Proto Workbench

Local AI. Scientific workspaces. Inspectable evidence.

Explore a question, work with scientific data, and keep the sources, calculations and review decisions together.

Explore the current source · Build and run · Documentation · Historical Windows preview

Open-source CLIs: Proto CLI · Chem CLI source and development guide

Source map: all operators, Skills and Harness

The September 24 unreleased managed Study increment adds frozen source-bound plans, project objects and retained evidence inspection. Its capability ledger distinguishes this implementation from the remaining next-architecture roadmap.

_{Actual source-browser capture · current method catalog · bundled open-source typography · no model inference performed for this screenshot}

Source and download versions differ. This page describes the September 2026 source update. The published v0.2.0-rc.1 Portable and Setup files are the earlier preview; they do not contain all the features below. Build from source for the current workspaces. Version and evidence guide →

Chat, Design, Compute

Chat	Design	Compute
Work with a local model, PDF/DOCX/XLSX documents, scientific tools, source quotations and reviewable claims. Retain conversations and working context across sessions.	Inspect DNA maps, edit source-bound occurrences and explore authentic protein coordinates with sequence mappings. Check and compile before reviewing the evidence.	Choose a method, inspect its inputs and assumptions, and save results with provenance. Reopen named research studies, compare recorded runs and compose figures.
Research conversations →	DNA editing → · Protein structures →	Scientific computing → · Saved studies →

A place to think. An empty, isolated Chat session. The selected model is not connected; no generated conversation is presented as evidence.	A calculation you can inspect. A local illustrative reaction model. The population scene follows calculated samples; it is not an atomistic trajectory.
Keep the study. A saved synthetic research project from the September 22 browser acceptance run.	Return to the evidence. A reopened supplied-alignment protein study from the same source-browser run.

These are actual application captures, with each capture's date, input scope and SHA-256 recorded in the gallery manifest. Browser views, fixture data and local calculations are identified separately from native packaged acceptance. Earlier native DNA, protein and Harness captures →

One research workflow, shared scientific tools

• Move between Proto and Chem. Both editions use Chat / Design / Compute. Chemistry adds structure inspection, analysis, chemical data and reaction networks, interfaces, reactors and kinetic studies through the shared tool registry. The current chemistry catalog contains 56 operators; availability and scientific scope remain operator-specific. Chem workspace →
• Keep data attached to its meaning. Saved studies retain input hashes, units, sample identities, method assumptions and result bindings. Protein comparison, RNA-seq studies and existing structure-prediction result import have their own explicit contracts. Data contracts →
• Build on recorded work. Reusable research workflows, claim review and figure panels refer to saved artifacts. Existing ColabFold result import reads supplied files; it does not launch a prediction job. Workflows → · Figures →
• Make recovery inspectable. Harness Slice B retains typed diagnostics, repair budgets, verdicts and negative memory. Unsupported work can end in abstention or human review. The shared execution journal preserves uncertain write effects for reconciliation. Harness implementation →

flowchart LR
    A[Question + sources] --> B[Chat / Design / Compute]
    B --> C[Shared tools + execution policy]
    C --> D[Saved inputs + results + receipts]
    D --> E[Review / compare / visualize]
    C -->|diagnostics + bounded repair| B
    D -->|uncertain effect| F[Reconciliation + human review]

Loading

Start where you work

I want to…	Start here
Run the current workspaces	Source development setup
Try the earlier Windows preview	Historical download · Desktop setup
Use the deterministic CLI	CLI installation and toy example
Set up a workspace with an API key or provider subscription	API-first initialization
Connect an AI host	MCP configuration and tools
Reproduce a scoped check	Current source evidence · CI profiles

LM Studio is an external application. Optional Python scientific packages, Psi4, XDL, WSL engines and OCI execution environments have separate setup and availability checks. A catalog entry or bundled UI does not establish an installed runtime. Public builds bundle Newsreader, Hanken Grotesk and Commit Mono; local self-use builds retain the supplied Anthropic families. Their font bytes stay Git-ignored. Typography profiles and licenses →

Try the CLI

From a source checkout, using Python 3.10 or later:

python -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install -e .
proto-agent check designs\toggle_switch.proto --json
proto-agent workflow run designs\toggle_switch.proto
proto-agent review run designs\toggle_switch.proto

This example uses a bundled toy development fixture. Outputs go under build/. Source-backed biological designs start with an eligible material search and a materialized selection. Full workflow guide →

What the evidence supports

Recorded check	Result	Boundary
Workbench offline suite	1,405 passed · 1 skipped	Local pre-publication source run; the skip retains its containment limitation
Python base suite	355 passed · 4 skipped	Separate local architecture run; Windows symlink skips remain explicit
Publishable source checkout	11 focused tests passed; TypeScript and desktop compilation passed	Exported staged files without local font bytes; reused installed Node dependencies
Frozen paired software campaign	Possible tasks: 24/24 direct · 23/24 Harness; correct abstention: 5/6 · 6/6	30 frozen tasks per arm; every attempt retained; software fixtures, not scientific-answer validation
Earlier Windows preview	Native Portable and extracted Setup-payload checks 8/8 each	Historical 0.2.0-rc.1 only; installation/upgrade/uninstallation untested
Typography follow-up	14 focused tests passed; local and public desktop source builds passed	Full offline run: 1,413 passed, 1 failed, 1 skipped; scan-suite recheck 6/6; both results retained

The paired campaign does not establish a completion or token-cost improvement. Its measured source predates the separately tested terminal-notification fix. Local checks do not establish a passing hosted CI run or clean-machine runtime acceptance. The hosted run at a3ff460 includes failures, recorded separately in the evidence guide. Current evidence and limitations · Machine-readable source summary · Hosted CI

Biological materials, with provenance

Public collection	Included	Boundary
iGEM DNA parts	1,046	Reviewed source records, sequence digests and per-record rights
UniProt proteins	20	Reviewed sequence records with source and license evidence
Quarantine metadata	1,795	Separate index; no sequence objects; activation denied; never model-visible

The checked-in distribution contains audited catalogs, licensed sequence objects, provenance and checksum inventories. The live external catalog and crawl caches remain local. Installing a public catalog does not activate it.

Public bundles · Data licenses · Material workflow

Explore, build, contribute

Documentation hub · Architecture upgrade · Research programme and open work · Changelog · Contributing · Security · Issues

Experimental research software. Workbench and CLI versions are independent. Software checks and policy eligibility do not establish experimental readiness; scientific interpretation requires human review. This project does not provide wet-lab execution instructions.

Software: MIT. Biological data and bundled third-party code retain their own license terms. Third-party notices.