macOSDMG
Apple Silicon
M-series Macs · signed and notarized
SHA256
Download Apple Siliconfb29ff801e1a86c1fa3fb5f08894f39fe36c0d4edb6b778fe4bd352a6c0a4755OPEN SOURCE · LOCAL-FIRST · SCIENTIFIC AGENT WORKSPACEVOL. 01 / v0.14.0
Wisp ScienceFrom research question to reproducible artifact.
Wisp Science brings papers, local data, Python, biological databases, and agent workflows into one traceable workspace—on desktop or through a headless CLI.
Bring your own model. Keep project state local.
RESEARCH RECORD · WS-014Analysis trace preserved
QUESTION
Which pathways distinguish the treatment-responsive cell population?
responsive population
01Evidence
12 papers · 3 local datasets
02Method
Scanpy workflow · persistent Python
03Result
Marker table · UMAP · methods note
0139
v0.14.0 bundled research SKILLs
0280+
scientific databases reached through MCP
03Python / R
persistent project runtimes
04Local
sessions, artifacts, and SQLite history
LATEST RELEASE
Install Wisp Science and begin with a local project.
Release files are served by GitHub. Package checksums are published here so each download can be verified before use.
Current releasev0.14.0
Published Jul 16, 2026. macOS packages are signed and notarized; Windows packages remain unsigned and may trigger SmartScreen.
Read release notesmacOSDMG
Intel
Intel Macs · signed and notarized
SHA256
Download Intel66641caf386d63955c7d29eb6b1ff008b1ae8d49240a879bddda549252b5d6b0WindowsEXE
Setup EXE
Standard Windows installer
SHA256
Download Setup EXE0548774ca0b0acffa30352c3119b43f56f480b68ecea38c6f3ab157d8de0123aWindowsMSI
MSI
Enterprise-style installer package
SHA256
Download MSI6fe544f526d9249ad2a8ce5dfb6cc5fc67a7c01f4622479465b19f746d423956RELEASE NOTES
What changed in v0.14.0
wisp-science v0.14.0 improves the conversation workspace, ACP collaboration flows, and desktop stability with ACP-backed side chat, a session mode selector, verifiable reviewer/harness access, durable message resource bindings, script file previews, one-click quote-to-chat selection, transcript/session pagination, math rendering, and a broad set of memory, rendering, preview, and UI consistency fixes.
01
What's changed
- Add ACP-backed side chat, a session mode selector, durable message resource bindings, safer reviewer lifecycle handling, and verifiable reviewer/harness backend selection.
- Confirm model switches that invalidate caches before applying them.
- Preview script files directly and send selected workspace text into the composer or side chat.
- Paginate long session history and transcripts, and consolidate top-bar, model-switcher, agent/context, and compute-host UI surfaces.
- Bound memory growth across UI events, tool snapshots, artifacts, processes, and long transcript rendering.
- Render Markdown math, fix JPEG2000 figures in WebView PDF preview, improve dark code highlighting, and clean up artifact/workspace UI inconsistencies.
WISP SCIENCE
Used by researchers from universities and research teams
Cornell University
University of Michigan
UCLA
Yale University
Tsinghua University
Affiliations are self-reported by individual users and do not imply institutional partnership, sponsorship, or endorsement.
WHAT STAYS WITH YOUR PROJECT
Keep sources, analysis, and results in one research workspace.
Work with papers and local files, run Python or R, query scientific databases, and keep code, figures, and decisions together. Your project stays on this computer; Wisp Science connects to an outside service only when you use one you configured.
What Wisp Science keeps together
- 01SOURCES
Papers and local files
Bring the material for one question into a project without losing track of the original files.
- 02PROCESS
Questions, plans, and decisions
Keep the reasoning behind each step beside the task, so you can review or continue the work later.
- 03ANALYSIS
Python, R, and research SKILLs
Run analysis and reusable workflows in the same project instead of moving data and context between separate tools.
- 04RESULTS
Code, tables, figures, and history
Return to the code, outputs, and conversation that produced a result when you need to check or share it.
BEFORE DATA LEAVES THIS COMPUTERLocal project files, explicit outside connections.
Project files, session history, and generated results are stored locally. If a task uses a model, database, ACP agent, or remote compute environment, the information needed for that step is handled under that service’s settings and credentials. Review sensitive material before enabling an outside connection.
