Gemini CLI gets serious
Google's Gemini CLI is being positioned as a real AI workspace — 60 free requests/min, a new plan mode, research/browser/generalist agents and an 'Import Memory' feature that defaults to Gemini 3.1 Flash in preview. Separately, LlamaParse paired with Gemini 3.1 Pro reportedly delivers ~15% better accuracy on PDFs and tables. ( )
Gemini CLI’s free-access quotas also include a 1,000 model-requests-per-user daily cap and a separate unpaid API-key tier that is limited to 250 requests per day and 10 requests per minute, according to the CLI quotas page. (geminicli.com) Google’s Plan Mode is turned on by default and forces the CLI into a read‑only workflow where the agent can map dependencies and produce a Markdown implementation plan but cannot modify project files. (developers.googleblog.com) Developers can launch Plan Mode manually with the /settings toggle or start a session in planning mode via gemini --approval-mode=plan, per the Gemini CLI documentation and GitHub docs. (github.com) Preview model routing and selection are exposed in the CLI’s /model command: enabling preview features and updating to the recommended CLI release lets users pick gemini-3.1-pro-preview or fall back to Flash variants when capacity or cost considerations demand it. (geminicli.com) The Memory Import Processor (memport) lets users modularize GEMINI.md with @file.md includes and returns an import tree for debugging; persistent memories are stored in the ~/.gemini GEMINI.md hierarchy and are managed via the CLI’s /memory and save_memory tools. (geminicli.com) (github.com) Google’s developer tutorial shows LlamaParse integrated with Gemini 3.1 Pro for the hard layout parsing step and Gemini 3 Flash for downstream summarization, and it links to a demo GitHub repo and sample code for a financial-assistant workflow. (developers.googleblog.com) LlamaParse’s documentation and the LlamaIndex blog describe the tool as targeted at complex PDFs and table extraction and show it as a supported integration for production ingestion pipelines alongside Gemini models via the google‑genai SDK. (llamaindex.ai) (pypi.org) Gemini 3.1 Pro’s model card and rollout notes cite large benchmark gains (77.1% on ARC‑AGI‑2) and list gemini‑3.1‑pro and Flash/Lite variants as the preview model family available to developers testing high‑accuracy parsing workflows. (deepmind.google) (geminicli.com)
