Read papers, run research Apps, and inspect the evidence behind a result.
Our approach · Paper · Research example · Try Pero · Development · Roadmap · 简体中文
Pero brings original papers, research Apps and Agent conversations into one desktop workspace. Start with a PDF: read it, ask a question, then follow the answer's citations back to the source.
- Keep the original evidence. Save papers and link report claims to observed passages. Earlier answers do not replace source evidence.
- Define the method before computing. Fix inputs, method and comparison rules, then check independently. The Fisher case applies this to one published number.
- Continue from a record. Retain findings and unresolved gaps for explicit follow-up questions.
Target workflow: shared cross-person project memory and the complete research loop remain planned.
To follow along in Electron, see Try Pero below.
These are simplified illustrations, not screenshots. Some illustrated navigation is newer than the November 4 package; the steps use that release's controls.
Open Papers. If prompted, choose Prepare environment and wait for Environment ready; first use may download uv, CPython 3.12 and PDF dependencies. Select Add papers → PDF → Import PDF, or drop in a public PDF with selectable text. Scanned pages are not automatically searchable. See offline/Python setup.
Read the saved original. Select a passage and choose Highlight to keep an annotation. Importing, reading and highlighting need no model account or API key.
Choose Ask AI or Discuss selection to open a conversation with paper context. It does not send automatically. Configure an Agent first: Settings → Built-in Agent needs an API provider and key; external Agents use their own login. Full-PDF research requires HTTP MCP support. See Agent setup.
Send a question such as: “What evidence supports this method, and which limitations should I check? Cite the original passages.”
When a report is saved, close the chat and open Research record (the reader's clock icon), then the saved report. Click a citation to inspect the original passage; if Citation evidence lists several matches, choose the relevant one. Continue from here prepares a follow-up for you to send.
More: Paper usage and boundaries · Recorded Paper journey.
- Data: inspect and chart dataset versions without a model. Try six synthetic records.
- Virtual Cell: explore the PBMC3k reference atlas and illustrative reaction simulations.
The Fisher (1936) case recomputes one Table III value from 100 reviewed Table I measurements: 19.1434 cm², the within-species sum of squares. It freezes inputs and comparison rules, checks the arithmetic in a separate process, and exports an evidence ZIP. After extraction, python -I -S verify.py checks it using only Python's standard library, without a model or network.
This is a curated single-result reanalysis, not general paper reproduction; the whole-paper verdict remains NOT_ASSESSED. Backend/API, export and restart checks exist; browser interaction, Native/Electron integration and installed-package acceptance remain unverified. See fixed-case instructions.
Note
Development preview. Capabilities follow current source; downloaded packages may lag behind. The walkthrough uses November 4 controls, with no new installed-package or real-model acceptance claimed here.
The 2026-11-04 prerelease provides these packages. Check Releases for newer builds and read the release notes.
| Product | Platform | Asset |
|---|---|---|
| Pero, Electron workbench | macOS Apple Silicon | Pero-macos-arm64.dmg |
| Pero, Electron workbench | Windows x64 | Pero-windows-x64.zip |
| Pero, Electron workbench | Linux x64 | Pero-linux-x64.tar.gz |
| Pero Native, GPUI workbench | macOS Apple Silicon | Pero-Native-macos-arm64.dmg |
Build metadata and SHA256SUMS accompany the downloads. macOS packages are ad-hoc signed, not notarized; Windows/Linux packages are unsigned. Electron and Native use separate profiles. Use E 7D68 lectron for the Paper walkthrough above; Native has a fresh-profile limitation. See release verification.
Complete development setup, including the pinned GPUI Kit, then run from the repository root:
npm run dev:paper # Open Paper in a separate development profilenpm run dev:apps opens Virtual Cell in a separate development profile. The November 4 release has no standalone Virtual Cell package.
Paper reading, cited reports, local Apps and research-project links are implemented. Autonomous research has controlled-stack checks; real-model and independent human quality acceptance remain pending. Shared App Task/Run integration, project memory, team sync and shared Web/mobile research clients remain planned. Quality gates and the Roadmap separate implementation from acceptance.
- Data: Apps store domain data on the execution machine. Remote models, literature services, MCP connectors and shared task services may receive selected prompts, sources or task data.
- Trust: App backends and local plugins run trusted code, not a general-purpose process sandbox. Review tools and the session's permission policy before using sensitive data.
- Cost: Model subscriptions/API calls and remote compute are billed by their providers. Missing usage is not zero cost; research call/time limits are not a billing cap. Local reading needs no model call after setup, and local App calculations need no control-plane service.
- Evidence: Citations locate sources; execution success or matching numbers do not establish scientific validity. Review conclusions and unresolved gaps yourself.
The Roadmap proposes verifying published results, comparing experiments, then sustaining research across people and environments. These are target workflows, not a completed end-to-end research loop.
Apps own domain methods and data; Host owns local processes, credentials and Agent sessions; the optional Go control plane coordinates shared tasks and Workers. In the diagrams below, solid boxes mark implemented foundations; dashed boxes and links mark planned capabilities or integration.
Architecture · Decisions · Development
Bring a research question, a small public dataset with provenance, an App or an independent result review. For substantial changes, open an Issue to agree on scope and checks. See CONTRIBUTING and join the community.
License: there is no top-level project license. Do not assume a blanket reuse grant; ask the maintainers about reuse/distribution. Third-party components retain their own licenses.