A TRACEABLE RESEARCH LOOP
Move from discovery to communication without losing the method.
Each stage remains connected to its evidence, code, tools, and generated outputs, so the final report is a record of the work—not just an answer.
- 01
Discover
Search papers and biological databases, inspect local files, and assemble evidence around the question.
- 02
Analyze
Run persistent Python or R, shell, MCP tools, and domain SKILL workflows while preserving project-scoped state.
- 03
Verify
Review plans before execution, inspect intermediate files, and trace every artifact back to inputs and code.
- 04
Communicate
Export figures, tables, methods, citations, and narrative reports from one coherent research record.
REPRODUCIBILITY, BUILT INTO THE SESSION
Evidence, code, and context—kept in the same research record.
Tables, figures, code blocks, LaTeX equations, and file paths can become artifacts. Each stays connected to its producing conversation, code, logs, inputs, and environment details.
@ file references
Attach the file already open or any file in the current project without breaking the research context.
Plan approval
Let the agent propose a multi-step method, pause for review, and only continue after approval.
Artifact provenance
Open tables, figures, code, formulas, and file references together with the context and inputs that produced them.
PROVENANCE LEDGERWS / TRACE-014
01Input
counts_matrix.h5adsha256 verified02Method
scanpy_marker_workflow29 steps logged03Figure
umap_response.pngcode attached04Report
methods_and_results.mdcitations linkedREPRODUCIBLE
RESEARCH PLATES
One workspace, different scientific questions.
Compose database retrieval, local computation, domain workflows, and structured writing around the task at hand.
PLATE 0101 / 04
Single-cell analysis
Run Scanpy or scVI-style workflows, annotate populations, and export UMAPs, marker tables, and methods notes.
PLATE 0202 / 04
Protein structure
Fetch sequences and structures, combine AlphaFold-, Boltz-, or OpenFold-style skills, and draft structured interpretation.
PLATE 0303 / 04
Cheminformatics
Search ChEMBL or PubChem, compare activity data, calculate properties, and prepare SAR-style tables.
PLATE 0404 / 04
Literature and writing
Search PubMed or Semantic Scholar, inspect PDFs, draft discussion sections, and keep citations with the text.
DEMO ATLAS
Inspect a complete agent trace before using API credits.
Bundled read-only sessions expose the path from question to evidence and exported output for common life-science tasks.
FIELD NOTES FROM THE BENCH
How researchers move work forward.
Not a vague “analyze this” prompt, but a concrete research task advanced with evidence, code, and a traceable record.
01 / 06
Start by QCing this single-cell dataset, then help me identify those clusters. Keep every parameter, figure, and rationale—I need to walk through it at lab meeting tomorrow.
02 / 06
Map the preclinical studies on this target from the past five years. Don’t just give me a conclusion—show the sources, conflicting findings, and evidence that is still missing.
03 / 06
Could these mutations affect protein stability? Check the databases and literature first, then give me a locally reproducible analysis plan without moving the raw data off my computer.
04 / 06
I added 20 papers to the project. Organize the controls by experimental system, flag inconsistent numbers, and make every finding traceable to the original text.
05 / 06
Don’t interpret these metabolomics results yet. Check missing values, batch effects, and outliers first, then turn the recommended statistical steps into a rerunnable script.
06 / 06
Search ChEMBL and PubChem for these candidate molecules, compare activity, selectivity, and known risks, then identify the three most valuable experiments to run next.
Anonymous early-user feedback. Identities are summarized, wording has been lightly edited for length and clarity, and portraits are illustrative.
QUESTIONS BEFORE YOU BEGIN
What Wisp Science is—and where its boundaries are.
01Is Wisp Science a new model?
No. It is an open-source desktop and CLI application that uses the model provider and credentials you configure.
02How is it different from a generic AI assistant?
It can execute local Python, shell and file tools, call MCP databases, follow domain SKILL workflows, and retain artifact provenance.
03Does research data stay private?
Project files, sessions, artifacts, and settings remain local. Prompts and responses still pass through your configured model provider.
04Is it production-stable?
Wisp Science is an active preview for local scientific workflows. Review critical methods and outputs, and check release notes for current signing and update status.