A terminal agent for de novo protein design. It plans, runs, and ranks design jobs — and (optionally) ships the survivors to a wet lab — from a single Go binary.

Install · Quickstart · LLM setup · Tools · Roadmap

---

❯ fova design --target PD-L1 --n 50
✓ planning            retrieved 12 papers · BindCraft + ProteinMPNN
✓ scaffolding         RFdiffusion · 200 backbones · 4m 12s
✓ sequence design     ProteinMPNN · 200 sequences · 2m 41s
✓ predict             AlphaFold3 · 1m 08s
✓ rank · ipSAE        shortlist: 47 designs · top score 0.84
▸ confirm wet-lab submission? [y/N] ▍

Status: alpha (v0.5.0). The core loop works end-to-end, but the agent still has rough edges and the wet-lab + antibody / enzyme tracks are incomplete. See Known issues before using it on anything you care about.

---

The loop · from a sentence to the bench

01 ▸ STEP	02 ▸ STEP	03 ▸ STEP	04 ▸ STEP
⌖ Plans	⠿ Orchestrates	≋ Ranks	✓ Ships
Turns a natural-language target into a grounded DesignPlan, backed by current literature from Europe PMC, OpenAlex, Semantic Scholar & bioRxiv.	Runs experimentally validated tools like BindCraft, RFdiffusion, ProteinMPNN, and AlphaFold3 — locally, or on your own Modal GPUs.	Scores every design on interface and confidence metrics (ipSAE primary), then filters to a small, trustworthy shortlist.	Sends the survivors to the bench through the Adaptyv Foundry API, then closes the loop with real results.

Why fova · principles, not features

• ⊹ Free by default. Local LLMs work out of the box. No account, no key, no card. Paid LLMs and wet-lab submission are opt-in.
• ✓ Validated tools only. Every built-in design tool has documented wet-lab success. No unproven methods ship in the box.
• ≋ Ranked on real metrics. fova scores every design on multiple interface and confidence metrics, never a single number, so the shortlist holds up at the bench.
• ⌗ Provenance everywhere. Every design carries its full lineage in local SQLite: intent → tool versions → wet-lab result.
• ◇ Local-first, cloud-elastic. Runs offline with Ollama. Scales out to your own Modal account for GPU when you need it.
• ▸ Human-in-the-loop. A confirmation checkpoint before anything slow, costly, or irreversible — including every wet-lab submission.

Orchestrated tools · only tools with wet-lab receipts

▸ Design · binders · antibodies · enzymes BindCraft · RFdiffusion · RFdiffusion2 · ProteinMPNN · LigandMPNN · RFantibody · BoltzGen

▸ Structure & scoring · predict · ipSAE · filter ESMFold · Boltz-2 · Chai-1 · ipSAE · metrics · filter

▸ Knowledge · free · no keys Europe PMC · OpenAlex · Semantic Scholar · bioRxiv · Crossref · UniProt · RCSB PDB · InterPro · BLAST · web search · web fetch · local PDFs · Bleve corpus

▸ Wet lab · close the loop Adaptyv Foundry · targets search · cost estimate · experiment status · results webhook

Install

One static binary. No runtime. No dependencies.

▸ install.sh · macOS · Linux curl -fsSL https://alvarogonjim.github.io/fova/install.sh | sh

▸ from source · requires Go 1.22+ git clone https://github.com/alvarogonjim/fova cd fova && make build && ./bin/fova

The one-line installer fetches the latest GitHub release for your platform (linux/darwin × amd64/arm64), verifies SHA256 against the release's checksums.txt, and installs to ~/.local/bin/fova. Override with FOVA_INSTALL_DIR=/somewhere/else, or pin a version with FOVA_VERSION=v0.5.0. fova is Unix-only — Windows users run under WSL.

Verify: fova --version → fova 0.5.0

Quickstart

fova                    # launches the TUI

On first run, an onboarding wizard walks you through picking a model and (optionally) entering an API key. After that, type into the chat and the agent will plan, call tools, run jobs, and rank designs.

A useful first turn:

Design a binder against PD-L1. Pull the top-3 recent papers, draft a plan, and stop at the plan-approval gate.

The agent will hit Europe PMC + OpenAlex, build a DesignPlan, and pause for /plan approve before launching any GPU jobs.

LLM setup

fova picks a provider based on the first API key it finds and your config.toml's [defaults] block. Set the env var for the provider you want.

▸ Cloud providers

export ANTHROPIC_API_KEY=sk-ant-…    # Claude
export OPENAI_API_KEY=sk-…           # GPT
export GOOGLE_API_KEY=…              # Gemini

Models are defined in ~/.config/fova/models.toml (materialized from the embedded defaults on first run). Edit it via /config edit models or switch the active model in-app with /model <id>.

