Rating certificate · AW/26-09/42126 Sept 2026
61of 100
The Register / Coding agents / Gemini CLI
Register entry · Coding agents · updated 26 Sept 2026
Gemini CLI
Gemini CLI is an Apache-2.0 terminal agent that reads and edits code, runs shell commands and uses MCP servers, extensions, skills, hooks and subagents with Gemini models. On 2026-06-18 Google stopped serving free, Google AI Pro and Ultra users, moving them to the closed-source, Go-based Antigravity CLI. It remains usable with Gemini Code Assist Standard/Enterprise licences, Vertex AI / Google Cloud, and paid Gemini API keys, and the repository continues to ship weekly releases.
Deprecated. Google announced on 2026-05-19 that it is 'transitioning our terminal experience from Gemini CLI to the brand-new Antigravity CLI'. From 2026-06-18 Gemini CLI stopped serving requests for free-of-charge users and Google AI Pro/Ultra subscribers (also Gemini Code Assist IDE extensions and Code Assist on GitHub free usage). Access remains for Gemini Code Assist Standard/Enterprise licences, Google Cloud, and paid Gemini / Gemini Enterprise Agent Platform API keys. The Apache-2.0 repo is still actively released (14 stable releases in the last 90 days, last v0.61.0 on 2026-09-23), with Google stating it will keep shipping model updates, bug and security fixes 'for our enterprise customers'. Marked deprecated because the consumer product was retired in favour of a successor; it is maintained, not dead, for paid/enterprise users. Sources: developers.googleblog.com post and GitHub Discussion #27274 (both by Google PM Dmitry Lyalin).
What reviewers and users praise
- Open-source Apache-2.0 codebase that users could audit and customise (DeployHQ review; HN commenters prefer it over the closed Antigravity CLI)
- Large context and fast single-turn responses (DataCamp review)
- Did not refuse security-audit tasks that its successor refuses (HN, SwellJoe)
What they criticise
- Free, Google AI Pro and Ultra users cut off on 2026-06-18; consumer users must move to Antigravity CLI (Google blog)
- Tool-call errors and retry loops on complex multi-file tasks (DataCamp review; multiple HN comments on loops)
- Two serious security issues in 24 months, including a CVSS 10 RCE in headless/CI mode (GHSA-wpqr-6v78-jr5g)
- Service reliability complaints from paid users (HN)
Experts vs the crowdcrowd cooler by 11
Experts56
Crowd44
| Price | Free |
|---|---|
| Use it via | CLI, IDE extension, API |
| Runs on | macos, linux, windows |
| Licence | Apache-2.0 |
| Updated | Evidence collected 26 Sept 2026 |
How the rating was worked out
Weights for Coding agents: adoption 35%, experts 40%, crowd 25%. Pillars without enough evidence are left out and the others re-weighted. The method.
| Adoption & momentumweight 35% | 80 | GitHub stars 107k → 92 · npm weekly downloads 364k → 69 · VS Code / Open VSX installs 1.1M → 55 · 14 releases in 90 days → 100reach × 0.7 + momentum × 0.3 |
|---|---|---|
| Expertsindependent reviews · weight 40% | 56 | 6 independent reviews (1 positive · 5 mixed · 0 negative), averaged toward 50 for small samples. Benchmarks, shown not scored: Terminal-Bench 2.0 No. 27 of 43 (Gemini CLI + Gemini 3.1 Pro) |
| Crowdcommunity sample · weight 25% | 44 | 21 dated posts from hackernews, reddit, github: 7 positive · 4 mixed · 10 negative (positive + ½ mixed) ÷ n, averaged toward 50 for small samples |
| Rating | 61 | high confidence (3 of 3 pillars, 21 community posts). |
Evidence checks. 2 reviews from tool directories, AI-written pages or unnamed authors not counted. 4 items could not be checked at the source (the page could not be fetched, or the quote is from a comment under the linked post); kept and marked “unverified”. How evidence is checked.
Editorial adjustment. Dates for three undated Reddit posts supplied by gapfill-1; assignment follows Reddit's increasing post IDs.
Trust & safety
Seven dimensions graded A–D from documented facts. Overall: B.
B
Permission model
Approval prompts before tool use with auto-edit and YOLO modes; sandboxing (Seatbelt/Docker/Podman) is opt-in, not on by default. Folder Trust feature is disabled by default. source
B
Data access scope
Works in the launched workspace; Trusted Folders scoping exists but is 'disabled by default'; .geminiignore supported. Broad access, documented. source
A
Data storage
Runs locally; prompts go to Google's Gemini/Vertex/Code Assist services per auth method; no required cloud storage of the codebase. source
B
Data retention & training
For the auth paths still supported (paid Gemini API, Vertex, Code Assist Standard/Enterprise) Google states it 'doesn't use your prompts ... or responses to improve our products'; logs kept 'for a limited period of time' for abuse detection with no stated duration. Unpaid API tier terms differ. source
C
Incident history (24 months)
Two serious issues in 24 months, both patched: Tracebit prompt-injection/command-allowlist bypass (reported 2025-06-27, patch released 2025-07-25) and a critical CVSS-10 RCE via headless workspace trust and --yolo allowlisting (CVE-2026-12537, fixed 0.39.1). Regraded from D: Revision 2: serious issues that were each fixed and publicly disclosed grade C; D is kept for unresolved issues or slow or undisclosed fixes. source
B
Compliance
Google Cloud lists SOC 2 and ISO 27001 for 'Gemini products in Google Cloud' (excluding Code Assist on GitHub); no attestation specific to the CLI client itself. Partial. source
A
Transparency
Apache-2.0 source with public GitHub releases and changelogs. source
The experts
6 named independent reviews found; dated reviews from the last 12 months are scored.
- mixed
Tool call errors and retry loops on complex multi-file tasks are a common complaint
Khalid Abdelaty (DataCamp) · 26 Mar 2026 · unverified · datacamp.com - positive
Gemini CLI is the clear winner for budget-conscious developers.
DeployHQ blog (editorial, author not named) · 12 Jun 2026 · deployhq.com - mixed
Gemini was the most ambitious but needed the most babysitting.
Sohail Saifi (Medium) · 8 Mar 2026 · unverified · medium.com - mixed
Gemini is most useful when given very precise instructions. It’s not as valuable for junior devs or vibe coders
Shipyard Team (shipyard.build) · 3 Apr 2026 · shipyard.build - mixed
Both tools produce code of good quality, but Claude Code generates code that is closer to production quality.
Amanda Adoyi (Real Python) · 20 Apr 2026 · unverified · realpython.com - mixed
The same behavior that makes Gemini CLI brilliant for interactive debugging makes it catastrophic for batch execution.
Mahendra Kutare (imaxxs.com) · 19 Jun 2026 · imaxxs.com
The crowd
21 coded posts from hackernews, reddit, github.
7 positive4 mixed10 negative
- mixed
For work, I primarily use Gemini CLI, and the one real awkward part of plan mode is that the harness can't run commands, even exploratory ones.
news.ycombinator.com · 26 Sept 2026 · news.ycombinator.com - positive
anyone who bothers to just do a side-by-side feature comparison can immediately see how many features antigravity is still missing compared to Gemini CLI even today.
news.ycombinator.com · 21 Sept 2026 · news.ycombinator.com - positive
Compared to gemini-cli that they took out behind the woodshed, I hate it.
news.ycombinator.com · 16 Sept 2026 · news.ycombinator.com - mixed
I was able to get good behavior out of Gemini CLI with overriding the system prompt. The default prompt crippled the harness with very poor instructions
news.ycombinator.com · 2 Sept 2026 · news.ycombinator.com - negative
the flash model was unusable for a whole month via gemini CLI. They could not fix it for a whole month and I was a paid customer.
news.ycombinator.com · 2 Sept 2026 · news.ycombinator.com - negative
I tested Gemini CLI while ago, and it was awful tbh.
news.ycombinator.com · 2 Sept 2026 · news.ycombinator.com - negative
antigravity CLI is much better than Gemini CLI now. Gemini would often get stuck in loops before.
news.ycombinator.com · 1 Sept 2026 · news.ycombinator.com - negative
They blew up my interest when the stole my money by cutting me off from Gemini CLI with no explanation or recourse.
news.ycombinator.com · 27 Aug 2026 · news.ycombinator.com
Pricing
As published, checked 26 Sept 2026.
| Plan | Price | Notes |
|---|---|---|
| Gemini CLI software | Free | Apache-2.0; model access billed separately. Free and Google AI Pro/Ultra model access via Gemini CLI ended 2026-06-18. |
| