gzip. Alignments: BAM, with its .bai index beside it; SAM, read as a coverage track. Signal: BigWig, bedGraph, Wiggle. Intervals: BED (BED3, BED6, BED12), BigBed. Variants: VCF, plain, gzip or BGZF. Interaction: BEDPE (paired contacts, such as Hi-C loops). Transcript integrity: INDEGRA's process table ({sample}_process.txt). Sessions: .vxs.

BED, bedGraph and Wiggle files may also be gzip-compressed. A .cool, .mcool or .hic file is refused with the reason: write its contacts as BEDPE (cooler dump --join). See docs/FILE_FORMATS.md for the full matrix.

Export formats

Static: PNG, JPEG and TIFF (the viewport at a chosen scale; through the API also the whole window and a single track as PNG), SVG (the viewport, as editable vector graphics), the circular view and the dot plot (SVG or PNG), the view's sequence (FASTA) or genes and features (GFF), a track's data (VCF, bedGraph or TSV), an aggregate track's rows (TSV or bedMethyl), analysis results (BED, BED12, bedGraph, TSV, bedMethyl, BEDPE, with their manifests), the Edit mode's reference (FASTA and .fai), annotation (GTF or GFF3 with its gene index) and features (BED), a Known Features tab (Markdown or JSON), session save / restore (.vxs). Animated: MP4 (H.264), WebM (VP9), GIF direct viewport capture via ffmpeg pipe.

---

Known limitations in v0.11.0

These are documented up front so you can decide whether v0.11.0 fits your workflow.

• Signing. The macOS bundle is signed ad hoc only (no Developer ID, no notarisation): macOS refuses its first start until you allow it in System Settings > Privacy & Security (Open Anyway). The Windows program is not signed, so SmartScreen may warn at its first start.
• No Windows installer. Windows is distributed as a ZIP that runs from any folder; there is no .exe or .msi installer in v0.11.0.
• First-region BAM fetch on a cold, very large BAM (about 20 GB and more) can take about 2 minutes the first time you navigate into a chromosome. The load-status strip shows that the fetch is running; later regions are fast.
• No web/cloud version. VX is a desktop application. If you need an embeddable browser-tab viewer, JBrowse 2 may suit you better.

See docs/TROUBLESHOOTING.md for runtime troubleshooting.

---

Repository layout

This repository is the public-facing distribution tree for VX. The D source code that produces the binaries is held in a separate private development tree under a proprietary licence (see Source code below).

Folder	Contents	Licence
mcp/	Open-source Python MCP server for Claude / agents	MIT
api/	OpenAPI 3.1 HTTP API spec + curl/Python/Claude examples	MIT
docs/	User guide, file-format guide, keyboard shortcuts, troubleshooting, build provenance, screenshots	CC-BY-NC-ND-4.0
binaries/	Release-asset notes + SHA-256 verification recipes	CC-BY-NC-ND-4.0
third_party/	The one change VX's build makes to GtkD, published as GtkD's licence asks	LGPL-3.0-or-later (the change)
Pre-compiled binaries	Attached to each tagged release under Releases	CC-BY-NC-ND-4.0

The per-folder licence matrix is reflected by LICENSE-binaries.txt, LICENSE-docs.txt, mcp/LICENSE and api/LICENSE; third_party/GTKD.md states its own.

---

Source code

The VX D source code is proprietary and confidential and is held under a separate licence by Biocodecs Group / Arnaroo Ribologicals. It is not distributed via this repository.

The pre-compiled binaries are licensed under CC-BY-NC-ND-4.0 (non-commercial, no-derivatives, with attribution). For commercial licensing, OEM integration, white-label deployment, or source-code access: contact@biocodecs.org.

VX is based in part on works of the GtkD project (http://gtkd.org). The libraries the macOS and Windows packages carry (GTK 3 and the libraries it uses; on Windows also libcurl with OpenSSL) keep their own licences: each package's THIRD_PARTY_LICENSES.txt names every library it holds, with its version, licence and source, followed by the licence texts. On Linux these libraries come from the distribution. Every package holds the font VX draws with, DejaVu Sans Mono, with its licence. The one change VX's build makes to GtkD is published in third_party/GTKD.md.

The MCP server (mcp/) and the HTTP API specification (api/) ARE open source under MIT and can be reused freely.

---

Citing

If you use VX in research, please cite the accompanying manuscript and the Zenodo DOI of the specific release you used.

Cite the project (all versions, always resolves to the latest release):

Shirokikh, N. E., & Cleynen, A. VX: an AI-agent-controllable desktop genome and transcriptome browser with a programmable analysis framework. Zenodo https://doi.org/10.5281/zenodo.20234450

Cite bioRxiv preprint:

VX: an AI-enabled desktop genome viewer and transcriptome browser with a programmable analysis framework Nikolay E Shirokikh, Alice Cleynen bioRxiv 2026.05.17.725790; doi: https://doi.org/10.64898/2026.05.17.725790

Cite a specific release (v0.9.0.1, snapshot of the v0.9.0 "Aardvark" public pre-release):

Shirokikh, N. E., & Cleynen, A. (2026). Arnaroo/VX: VX 0.9.0.1, archival re-publish. Zenodo. https://doi.org/10.5281/zenodo.20234451

BibTeX:

@software{vx_genome_viewer,
  author       = {Shirokikh, Nikolay E. and Cleynen, Alice},
  title        = {{VX: an AI-agent-controllable desktop genome
                   and transcriptome browser with a programmable
                   analysis framework}},
  year         = {2026},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.20234450},
  url          = {https://doi.org/10.5281/zenodo.20234450}
}

Once the accompanying manuscript is published, this block will gain an additional reference to the journal article; the Zenodo DOIs above will continue to resolve to the archived software snapshots.

---

Build provenance

For full transparency on how each release binary was produced, including compiler version, flags, LTO settings, bundled libraries, signing flow, recommended downstream verification, see docs/BUILD_PROVENANCE.md.

---

Keywords

genome browser · transcriptome browser · sequence viewer · read alignment viewer · NGS · next-generation sequencing · RNA-seq · ChIP-seq · ATAC-seq · CLIP-seq · Hi-C · long-read sequencing · Nanopore · direct RNA · PacBio · Illumina · variant viewer · VCF viewer · BAM viewer · GTF viewer · GFF viewer · FASTA viewer · BigWig · BigBed · BedGraph · OpenGL · GPU-accelerated visualisation · GTK3 · D language · LDC · cross-platform bioinformatics · interactive genome visualisation · genome visualization · transcriptome visualization · MCP · Model Context Protocol · Claude · Anthropic · AI agent · LLM-driven analysis · agentic bioinformatics · HTTP API · peak calling · GC content · CpG density · motif search · Ts/Tv · variant density · alignment QC · MAPQ · insert size · strand bias · soft clip · TAD boundary · contact-vs-distance · contact matrix · heatmap · profile plot · signal correlation · feature counts · RPKM · TPM · CPM · differential expression · splice junctions · Sashimi plot · transcript quantification · gene body coverage · ORF prediction · copy number · CNV · base modifications · m6A · 5mC · modkit · bedMethyl · poly(A) tail · coverage anomaly · gene structure · exon · intron · UTR · CDS · isoform · transcript · circular genome view · dot plot · annotation editing · subsampling · async loading · video recording · screenshot · screencast · SVG export · reproducible figures · publication-grade · desktop bioinformatics · IGV alternative · JBrowse alternative · Tablet alternative · SeqMonk alternative · Pileup.js alternative.