開源工具
Open Science v0.12.1 Keeps Research-Agent Compaction, Execution, and Artifact Evidence in the Local Workspace
AIPOCH’s Apache-2.0 desktop workbench integrates model agents, Python and R kernels, scientific data connectors, and immutable artifact versioning. The August 9 update makes context compaction visible and fixes Codex MCP approval and Windows notebook execution issues, but still falls short of fully portable replay.
AIPOCH released Open Science v0.12.1 on August 9, continuing its effort to bring research agents, code execution, and chains of evidence into a single local workspace. This Apache-2.0 Electron application runs on macOS, Windows, and Linux, with a stack comprising React, TypeScript, Prisma/SQLite, and an ACP-compatible agent execution layer. Users can connect Claude Code, OpenCode, Codex, cloud model providers, or custom gateways instead of being locked into a single model.
Its main technical contribution is not simply another chat interface. With approval, agents can read and write files, execute shell commands, maintain Python and R notebook kernels, search the web, and invoke 24 scientific data connectors. Its 18 built-in research skills focus on bioinformatics and drug discovery, including AlphaFold2, Boltz, DiffDock, ESM, OpenFold3, scGPT, and remote computation over SSH. Every output artifact is stored as an immutable, checksummed version, and the interface links it back to the code that generated it, execution logs, inputs, environment manifests, and conversation branches. Missing evidence is marked as unavailable rather than reconstructed by the model.
v0.12.1 adds a session sidebar grouped by recent activity and date, displays context-compaction events directly in transcripts, and improves keyboard-based search. Fixes cover Codex MCP approvals, the presentation of agent questions to users, stalls in the artifact-completion workflow, and crashes in the Windows notebook runner. For long-running research agents, knowing when context was compacted matters because compaction can cause a model to lose assumptions, data fields, or approval conditions. Including these events in an auditable record is therefore more useful than merely displaying token usage.
However, local-first does not mean that data never leaves the device: model requests, web searches, and remote connectors still send required parameters to external services. The project also does not yet provide deterministic reconstruction, portable environment restoration, or complete session replay. Its built-in reviewer remains a secondary audit performed by an agent using its own records and cannot replace statistical review, clinical validation, or human reproduction.