Teamday
Gemini logo

Harness · Google

Gemini CLI.

Google's agentic harness. The one we reach for when a mission needs a 2M token context window, deep ARC-AGI-style reasoning, or tight integration with the Google stack.

Meet the agents →All harnesses

What it is

Gemini CLI is Google's agentic harness for the Gemini family of models. It pairs the largest context window in the industry with native access to Google Workspace, Drive, Gmail, and BigQuery — the surfaces where most enterprise data already lives.

Models it runs

  • Gemini 3.1 Pro — released February 19, 2026. 2M token context, ARC-AGI-2 score of 77.1%, biggest reasoning leap in the family.
  • Gemini 2.5 Pro — the previous flagship; still strong on cost-per-token.
  • Gemini 2.5 Flash — the fastest, cheapest option for high-volume loops.

What makes it distinct

  • 2M context window. Read entire books, multi-year email threads, or a complete data warehouse schema in one prompt.
  • ARC-AGI-2 reasoning. Strongest on novel puzzle-style problems where pattern memorization fails.
  • Google-stack integration. Workspace, Drive, Gmail, BigQuery, Vertex AI, Calendar — all first-class.
  • Cost-efficient. Some of the lowest per-token pricing among frontier reasoning models.

Capabilities at a glance

  • Sub-agents: yes — built-in (codebase_investigator, cli_help, generalist, browser_agent) and custom agents in .gemini/agents/*.md.
  • Skills: yes — Agent Skills standard; auto-activated; tiered discovery (workspace > user > extensions).
  • MCP servers: yes — stdio, SSE, and HTTP transports; tools, resources, and prompts.
  • Hooks: yes — the richest lifecycle of the bunch: SessionStart, BeforeAgent, BeforeToolSelection, BeforeTool, AfterModel, AfterAgent, SessionEnd, with structured decision returns.
  • Slash commands: yes — built-ins plus custom TOML commands in .gemini/commands/ with {{args}} placeholders.
  • Permissions / sandboxing: Policy Engine (TOML), trusted folders, per-subagent policies, browser-agent domain restrictions.
  • Plugins: yes — extensions via gemini-extension.json bundling MCP servers, commands, skills, hooks, themes.
  • Multi-model: Gemini-only (2.5/3 Flash and Pro, OAuth or API key, Vertex AI). No third-party providers.
  • Sessions: checkpointing saves git snapshots + conversation history before file mods; /restore and /rewind.
  • Surfaces: Ink/React TUI, VS Code companion with native diffing, IDE-integration spec, GitHub Actions.
  • Headless / SDK: yes — -p/--prompt, JSON / streaming JSONL output, @google/gemini-cli-sdk.
  • License: Apache 2.0; github.com/google-gemini/gemini-cli.

How TeamDay uses it

Selecting Gemini CLI as a TeamDay agent's harness unlocks every Gemini model in the dropdown. Pair it with Google Workspace MCP servers for end-to-end automation across docs, sheets, mail, and calendar.

  1. Open an agent → Settings → Harness → Gemini CLI.
  2. Pick the model. Default is Gemini 3.1 Pro for reasoning; Flash for high-volume.
  3. Attach Google MCP servers (Workspace, BigQuery, Search Console).
  4. Run a mission.

When to pick Gemini CLI

  • Giant-context analysis — entire codebases, multi-year email archives, full PDF books.
  • Google Workspace automation — Docs, Sheets, Slides, Gmail, Calendar.
  • BigQuery agents that need both schema-level and row-level reasoning.
  • Cost-sensitive loops at frontier quality.

How it's benchmarked

Gemini CLI is evaluated on Terminal-Bench (tbench.ai) — the standard suite for end-to-end terminal tasks. Useful for comparing Gemini-family models on the same coding workload as Claude Code and Codex.

When to pick something else

  • Claude Code — for long-horizon coding with self-verification.
  • Codex — for unified multimodal and computer-use missions.