opencolab

Generate accurate, readable D2 block diagrams for software, research, and system explanations from a text brief. Normalize components and edges, write a canonical .d2 source file, render compact sketch-style SVG and optional PNG artifacts by default, support optional LaTeX equation blocks when the diagram genuinely needs them, keep arrows unlabeled unless a label adds concrete meaning, and emit Telegram file directives when needed.

Create, edit, Git-version, compile, and deliver scientific LaTeX papers, reports, and research-derived PDF summaries. Use for venue-aware templates such as ICLR, NeurIPS, ICML, CVPR, ACL, ACM, and IEEE; experiment-result tables; architecture figures or diagrams; deep-research/pro-research/fast-research to PDF summaries; and .tex to PDF builds with latexmk.

Generate and edit images with Google Gemini Nano Banana Pro from prompts and optional reference images, using GEMINI_API_KEY from .env.local when available.

Extract and return figures from already-downloaded local PDFs with PyMuPDF, optionally reusing PageIndex artifacts for page selection and verifying shortlisted candidates multimodally before delivery.

Automated browser control via the `browser-use` CLI. Navigate pages, fill forms, click elements, take screenshots, and extract content from the command line. Used by `apply-to-job` to drive real ATS portals end-to-end while the human watches.

Author, validate, run, and oversee OpenColab project workflows that coordinate multiple agents through bounded loops, decisions, and human gates.

Deep scientific investigation with papercli. Iterative search, broad PDF corpus download and reading, equation-level analysis, and exhaustive referenced markdown findings.

Fast scientific paper scouting with papercli. Search, download, read, and produce a referenced markdown findings file with core ideas, concepts, and key math.

Local-first grounded follow-up QA over already-downloaded papers. Build and cache per-paper PageIndex trees, keep paper selection bounded, and answer with exact paper or page references when the evidence supports it.

Deterministic PDF-to-markdown paper summarization for papercli workflows. Given one paper PDF or a directory of paper PDFs, produce schema-conformant markdown summaries and optionally update summarized_ids.txt.

Professional paper research with papercli. Multi-pass search, PDF download and reading, math-aware synthesis, and a detailed referenced markdown findings report.

Iterative keep/discard experiment workflow over one explicitly configured repo, editable file, run command, and metric rule. Any agent may use it, but the `autoresearch` specialist is the default owner for sustained experiment-loop work.

Default Runpod workflow for user-managed Pods. Ask the human to create a Pod with the desired GPU, get the `pod_id` and SSH access details, save a manual Pod profile, and use `opencolab gpu ssh` sessions as the default control path. Use OpenColab `gpu server` and `gpu job` only when the user explicitly wants managed provisioning or `run_id` tracking.