Composable research apps and distributed agent collaboration.
Download dev build · Roadmap · Development · Status · Contribute · 简体中文
Note
Development preview. macOS arm64 builds are available. There is no stable release yet.
A research workflow often starts as a script, notebook or model demo. Pero turns it into an App that can run on its own or alongside other tools in a shared workbench. Virtual Cell is the first example, with cell-atlas exploration and reaction simulation.
The longer-term goal is to connect these Apps with agents, collaborators and compute while keeping inputs, methods and results traceable.
| Area | Current status | Roadmap |
|---|---|---|
| Desktop & Agent sessions | ACP conversations, tool results, approval UI, history and workspaces implemented | Sessions |
| Tasks & Workers | Dispatch, leases, cancellation and recovery implemented; Linux real-provider acceptance pending | Tasks |
| GitHub & contributions | Intake, claims and local coding implemented; real GitHub PR publication acceptance pending | Contributions |
| Apps & products | Virtual Cell runs locally and in a macOS development package; Agent tools and task integration pending | Apps & products |
| Projects, knowledge & teams | Project journals, memory, Pages and team sync planned | Projects & teams |
| Playbooks, plugins & connectors | Method orchestration, multi-agent teams and MCP integrations planned | Methods |
| Trusted delivery | Artifact versions, independent verification and acceptance decisions planned | Verification |
| Execution environments | Environment templates, Use lifecycle, isolation and resource matching planned | Environments |
| Web & mobile workbench | Shared workbench clients planned; a static WASM preview is published at /workbench/. Other Web pages are demos or task browsing |
Clients |
The intended workflow is to define a task with fixed inputs and checks, assign it to a collaborator or approved worker, then review the recorded result. Compute can stay on a personal machine, a lab server or the cloud; coordination can be self-hosted.
Agents propose work; people approve consequential actions and accepted conclusions. Task intake alone grants no execution permission, and runs retain the evidence needed for review.
Solid boxes mark implemented foundations; dashed boxes and links mark planned capabilities or integration. The end-to-end research workflow remains in progress. See the architecture and development guide for details.
Bring a research use case, contribute a method or App, or review results. Pick a bounded item from the Roadmap; for substantial changes, open an Issue to agree on scope and acceptance criteria. Include a reproducible check with code changes. AI-assisted contributions follow the same review process. See CONTRIBUTING.
