Google's gemini-api CLI gives an existing coding agent a shell route to Gemini API models and managed agents. The agent can inspect a bundled schema and preview the resolved HTTP request before a developer permits an authenticated call. That preview checks the proposed command's shape; account access and actual results remain separate checks.

The v0.4.3 release was published on October 19, 2026. An October 24 feed pickup is discovery of that earlier release, not a new launch. Google describes this as a beta, stateless tool for an agent that already has shell access. It is a separate product from the conversational Gemini CLI, and v0.4.3 is a CLI release rather than a new model.

The big change

  • What changed: The version-pinned README documents a command interface that exposes flags, machine-readable usage, bundled request schemas and an offline dry run. An agent can use that interface to assemble a request for developer review before an API key enters the workflow.
  • Why it matters: A model call, media generation or managed-agent run may involve different inputs, output handling and charges. Previewing the resolved request lets a developer catch an unexpected model, payload or destination before deciding to make the call. This is our reading of the documented interface, not a measured reduction in mistakes.
  • What to watch: The dry run makes no network call and does not read the OS keychain, according to Google. It therefore cannot establish live model access, quota, billing status, latency, safety behavior or result quality. Those require separate account checks and, if justified, an authorized real request.

What the agent can inspect

In v0.4.3, gemini-api --usage emits a machine-readable command tree with flags, defaults, environment variables and configuration keys. Commands that accept a body expose a bundled JSON Schema through --schema. The README's examples use:

Terminal
gemini-api agent run --schema
gemini-api generate "Hello from my agent" --dry-run --output-format json

The first command inspects the local schema. The second validates the supplied command and prints a JSON preview of its resolved HTTP request, with credentials redacted and without contacting the API. These commands are documented examples, not calls BIG CHANGE ran. A schema describes accepted request structure in the installed binary; it is not a live entitlement check. A dry-run preview may itself contain the prompt or other request data, so a developer still has to control where that output is saved or pasted.

Google also documents --no-interactive for headless execution and --agent-mode for structured error envelopes. The CLI distinguishes success, runtime/API failure, invalid usage and authentication/authorization with exit codes 0, 1, 2 and 3 respectively. That separation gives a host agent a way to stop on validation or access errors instead of treating every failure as a model answer. It does not repair a rejected request automatically.

Where credentials and results enter

Google says a developer can supply GEMINI_API_KEY in the subprocess environment or configure a key through the CLI, including an OS keychain option. gemini-api whoami reports the active configuration and credential source. A dry run needs neither credentials nor a keychain read; the actual API call is the later, distinct step. The host agent still needs an explicit permission boundary around that step, especially if it can choose prompts, upload files or launch a managed agent.

For results, the pinned README shows --stream=false --output-format json for a complete structured response and --jq to select a field. Media commands can write a returned asset to a local path; a background managed-agent call returns an interaction ID for later status checks. Those are different output contracts. The host agent must retain the appropriate file path or interaction ID and check the resulting content before using it. The CLI keeps no local conversation state; Google says server-side interactions and uploaded files have their own retention.

The bundled gemini-api models catalog is also different from gemini-api models list --all, which queries models available to the caller's API key. The offline catalog and a clean dry run cannot prove that a specific account has model access. Google's rate-limit documentation says limits vary by model and project tier and tells users to view their active limits in AI Studio. The pricing page separates free and paid access and charges by capability and usage. A developer should check the current model, account tier, price and limits before enabling a paid call; a CLI release does not grant a free allowance for every command.

A useful boundary, with a narrow claim

For a team adding Gemini to an existing coding agent, the useful sequence is to inspect the installed v0.4.3 command and schema, review an offline request preview, then decide separately whether the agent may use credentials and execute the request. A valid request can still send sensitive input, incur a charge, hit a rate limit or return an unusable result.

Google labels the CLI beta and warns that breaking changes may occur between releases. The company recommends pinning a release in automated workflows. This article describes the v0.4.3 documentation and current Google API cost/limit guidance as checked on October 25, 2026; it does not claim a local installation, API call, integration test or observed output.

Sources & further reading

  • Google Gemini API CLI v0.4.3 release — released October 19, 2026, with commit 72f9764 and version-specific changelog. The October 24 feed pickup is not its release date.
  • Gemini API CLI README pinned to v0.4.3 — the primary source for command syntax, schema, dry run, authentication, output behavior and beta status. It documents behavior; BIG CHANGE did not execute the CLI.
  • Google Gemini API pricing — current model and capability pricing and tier distinctions. Prices can change and must be checked for the account and intended call.
  • Google Gemini API rate limits — project, model and tier dependent limits; the account's active limits are shown in AI Studio rather than proved by an offline CLI preview.