aginti-labcanvas

AgInTi LabCanvas is a local control plane for agent-assisted scientific visuals and app automation. Its stated goal is to keep generated figures editable: an overview image may start an idea, but final outputs are rebuilt from atomic parts, scene specs, CAD files, manifests and tool-specific artifacts. Core workflow 1. Start from chat or a saved JSON scene spec. 2. Generate overview concepts through AgInTi image payloads or another image backend. 3. Split the figure into editable atoms such as panels, icons, labels, CAD parts, renders and TeX assembly layers. 4. Route work to BioRender for academic assets, OpenSCAD for mechanical layout, Blender for 3D setup renders and KiCad for PCB artifacts. 5. Keep every artifact in a canvas manifest so later chat edits can target one part instead of flattening the whole figure. Documented capabilities - Workspace agent: persistent direct chat, dynamic low-to-xhigh model routing, durable tasks, cancellation and artifact return across lab tools. - Web studio: agent chat, room-based persistent sessions, artifact canvas, backend settings and a multilingual UI; `/rooms` supports persistent conversations, with the local owner getting the full tool set and invited participants restricted to plan-only turns. - Presentations: manifest-driven editable PPTX, selective visual assets, generated-image policy gates and PDF/PNG preview inspection. - Paper figures: exact NxM SVG grids, AgInTi image dry-run payloads and editable artifact manifests. - Grant projects: durable goal workspace, traceable evidence, editable BioRender/SVG/TeX figures, checked LaTeX/PDF and chat delivery. - Protein structures: a reused ProteinStructure AlphaFold Server pipeline with a persistent logged-in CDP/noVNC browser, downloads, metrics, plots and screenshots. - 3D setup renders: JSON scene specs to Blender PNG and `.blend` output. - CAD devices: OpenSCAD exports and a C-mount reflector adapter CAD example. - Board/CAD tasks: shared CLI and web-chat workflow for KiCad, OpenSCAD, renders and manufacturing prep. - PCB manufacturing: KiCad HYBEC and Lumileds boards, DRC/ERC checks and JLCPCB Gerber ZIPs. - LabVIEW automation: Linux install probe, MCP candidate research and a stdio-to-HTTP bridge. - Android control: a dedicated noVNC/scrcpy desktop and ADB wrapper for a Mi MIX 2S device plus a matching AVD profile. - WeChat chatops: isolated Linux GUI, local message mirror, per-chat media isolation, save-only default for unaccompanied videos, serialized text/file delivery, bounded recovery and CAD/PCB/Blender artifact return. - WeCom bridge: official AI Bot and `wecom-cli` transports, allowlisted GUI/Android relays, voice/media transcription, idempotent file sends, per-chat sessions, member archives and LaTeX report recovery. - Video handoffs: story drafting and browser generation with MP4 download for WeChat workers; a separate publish handoff delegates subtitle correction, metadata, logo/subtitle burn, packaging and posting to LazyEdit/AutoPublish. - Social content agent: persistent campaigns, source-grounded platform drafts, SQLite history, Postiz/X MCP adapters and exact-content publication approvals. - App routing: Blender, BioRender, Unity, Unreal and custom target dispatch. Interfaces and architecture Entry points include a CLI (`labcanvas`, with `app-auto-action` and `agenticapp` as compatibility aliases), a web studio, an MCP-ready bridge and a persistent web chat. The README shows source-checkout usage via `PYTHONPATH=src python -m agenticapp ...` and notes that the npm package has been renamed to `@lazyingart/labcanvas`, pending a fresh authenticated publish. Every target dispatch receives a reviewable JSON envelope containing target, kind, instruction, payload and metadata. A `configs/targets.example.json` file can be copied to `labcanvas.targets.json` for local ports, commands and tokens; the override file is git-ignored. Model policy is configured in `configs/model-policy.json`, with AgInTiFlow described as the primary persistent backend, DeepSeek first and LocalLLM as a same-session provider handoff; Codex and Claude Code are explicit opt-in compatibility backends. Validation and security Validation commands include `npm test`, `npm run pack:dry-run` and `PYTHONPATH=src python -m agenticapp doctor`. The README advises keeping transport behavior covered by tests before adding live editor features and reviewing SECURITY.md before enabling live dispatch to editor bridges or browser sessions.