Genomix CLI

AI-powered CLI for DNA sequence and genome analysis.

Genomix is an intelligent command-line tool that helps biologists, bioinformaticians, and researchers analyze genomic data through natural language. Ask questions about your VCF, FASTA, or FASTQ files — the AI reads them, queries real databases (NCBI, Ensembl, ClinVar), and explains results in accessible language.

Local-first. Runs with Ollama by default — your genomic data never leaves your machine.

What It Does

❯ Read raw_variants.vcf and give me a clinical summary

  ⚡ read_file(path='raw_variants.vcf')
    ↳ ##fileformat=VCFv4.2 ...

  1. BRCA1 missense (chr17:43094464): Pathogenic — increased breast/ovarian cancer risk
  2. CFTR deletion (chr7:117559593): Pathogenic — cystic fibrosis (homozygous)
  3. HBB missense (chr11:5226773): Pathogenic — sickle cell trait (carrier)
  4. APOE missense (chr19:44908822): Risk factor — Alzheimer's disease
  ...

❯ What does this reveal about the person's ancestry?

  Based on the variant profile:
  - HBB/rs334 (sickle cell trait): high frequency in African/Mediterranean populations
  - CFTR deltaF508: most common in Northern European populations
  - Combined profile suggests mixed European/African ancestry

Features

• Natural language interface — ask questions about your genomic data in plain English or French
• 18 MCP servers — 5 biotools (samtools, BWA, GATK, BLAST+, FastQC) + 13 databases (see below)
• 27 slash commands — /qc, /align, /variant-call, /blast, /msa, /explain, /report, /structure, and more
• 21 built-in skills — specialized AI instructions for sequencing, comparative genomics, clinical, oncology, pharmacogenomics, and more
• Protein structure analysis — AlphaFold predictions, AlphaMissense pathogenicity, PDB experimental structures
• Streaming responses — token-by-token display with thinking spinner
• Clinical HTML reports — /report generates styled variant reports
• Smart analysis — reads raw VCFs (no annotations needed), identifies genes from coordinates, infers clinical significance
• Ancestry inference — population frequency analysis via gnomAD/1000 Genomes
• 3 AI providers — Ollama/local (default), Claude (Anthropic), OpenAI
• Privacy mode — automatically active with local models, raw sequences never sent to cloud
• MCP management — /mcp to view, connect, and manage bioinformatics tool servers

Installation

# Install
pip install genomix-cli

# Check dependencies
genomix setup

# Initialize a project
cd my-analysis/
genomix init

Requirements

• Python 3.11+
• Ollama with a model (e.g., ollama pull qwen3-coder:30b)
• Optional: samtools, BWA, GATK, BLAST+ for bioinformatics tools

Quick Start

# Start interactive mode
genomix

# Non-interactive usage
genomix ask "What is the BRCA1 gene?"
genomix ask "Read sample.vcf and summarize the variants"
genomix run /qc data/reads.fastq.gz

Interactive Session

   ██████╗ ███████╗███╗   ██╗ ██████╗ ███╗   ███╗██╗██╗  ██╗
  ...
  v0.4.2 — AI-powered genome analysis

  ┌──────────────────────────────────────────────────────┐
  │  Project    BRCA Analysis - Cohort 2026              │
  │  Organism   Homo sapiens                             │
  │  Reference  GRCh38                                   │
  │  Provider   ollama (qwen3-coder:30b)                  │
  │  Privacy    🔒 Local — data stays on this machine    │
  │  MCP        18 registered (4 connected, 14 missing)   │
  └──────────────────────────────────────────────────────┘

  Connecting MCP servers...
  Connecting to ClinVar... ✓ (3 tools)
  Connecting to dbSNP... ✓ (3 tools)
  Connecting to Ensembl... ✓ (5 tools)
  Connecting to NCBI... ✓ (4 tools)

❯ _

Slash Commands

Full reference with examples: docs/commands.md

Command	Description
Analysis
/qc	Quality control (FastQC)
/align	Align reads to reference genome
/variant-call	Call variants (GATK/FreeBayes)
/annotate	Annotate variants (SnpEff/VEP)
/pipeline	Full pipeline: QC → align → call → annotate
/report	Generate styled HTML clinical report from VCF
Databases
/lookup	Look up a gene or variant across databases
/frequency	Population allele frequencies (gnomAD)
/disease	Disease associations (OMIM)
/cancer	Somatic mutation context (COSMIC)
/drug	Pharmacogenomics annotations (PharmGKB)
/literature	Search biomedical literature (PubMed)
Structure
/structure	Protein structure and AlphaFold predictions
/domains	Protein domain mapping (InterPro)
Comparative
/blast	BLAST similarity search
/msa	Multiple sequence alignment
/phylo	Phylogenetic tree construction
Exploration
/summary	Summarize a genomic file
/search	Query databases (NCBI, Ensembl...)
/explain	Explain a variant, gene, or region
Session
/mcp	Manage MCP servers (connect, status)
/swarm	Show background analyses
/provider	Switch AI provider
/model	Switch model
/help	Show available commands

Supported Databases

Database	Description
NCBI	Gene, nucleotide, and protein search
Ensembl	Genome browser, gene annotations, variants
ClinVar	Clinical variant interpretations
dbSNP	SNP identifiers and allele frequencies
gnomAD	Population allele frequencies
OMIM	Mendelian disease catalog
PharmGKB	Pharmacogenomics annotations
COSMIC	Somatic mutations in cancer
InterPro	Protein domains and families
PubMed	Biomedical literature search
AlphaFold	Protein structure predictions
UniProt	Protein sequences and annotations
PDB	Experimental protein structures

Protein Structure Analysis

Genomix integrates with Google DeepMind's AlphaFold for structural variant interpretation:

❯ /structure TP53

  ⚡ uniprot_gene_to_accession(gene_name='TP53')
  ⚡ alphafold_prediction(uniprot_id='P04637')
  ⚡ pdb_search_gene(gene_name='TP53')

  TP53 (Cellular tumor antigen p53)
  UniProt: P04637 | AlphaFold pLDDT: 75.06
  PDB: 172 experimental structures

  Domains: DNA-binding (IPR008923), Tetramerization (IPR003106)
  Hotspot mutations: R175H, R248W, R273H (DNA-binding domain)

When analyzing missense variants, Genomix automatically checks:

• AlphaFold confidence at the variant position
• Protein domain context (via InterPro)
• AlphaMissense pathogenicity score

Architecture

┌─────────────────────────────────────────────┐
│              genomix-cli                     │
│                                              │
│  CLI/TUI ── Agent Loop ── Swarm Manager      │
│                 │                             │
│    ┌────────────┼────────────┐                │
│    ▼            ▼            ▼                │
│  Tool       Skills       Project              │
│  Registry   System       Manager              │
│    │                                          │
│    ▼                                          │
│  MCP Servers                                  │
│  ├── biotools: samtools, BWA, GATK,           │
│  │   BLAST+, FastQC                           │
│  └── databases: NCBI, Ensembl, ClinVar,        │
│      dbSNP, gnomAD, OMIM, PharmGKB, COSMIC,   │
│      InterPro, PubMed, AlphaFold, UniProt, PDB │
│                                               │
│  AI Providers                                 │
│  Ollama (local) │ Claude │ OpenAI             │
└───────────────────────────────────────────────┘

AI Providers

Genomix supports 3 AI backends. Switch anytime with /provider in the chat.

Option 1: Ollama (local, default)

Everything stays on your machine. No API key needed. Best for sensitive/patient data.

# Install Ollama
brew install ollama

# Pull a model (pick one)
ollama pull qwen3-coder:30b    # Best quality, needs 18GB RAM
ollama pull qwen3.5             # Faster, lighter, 128K context
ollama pull llama3.3:70b        # Alternative, needs 40GB RAM

# Start Ollama (runs in background)
ollama serve

Config (~/.genomix/config.yaml):

provider:
  default: ollama
  model: qwen3-coder:30b

No secrets file needed. Privacy mode is automatic.

Option 2: Claude (Anthropic)

Best reasoning quality. Requires an API key from console.anthropic.com.

# 1. Get your API key at https://console.anthropic.com/settings/keys

# 2. Create config
cat > ~/.genomix/config.yaml << 'EOF'
provider:
  default: claude
  model: claude-sonnet-4-6
EOF

# 3. Store your API key (secure file, never committed to git)
cat > ~/.genomix/secrets.yaml << 'EOF'
anthropic_api_key: "sk-ant-your-key-here"
EOF
chmod 600 ~/.genomix/secrets.yaml

# 4. Launch genomix
genomix

Available Claude models:

Model	Best for
claude-sonnet-4-6	Fast, good quality (recommended)
claude-opus-4-6	Best reasoning, slower
claude-haiku-4-5-20251001	Fastest, cheapest

Option 3: OpenAI

Requires an API key from platform.openai.com.

# 1. Get your API key at https://platform.openai.com/api-keys

# 2. Create config
cat > ~/.genomix/config.yaml << 'EOF'
provider:
  default: openai
  model: gpt-4o
EOF

# 3. Store your API key
cat > ~/.genomix/secrets.yaml << 'EOF'
openai_api_key: "sk-your-key-here"
EOF
chmod 600 ~/.genomix/secrets.yaml

# 4. Launch genomix
genomix

Available OpenAI models:

Model	Best for
gpt-4o	Best overall (recommended)
o3	Strongest reasoning
gpt-4-turbo	Fast, 128K context

Switching providers on the fly

Inside a genomix session, switch without restarting:

❯ /provider claude
  Switched to provider: claude

❯ /model claude-opus-4-6
  Switched to model: claude-opus-4-6

❯ /provider ollama
  Switched to provider: ollama

Privacy considerations

Provider	Data location	Best for
Ollama	100% local	Patient data, GDPR, confidential
Claude	Anthropic servers	Research, best analysis quality
OpenAI	OpenAI servers	Alternative cloud option

With Ollama, raw sequences never leave your machine. With cloud providers, only tool result summaries are sent (not raw genomic data) when privacy mode is active.

Contributing

Contributions welcome! See CONTRIBUTING.md for development setup, project structure, and how to add new MCP servers, skills, or AI providers.

The easiest way to contribute is adding a new database MCP server — each one is a single self-contained file. See the architecture docs for an overview of the system.

License

Apache 2.0