Ian Kellar.

Behavioural and implementation science

ian@k3llar.com
ORCID
Bluesky
GitHub

Writing

Claude Science, and the open equivalents that only match part of it

9 November 2026

This is one part of a series on research-agent systems. The landscape piece states where the field stands and sets all twenty-seven systems side by side: Agentic science harnesses: the landscape.

Claude Science and the open equivalents

Full dossier from research card t_03745838. Anthropic's AI workbench for scientists, plus the open projects that claim or approach its ground: Runcell Science, OpenScience, AIPOCH Open-Science (newly profiled 14 Sep 2026), Biomni and Eigent. Ends with the "which is the open equivalent" assessment.

Claude Science: the incumbent workbench (Anthropic)

Team/origin: Anthropic's third flagship product line after Claude Code and Claude Cowork, announced 30 Jun 2026 in beta for macOS/Linux on Claude Pro/Max/Team/Enterprise plans. Anthropic stated it is "not a new AI model and not a more capable model for biology", it runs the same Claude models in the plan. Architecture: a generalist coordinating agent with 60+ curated skills and connectors (genomics, single-cell, proteomics, structural biology, cheminformatics), which can spawn specialist sub-agents and user-created agents; a reviewer agent checks citations, calculations, figures vs code and self-corrects during pipelines. Runs as a local server with a browser UI, persistent Python/R kernels, jobs on local kernels, Slurm clusters over SSH, or Modal sandboxes. Every artifact carries code + environment + plain-language description + message history (provenance). Connects to NVIDIA BioNeMo Agent Toolkit models (Evo 2, Boltz-2, OpenFold3). HN analysis (gjuggler): architecture is TRE-friendly: the server can run inside a Trusted Research Environment with UI proxied to the browser, which matters for pharma/genomic-biobank adoption.

Lit review: literature analysis is a core stage; connectors include PubMed/Crossref/Semantic Scholar/arXiv/FDA (HN ml_more). Allen Institute's Jerome Lecoq built a ~20-skill multi-agent "computational review template" that reads thousands of papers, stores claims+key numbers in an evidence database, then writes 100+ page reviews section-by-section with actor-critic pairs. Citations checked by reviewer agents. Limitation: cannot access paywalled papers (Lecoq, The Scientist). HN testing (ml_more) estimates citation confabulation drops from ~50-60% base rate to ~5-15% with RAG + citation-check instruction, but warns residual hallucinations are still dangerous.

Pricing/hosting: no separate price: included in Pro ($20/mo), Max (from $100/mo), Team (from $20-25/seat), Enterprise; usage counts against the same 5-hour/weekly limits as Claude Code/Cowork. Academic: "Claude Team plan for scientists": Standard seats $0/mo, Premium $15/mo for verified PIs at academic/nonprofit institutions (12-month promotional pricing). AI for Science: up to 50 projects with up to $30k in credits; Modal adds up to $2k compute per selected project (Modal committing up to $100k total). Applications closed 15 Jul 2026, projects ran Sep-Dec 2026. Local-first: macOS 13+, Windows (support added 10 Sep 2026, changelog 0.1.47) and Linux x64; raw datasets stay local. Provider tie: Anthropic models only (no BYO-model). GitHub: none (closed). X: no dedicated account; promotion via @AnthropicAI.

November momentum (added 9 Nov 2026): changelog 0.1.72 (5 Nov) makes LaTeX previews render tables and body text more accurately, completes a Directory connector such as Google Drive in the browser when Claude Science runs in its own window, and extends the macOS and Linux sandbox to more developer caches and startup files (Linux starter environments take about 5 GB more disk after a background copy). Anthropic's science page added the first 3D structure of a vault particle, resolved with Claude (4 Nov), and a teaching post on AI-built scientific models (6 Nov). On the platform side, an IP allowlist beta opened (5 Nov), text watermarking reached Mythos 5, Opus 4.7 and Sonnet 4.6 (3 Nov), message threads entered beta (2 Nov), and fast mode for Opus 4.8 was deprecated on 30 Oct for removal on 30 Nov.

Reception: positive, The Scientist (15 Jul 2026): early users Stephen Francis (UCSF), Jerome Lecoq (Allen Institute), Kelvin Lau (EPFL) all said they would continue using it; faster workflows, transparency ("It really shows you step-by-step, 'I am doing this. This is the code I'm using'"), strong provenance; WGS analysis in minutes (HN packeted); Medium (Champ18ion): "a real, useful, well-timed product". Negative: guardrails "too stifling": Francis' lab got biosecurity-style refusals on non-biosecurity prompts (The Scientist); paywalled-paper gap (Lecoq); reviewer doesn't re-run analyses and "reduces, but doesn't eliminate" errors (docs; Coursiv); "zero connectors for non-biology" (HN throwaway219450); citation hallucination still 5-15% (ml_more); "Claude Cowork in a lab coat" framing (Medium). Neutral: MIT Technology Review: "Anthropic's newest flagship product"; Stephen Turner test drive: "nothing revolutionary here, but some nice cohesion my stack lacks". The Life Sciences Verification Program (17 Sep) responds directly to the guardrail complaints above: enforcement shifts from real-time blocking to offline monitoring, with 30-day retention of flagged traffic.

Model layer + momentum (updated 12 Oct 2026): Claude Science runs on the plan's Claude models, and that layer moved on 1 Sep with Fable 5.1 and Mythos 5.1 (same model, different safeguards): up to ~25% cheaper for typical workloads and up to ~45% savings on agentic work via reduced cache-read pricing; the Mythos biology access program landed on 17 Sep as the Life Sciences Verification Program (public beta): biology-permissive safeguards on Mythos 5.1, Opus 5 and Sonnet 5 via Standard Use grants (team-wide, yearly) and High-risk Use grants (single dual-use projects, 6-monthly), usable across Claude Science, Claude.ai, Claude Code and the API, with dozens of organisations onboarded in early access; Anthropic frames it as the answer to its September threat report (10 Sep), which described attempted misuse including biological-weapons-adjacent work. Anthropic's science page now leads with the formalisation of Fermat's Last Theorem (4 Sep 2026: first complete computer-checked proof, written in Lean largely autonomously over 11 days) and an independent test (vals.ai) in which Fable 5.1 cracked Sir Thomas Urquhart's 370-year-old "Cyphral Distich" cipher, viral on HN on 13 Sep (906 points, 382 comments). A science-results post on 17 Sep added: Claude optimised 30+ open-source biomolecular models to run ~4x faster on average (FlashPairformer kernels), plus a low-memory "Big" mode that folds systems above 10,000 tokens on a single GPU node; the optimised code is open-sourced and a protein design competition with Adaptyv Bio carries up to $1M in Claude credits and wet-lab validation for 5,000+ designs. Access: 10,000 free/discounted seats for scientists opened 27 Aug via the Claude team plan for scientists (standard seats free; premium $15/mo with 5x usage); the Model Hardware Standard research preview also opened to a first group of scientific research labs. Week of 22-25 Sep 2026: Claude Opus 5.5 shipped (22 Sep): first model of the 5.5 family, at Fable 5.1 level on most work at ~40% lower running cost than Opus 5 (API $4/$20 per 1M tokens, cache reads $0.20; ~30% faster output), leading agentic-coding benchmarks. Two science results followed: Claude autonomously discovered a novel enzyme system with CRISPR-like repeats (23 Sep): ~950 agents searched DNA databases for 21 hours / 210M tokens, gathering 200,000+ reverse transcriptases, narrowing 3,500 candidates to 20 reports, and finding "array-associated reverse transcriptases" (ART) in bacteriophages: an RT, a partner gene and a repeat array expressed as distinct short RNAs; pre-print released, function not yet known, Feng Zhang calls it "genuinely intriguing"; the post also introduced Anthropic's new life-sciences research group and lab (Bay Area, BSL-1/BSL-2, human-run wet lab). And Claude computed the nine-loop N=4 super Yang-Mills amplitude (25 Sep guest post): a challenge set by physicist Matt von Hippel in August, solved on an academic-scale budget inside Claude Science with Fable 5.1, validated by Lance Dixon; a concurrent result arrived from Song He's group (using GPT-6 for some constraints). Claude Science changelog reached 0.1.53 (24 Sep), then 0.1.54 (28 Sep) and 0.1.55 (29 Sep): covering the Sonnet-5.5 transition (sessions default to Extra-high reasoning) and follow-up fixes. Week of 12-19 Oct 2026: the changelog ran 0.1.61 (13 Oct: Linux now needs system-wide bubblewrap and socat; a renewed Windows signing certificate, SHA-1 60C2D9C5...51292D1), 0.1.62 (13 Oct: that system-wide requirement withdrawn after Linux start failures, large-file unresponsiveness fixed) and 0.1.63 (14 Oct: GPU detected only when nvidia-smi runs from a fixed folder set, shared package-cache environments now read-only, provenance lists more input files). The marquee result: Claude designed quantum circuits that beat the best published ones on eight specifications drawn from chemistry, optimisation, physics simulation and number theory, working largely on its own against an automated evaluator, in some cases by an order of magnitude (16 Oct; led by Andi Gu in Anthropic's STEM Fellows Program with Alek Kemeny, Susanne Yelin and Mikhail Lukin; arXiv 2610.22367). Week of 5-12 Oct 2026: Claude Haiku 5.5 shipped (7 Oct): the third of the 5.5 family, from $0.10/$0.50 per 1M tokens for prompts up to 100K (~75% cheaper to run than Haiku 4.5), 1M-token context, adjustable effort; the Claude Science changelog advanced to 0.1.60 (7 Oct) via 0.1.56 (2 Oct: an automatic single retry on another model when safeguards flag life-sciences content; Excel preview upgrades) and 0.1.59 (6 Oct: scoped plugin approvals; analysis-sandbox hardening on Mac, Windows and Linux). And the marquee result: the first complete map of the sky in ultraviolet light (8 Oct), produced with Claude Science by Brice Ménard (Johns Hopkins University and Anthropic): ~38,000 GALEX observations plus Swift and FIMS/SPEAR data cross-calibrated and merged onto one grid; the third of the sky no UV telescope has observed reconstructed by inpainting from visible/infrared/radio correlations, plus Gaia-based estimates for 100M+ stars; hold-out tests within ~10% of real measurements; every pixel labelled measured or predicted with an uncertainty. Data CC BY 4.0, processing code MIT.

Runcell Science: self-styled open alternative (runcell-ai)

GitHub (API, 9 Nov 2026): runcell-ai/runcell-science: 24 stars, 5 forks, 1 open issue, Apache-2.0, TypeScript, created 2026-07-02, last push 2026-10-14 (no push since). The eleven-week quiet spell ended on 14 Oct with two commits: an OpenCode provider (version-matched v2 SDK, authenticated local HTTP server, server-sent event stream) and a runtime reconciliation that shares the process launcher with the other providers. Only one release exists (0.0.1, 6 Jul); adoption is still negligible, and the companion runcell-desktop repo was also touched on 17 Oct. Description verbatim: "An open souece alternative to claude science. AI Agent for science supporting codex, claude code, etc." (typo in original).

Team/origin: Runcell team (runcell-ai org, parent company Kanaries Data Inc. per runcell.dev footer). Runcell itself is a separate commercial Jupyter-notebook AI agent; Runcell Science is independent, Apache-2.0, no Runcell account needed. Launch post on X by Elwynn Chen (@ob12er): "I built an open-source alternative to Claude Science - Runcell Science - with half my weekly Fable 5 Max quota... runs on your own Claude Code or Codex subscription - no extra API keys".

Architecture: TypeScript monorepo: apps/web (browser workspace), apps/server (API, session persistence, provider runtimes, Jupyter, MCP), apps/desktop (Electron), packages/ui, packages/science-connectors (MCP-style connectors: PubMed, ChEMBL, BioMart, Ensembl, UniProt, AlphaFold, OpenAlex, GTEx, ZINC, CellGuide), packages/nbcli. Runtimes: Codex via JSON-RPC app-server, Claude via Agent SDK, Grok Build via Agent Client Protocol, and (added 14 Oct) OpenCode via its v2 SDK. Local-first web app at 127.0.0.1:27183; drives the Claude Code / Codex CLI already installed and signed in: no separate API keys or per-token billing.

Lit review: reads papers, extracts methods, summarizes assumptions, compares methods, turns findings into reusable notes; literature workflows are an explicit skill area; PubMed/OpenAlex connectors; artifacts (figures, notebooks, molecule edits) stay attached to the session for evidence.

Pricing/hosting: free, open source (Apache-2.0); local web app; Electron desktop; Windows currently via WSL. Provider tie: cross-vendor: Claude Code + Codex (and Grok Build), model and runtime neutral; uses existing subscriptions. X: @runcellai (8 followers, product posts); author @ob12er posted the Science launch. Reception: minimal so far (22 stars, 1 contributor): too early for meaningful community reception; no HN coverage found.

OpenScience: largest open Claude-Science-style workbench (synthetic-sciences)

GitHub (API, 9 Nov 2026): synthetic-sciences/openscience: 4,358 stars (+36 since 2 Nov), 550 forks, 13 open issues, Apache-2.0, Bun + TypeScript (72%) / Python (16%), created 2026-07-03, current release v2.0.153 (30 Oct 2026); last push 9 Nov 2026 (active pull requests, including opt-in budgets and runtime limits for delegated subagents, PR #1044, and DaKES open-access benchmark datasets). Homepage openscience.sh. Description verbatim: "The open-source AI workbench for scientific research". Added to this report 7 Sep 2026.

Team/origin: Synthetic Sciences org; lead committer Ishaan Gangwani; contributors include Satya Arjunan (Victor Chang Cardiac Research Institute): wider team and funding unverified. Launched as an open-source AI research agent on 4 Jul 2026 (Show HN id 48786827, 13 points). Launch positioning mirrors Runcell Science's: an open Claude Science alternative, launched one day after Runcell's repo appeared.

Architecture: model-agnostic "AI co-scientist": give it a goal and it reads the literature, writes and runs code, runs experiments, and writes up results. Desktop app (macOS/Windows/Linux), browser workspace, terminal CLI (npm @synsci/openscience / npx synsci). Model access: managed "Ace" option (pay-as-you-go wallet, no subscription), your own API keys or provider sign-in (Claude, GPT, Gemini, DeepSeek, GLM, Kimi per third-party overviews), or local models (Ollama, LM Studio). MCP connections, plugins, custom agents/commands, SDK. Turned "fully user-owned" and local-first in the v2.0.x line at end Aug 2026 (PR #446 retired managed-service paths).

Lit review: literature review is a headline workflow: search scientific sources, compare findings, save cited evidence; 42 scientific connectors (science_fetch retrieves records/files), ~312 bundled skills (README), docs covering literature review, data analysis, reproduction and writing. No dedicated verifier agent documented; the README's own guidance is to "review sources, assumptions, code, and outputs before relying on a scientific conclusion".

Pricing/hosting: free, open source (Apache-2.0); local-first desktop/CLI; optional Ace managed models/research search (pay-as-you-go); own keys or local models without an account. Provider tie: none (cross-vendor + local). X: none verified (unverified). Reception: early but a strong adoption signal: ~4.0k stars within three months and a near-daily commit cadence; community coverage is still thin (no significant independent reviews found; third-party write-ups are SEO-reworked versions of its own material).

AIPOCH Open-Science: largest open workbench by adoption (aipoch)

GitHub (API, 9 Nov 2026): aipoch/open-science: 5,501 stars (+83 since 2 Nov), 553 forks, 40 open issues, Apache-2.0, created 2026-07-03, pushed 9 Nov 2026 (daily activity); releases v1.0.8 (4 Nov), v1.1.0 (6 Nov) and v1.2.0 (9 Nov): approved jobs now run on user-connected Kubernetes, Hugging Face Jobs, Modal and Ray beside SSH and Slurm; the standalone Open-Science agent became the default runtime; step-linked discussions arrived; macOS updates are differential. Homepage aipoch.com/open-science. Description verbatim: "AIPOCH Open-Science is an open-source, local-first, model-agnostic AI research workbench for macOS, Windows, and Linux, with scientific agents, Python/R notebooks, data connectors, and reproducible provenance." Added to this report 14 Sep 2026.

Positioning: the same wave as Runcell Science and OpenScience (repo created 2026-07-03, three days after the Claude Science launch). Its own comparison material positions it as an independent open-source workbench in the same broad category as Claude Science, without claiming identical integrations or performance. Local-first, model-agnostic, with reproducible provenance named as a first-class feature.

Reception: the fastest-rising adoption of any open workbench (5,501 stars; +83 this week, ahead of OpenScience), but independent reception is still thin; no significant third-party reviews located by the 12 Oct check; treat capability claims as self-reported for now.

Biomni: closest domain-native equivalent (snap-stanford)

GitHub (API, 9 Nov 2026): snap-stanford/Biomni: 4,069 stars (+27 since 2 Nov), 738 forks, 131 open issues, Apache-2.0 (with mixed-dependency caveat), Python, created 2025-03-19, last push 2026-11-02 (branch activity). Homepage biomni.stanford.edu.

Team/origin: general-purpose biomedical AI agent from Stanford SNAP (Jure Leskovec lab), lead author Kexin Huang. 20+ authors across Stanford, Genentech, Arc Institute, UCSF, Princeton, UW. Preprint bioRxiv 2025-05-30; peer-reviewed version published in Science (10.1126/science.adz4351, "Autonomous biomedical research with an artificial intelligence agent", supplementary correction 21 Aug 2026). Commercial layer (added 12 Oct 2026): the public platform now runs as Phylo's Biomni Lab: Phylo is the Stanford spin-out founded by Biomni authors Kexin Huang (CEO) and Yuanhao Qu (President), with Jure Leskovec among the scientific co-founders; launched 3 Feb 2026 with a $13.5M seed (a16z and Menlo Ventures' Anthology Fund with Anthropic); 300+ integrated databases, software systems and tools; Bayer, Ono Pharmaceutical and Chugai deployments (Bayer announced 22 Oct 2026, for literature and omics evidence workflows in drug discovery); Target Atlas launched 29 Sep 2026, an agentic target-prioritisation app at biomni.phylo.bio with a Michael J. Fox Foundation Parkinson's edition; Sage Bionetworks' Synapse platform (~3PB of data) integrated Sep 2026; free academic tier. The open-source project continues under snap-stanford; Stanford-reported usage at nine months: 15,000+ scientists and ~100,000 workflows.

Architecture: LLM reasoning + retrieval-augmented planning + code-based execution, composing workflows without predefined templates. Action space of 150 specialized tools, 59 databases, 105 software packages mined across 25 biology domains; ~11GB data lake auto-downloaded on first run; Know-How Library (auto-retrieved protocols); MCP support; PDF reports of execution traces; Biomni-R0 reasoning model (Qwen-32B-based, local via SGLang); Biomni-Eval1 benchmark (433 instances, 10 tasks; Science paper reports 443 questions with average accuracy 57%, matching specialists on 3 expert-level tasks).

Lit review: literature review is a core stated capability: automates literature reviews, hypothesis generation, protocol design; semantic search + summarization tools; AWS ML blog builds a systematic-literature-review agent on Biomni tools with strict numbered citation requirements per query.

Pricing/hosting: free, open source (pip install biomni); brings your own LLM API keys; local conda env; web UI; used on HPC. Provider tie: any: Anthropic, OpenAI, Azure, Gemini, Groq, Bedrock, Ollama, or custom; default claude-sonnet-4. X: @ProjectBiomni. Reception: positive: peer-reviewed Science publication; Stanford HAI coverage; HN lebovic prefers Biomni over Claude-Code-custom-setup for bio ("stays on the frontier for bio", HPC reawakening); AWS Bedrock integration blog. Negative/caveats: README security warning: executes LLM-generated code with full system privileges; sandbox recommended; Apache-2.0 headline does not extend to every bundled tool/database (mixed commercial terms, Eigent and README both flag this); release frozen 15 Apr 2025 vs current web platform.

Eigent: open orchestration layer (eigent-ai)

GitHub (API, 9 Nov 2026): eigent-ai/eigent: 15,780 stars, 1,886 forks, 264 open issues, Apache-2.0, TypeScript, created 2025-07-29, last push 2026-11-05 (latest release v1.0.7, 30 Oct). Description: "Eigent: The Open Source Cowork Desktop - Local and Free Alternative to Claude Cowork and Codex".

Team/origin: open-source Cowork desktop for multi-agent workflows, built on CAMEL-AI (founder Guohao Li; LinkedIn Sam Witteveen). Org claims 35K+ GitHub stars and 200+ contributors across projects (eigent.ai/about). Gained ~2K stars in a single day on 17 Jan 2026 (sulat.com). General-purpose: not science-specific.

Architecture: React/Electron/TypeScript frontend; FastAPI + uv + Uvicorn backend; CAMEL multi-agent framework; MCP integration; built-in browser + terminal toolkits; skill integration; multi-agent workforce with parallel execution; local deployment with local inference engines (vLLM, Ollama, LM Studio); cloud version; enterprise (SSO, access control). README: "100% open source from day one".

Lit review: none built-in for science. Its own comparison blog is explicit: use PaperQA2 for the literature layer, Biomni for domain reasoning, Jupyter+MCP for kernels, Eigent/OpenHands for orchestration, and accept native scientific rendering + provenance as open-stack gaps.

Pricing/hosting: free, open source (Apache-2.0); cloud platform paid (public beta); enterprise by contact; model-agnostic (cloud APIs, enterprise gateways, local inference). X: @Eigent_AI. Reception: 15.5k stars; HN and blog coverage mostly around Claude Cowork positioning; "New FREE Eigent AI DESTROYS Claude Cowork?" YouTube; its own blog positions it as the orchestration layer in an assembled open science stack.

wisp-science (added 2 Nov 2026)

xuzhougeng (Zhougeng Xu) and 18 other contributors; open-source project

1,068★

AGPL-3.0

Local-first · v1.24.0

What it is

An open-source, local-first desktop AI research workbench for scientific computing. It pairs a chat agent with a real computing environment: Python and R runtimes reached over SSH, WSL or a GPU host, MCP servers for bioinformatics tools, and OpenAI or Anthropic models by key. Desktop applications ship alongside iOS and Android clients, and remote access runs through a one-click Cloudflare Worker relay with QR pairing. First release 1 July 2026; 73 releases to v1.24.0 (3 Nov 2026).

How it works

  • Local-first execution: code runs on your own machine or on a server you name, not in the vendor's sandbox; SSH, WSL and GPU targets are first-class runtime choices
  • Tools in the same session: MCP bioinformatics servers sit beside general scientific computing, so a tool call and the code that consumes it share one workspace
  • Model-agnostic: OpenAI or Anthropic keys, so the harness is not tied to one vendor
  • Reach: desktop apps plus iOS and Android, and a relay that puts the same session in a browser on another device

Pricing / hosting / provider tie

Free, open source (AGPL-3.0); self-hosted or local desktop; costs are your own model API usage. Provider tie: none enforced (OpenAI or Anthropic keys). GitHub: xuzhougeng/wisp-science: 1,068 stars, 107 forks, 0 open issues (API, 9 Nov 2026; created 1 Jul 2026; latest release v1.24.0, 3 Nov 2026).

Positive

  • Steady, fast cadence: releases through v1.24.0 on 3 Nov 2026, with commits landing on 9 Nov; the desktop-plus-mobile shape and SSH/GPU runtimes match how computational scientists already work

Neutral

  • No independent evaluation, press coverage or third-party review located in this pass; reception is repo-level only (1,068 stars, 107 forks). The 29 Oct pass missed the project entirely

Which is the open equivalent?

  • No single open project replaces Claude Science one-for-one. That is the honest verdict, and Eigent's own comparison makes the same call.
  • Runcell Science self-styles as "the open-source alternative to Claude Science" and matches the positioning: science agent for papers/code/experiments, local-first, reuses your Claude Code/Codex subscription, but is very early (23 stars, 1 contributor, minimal reception).
  • OpenScience (added 7 Sep 2026) is close behind on adoption: ~4.3k stars in three months, Apache-2.0, cross-vendor and local models, 42 scientific connectors, frequent releases (latest v2.0.152), but it is young: no deep independent reception yet, provenance/verifier tooling less developed than Claude Science's, and its own docs tell users to review outputs before relying on them.
  • AIPOCH Open-Science (profiled 14 Sep 2026) is now the largest open workbench by adoption: 5,501 stars in eighteen weeks (+83 since 2 Nov; releases through v1.2.0, 9 Nov), Apache-2.0, local-first, model-agnostic, with scientific agents, Python/R notebooks and data connectors; independent reception remains thin so far. The field has widened further: ai4s-research/open-science (2,083★, "Open Science Desktop"; v0.7.1 on 29 Oct moved terminals outside the app so a Claude Code or Codex session survives an update or a relaunch, and added Claude Code, Codex and pi as conversations beside OpenCode) and exon-research/genomi (491★, genomics-focused) are also live, alongside Runcell Science.
  • wisp-science (added 2 Nov 2026) is a fourth workbench, and the most active of the smaller ones: 1,068 stars in four months, releases to v1.24.0 (3 Nov 2026), AGPL-3.0. It is a local-first desktop workbench for scientific computing: Python and R over SSH, WSL or a GPU host, MCP bioinformatics tools, and OpenAI or Anthropic models, with desktop plus iOS and Android apps and remote access through a one-click Cloudflare Worker relay. No independent evaluation located in this pass.
  • Biomni is the closest domain-native equivalent for biomedical work: peer-reviewed in Science, 150 tools / 59 databases / 105 software, Biomni-Eval1 benchmark, any-provider model support, and its public platform now runs as Phylo's Biomni Lab (commercial; the open-source project continues), but it is a biomedical research agent, not a full workbench: no native scientific rendering of proteins/structures, and provenance is PDF execution traces rather than artifact-level code+env records.
  • Eigent is the orchestration layer only; it belongs in an assembled open stack (Eigent or OpenHands + Jupyter/MCP kernels + PaperQA2 for literature + Biomni for domain reasoning), not as a science workbench itself.
  • Genuine open gaps vs Claude Science: native scientific rendering, provenance-by-default, and 60+ pre-wired database connectors.