blatant-why

Open-source protein design agent for Claude Code

Commercial platforms wrap open-source tools behind paywalls and call it a revolution.
BY gives you direct access through Claude Code. No platform fees. Your tools, your compute, your designs.

_{Source: trust us bro}

Quick Start (5 minutes)

You don't need to be a developer. If you can open a terminal and paste commands, you can run BY.

1. Install prerequisites

2. Create your project

mkdir my-campaign && cd my-campaign
npx blatant-why init

This scaffolds everything: 11 MCP servers, 21 agents, 19 skills, 13 slash commands, and a CLAUDE.md orchestration file. Takes about 30 seconds.

3. (Optional) Configure compute

BY defaults to local GPU if one is available. Otherwise the first-run questionnaire will help you pick between local, HPC (RunPod / Modal / SLURM), or Tamarind cloud. See Compute Options below.

cp .env.example .env
# Add keys for whichever compute provider you'll use

4. Start designing

claude

Then just tell it what you want:

> "Design VHH nanobodies against PD-L1"

Or use the guided workflow:

> /by:plan-campaign

Or if it's your first time:

> /by:welcome

That's it. Claude Code handles the rest -- research, design, screening, and ranking.

What It Does

Give it a target protein. It researches across PDB, UniProt, and SAbDab. It plans a design campaign with statistical-strategy debate when the target is novel. It runs compute jobs on your local GPU (default), on your HPC (RunPod / Modal / SLURM via the by-deploy-compute skill), or on Tamarind Bio cloud. It screens every design for structural quality, sequence liabilities, and developability. It ranks candidates by composite score. When you submit them to the lab and the results come back, it ingests the CSVs, diagnoses which in-silico features predicted reality, and feeds the calibration back into the next round.

The whole pipeline runs inside Claude Code. No platform. No dashboard. No vendor lock-in.

What's Inside

Setup

API Keys

Claude Code handles its own authentication. No separate Anthropic API key needed.

No keys needed for local-GPU mode — if you have an NVIDIA card with enough VRAM, BY can run the whole pipeline without any cloud service.

Configure your environment

After npx blatant-why init:

1. Copy .env.example to .env.
2. For local GPU (default) — set the tool paths:

   PROTEUS_FOLD_DIR=/path/to/Protenix
   PROTEUS_PROT_DIR=/path/to/PXDesign
   PROTEUS_AB_DIR=/path/to/boltzgen
   

3. For HPC (RunPod, Modal, SLURM) — add your provider key and let the by-deploy-compute skill handle deployment:

   RUNPOD_API_KEY=your_key_here   # or modal token, etc.
   

4. For Tamarind cloud fallback — add the API key:

   TAMARIND_API_KEY=your_key_here
   

5. For SSH remotes — add host configs to .by/config.json (compute.ssh_hosts).

The first-run questionnaire (run by the by-session skill on session open) walks through this interactively. Default compute.default_provider is "local"; priority order is ["local", "hpc", "tamarind"].

Compute Options

BY defaults to local GPU. The by-deploy-compute skill knows how to deploy Protenix / BoltzGen / PXDesign (and supplementary tools like AlphaFold, RFAntibody, ImmuneBuilder, ThermoMPNN, Boltz-2) on any of the targets below.

git clone https://github.com/bytedance/protenix
cd protenix && pip install -e .

PROTEUS_FOLD_DIR=/path/to/protenix

# Follow PXDesign installation guide

PROTEUS_PROT_DIR=/path/to/pxdesign

git clone https://github.com/jostorge/boltzgen
cd boltzgen && pip install -e .

PROTEUS_AB_DIR=/path/to/boltzgen

Your First Campaign

After setup, here's what to expect when you run your first design campaign.

Start Claude Code:

claude

Tell it what you want:

> "Design VHH nanobodies against PD-L1"

What happens next:

1. Research (~1-2 min) -- BY searches PDB, UniProt, and SAbDab for your target. It pulls crystal structures, known binders, epitope data, and literature context.
2. Campaign plan (~30 sec) -- You get a plan showing how many designs will be generated, which models will be used, and estimated compute cost. You approve before anything runs.
3. Design generation (~5-15 min) -- Compute jobs run on Tamarind Bio (or your local GPU). BY monitors progress and reports back.
4. Screening (~2-5 min) -- Every design is scored for structural quality (ipSAE), binding confidence (ipTM), and sequence liabilities. Problem candidates are flagged.
5. Ranking -- You get a ranked table of candidates with composite scores, ready for lab ordering.

Typical first campaign: 5-10 nanobody designs, ~20 minutes end-to-end, zero GPU required (Tamarind free tier).

You can also use slash commands for more control:

Architecture

flowchart TB
    User([User]) -->|prompt| Claude[Claude Code + CLAUDE.md]

    Claude -->|delegates| Agents[21 Agents]
    Claude -->|invokes| Skills[19 Skills]
    Claude -->|slash cmds| Commands[13 Commands]

    Agents --> MCP[11 MCP Servers]
    Skills --> MCP

    subgraph Data["Biological Databases"]
        PDB[(PDB)]
        UniProt[(UniProt)]
        SAbDab[(SAbDab)]
    end

    subgraph Compute["Compute (local-first)"]
        LocalGPU["Local GPU -- Default"]
        HPC["HPC: RunPod / Modal / SLURM"]
        Tamarind["Tamarind Bio -- Cloud Fallback"]
    end

    subgraph Models["Models"]
        BoltzGen["BoltzGen -- Ab/Nb Design"]
        Protenix["Protenix v1 -- Structure Prediction"]
        PXDesign["PXDesign -- Binder Design"]
    end

    subgraph Screening["Screening"]
        ipSAE["ipSAE Scoring"]
        Liabilities["Liability Scan"]
        Developability["Developability"]
        Diversity["Diversity Selection"]
    end

    subgraph Lab["Lab Integration"]
        Adaptyv["Adaptyv Bio -- Triple-Gated"]
    end

    MCP --> Data
    MCP --> Compute
    Compute --> Models
    MCP --> Screening
    MCP --> Lab

    Knowledge[("Knowledge Store")] <--> MCP

blatant-why/
├── assets/                  # Banner and diagrams
├── src/
│   ├── init-cli/            # npx blatant-why init CLI
│   └── proteus_cli/         # Python CLI (scoring, screening, campaign)
├── templates/               # Deployed by init CLI
│   └── .claude/
│       ├── agents/          # 21 specialized agents
│       ├── commands/by/     # 13 slash commands
│       ├── skills/          # 19 skills (see skills/README.md for catalog)
│       └── mcp_servers/     # 11 MCP server implementations
├── tests/                   # Test suite
├── CLAUDE.md                # Agent orchestration rules
├── package.json             # npm package
├── pyproject.toml           # Python package (uv)
└── README.md

Credits

Built by Tristan Farmer

• Hannes Stark and the MIT team for BoltzGen
• Deniz Kavi and Sherry Liu at Tamarind Bio
• Julian Englert at Adaptyv Bio
• The Claude Code team

License

MIT