▸ Ollama · local, easiest

Install Ollama, pull a tool-capable model, then point fova at it:

ollama pull llama3.3:70b
ollama serve                         # default: http://localhost:11434
fova                                 # /model llama3.3:70b

Ollama is pre-wired in models.toml as the ollama provider with base_url = "http://localhost:11434/v1". No API key needed.

▸ vLLM · local, throughput-tuned

Run any OpenAI-compatible vLLM server:

vllm serve Qwen/Qwen3.6-27B \
  --host 0.0.0.0 --port 8000 \
  --api-key secret-token-please-change
export VLLM_API_KEY=secret-token-please-change

vllm is pre-wired as a provider in models.toml. Add or edit [[model]] entries to point at your served model. The base URL defaults to http://localhost:8000/v1.

▸ LM Studio · local, GUI-friendly

LM Studio exposes an OpenAI-compatible server at http://localhost:1234/v1. Add a provider block to models.toml:

[[provider]]
name     = "lmstudio"
kind     = "openai"
base_url = "http://localhost:1234/v1"

[[model]]
id             = "your-model-name"
display_name   = "LM Studio (local)"
provider       = "lmstudio"
context_tokens = 32768
supports_tools = true

Reload with /reload.

Design tools · install on demand

Design tools run inside isolated uv environments managed by fova. Install them from inside the TUI:

/tools              list available tools and their install status
/install bindcraft  install one
/doctor             diagnose the local environment

The local backend assumes a container runtime (Docker or Podman) is available on PATH for tools that need CUDA — /doctor will tell you what's missing.

For elastic compute, deploy the same tools to Modal:

/modal deploy       writes functions.py and runs the Modal CLI for you

You bring your own Modal account; fova does not run a proprietary cloud backend.

Configuration

fova materializes its config on first run. Default locations follow XDG.

File	Purpose
~/.config/fova/config.toml	UI theme, default provider/model, backend, knowledge tunables, webhook, budget
~/.config/fova/models.toml	Provider definitions + model catalog
~/.config/fova/system.md	Base system prompt — edit to steer the agent
~/.config/fova/skills/	Markdown skills surfaced via /skills
~/fova/	Data dir: workspace DB, project corpora, jobs

Open any of these with /config edit <name>, print the path with /config path, validate with /config validate.

▸ Environment variables

Variable	Effect
FOVA_HOME	Override the data dir (~/fova).
FOVA_CONFIG_DIR	Override the config dir (~/.config/fova).
FOVA_COMPUTE_BACKEND	local or modal — overrides [defaults].
ANTHROPIC_API_KEY	Claude API key.
OPENAI_API_KEY	OpenAI API key.
GOOGLE_API_KEY	Gemini API key.
VLLM_API_KEY	Bearer for the local vLLM server (if you set one).
ADAPTYV_API_TOKEN	Adaptyv Foundry token for wet-lab submission.
PAPERCLIP_TOKEN	Enable the optional knowledge.paperclip MCP forwarder.

Slash commands

Type / in the chat to see the menu. Highlights:

Command	What it does
/plan approve · /plan cancel	Gate the agent at the plan checkpoint.
/model <id>	Switch model (and provider) mid-session.
/tools · /install · /uninstall · /doctor	Manage local tools.
/modal deploy	Push the tool registry to Modal.
/auth <provider>	Store an API key in the OS keyring.
/skills list|show|new|edit|validate|reset|path	Manage skills.
/config edit|reset|validate|path	Manage config assets.
/reload	Pick up edits to config.toml / models.toml without restarting.
/keys · /help	Keybindings and command reference.
/quit	Save and exit.

Status

• v0.5 "Public alpha" · current — cleanup pass, public README, version alignment, going public.
• v0.4 "Closing the loop" · in progress — TUI redesign with a design-token palette, slash-command autocomplete, animated thinking indicator, and a context-meter status footer. Adaptyv wet-lab integration and the antibody / enzyme tracks are still to come.
• v0.3 "Plan from target" · done — free knowledge stack, per-project literature corpus, structured DesignPlan and /plan view, Google Gemini provider, pkg/proteinio FASTA / PDB / mmCIF helpers.
• v0.2 "Real designs" · done — SQLite persistence, async jobs system, uv + Modal backends, BindCraft / RFdiffusion / ProteinMPNN with ipSAE scoring, jobs + designs TUI panels.
• v0.1 "Hello, sequence" · done — chat TUI, agent loop, and fold.esmfold.

See docs/SPECS.md for the full implementation spec and docs/DESIGN.md for the visual identity.

Contributing

Early-stage alpha — issues, bug reports, and PRs welcome. Expect rough edges and a fast-moving codebase. Before opening a PR:

go vet ./...
go test ./...

If you're adding a tool, follow the existing adapters under internal/tools/ and internal/backends/local/. Tool docs live in docs/tools/.

License

AGPL-3.0-or-later. If you run a modified fova as a network service, the AGPL requires you to make your modifications available.

---

_{fova · design proteins in your terminal · github.com/alvarogonjim/fova}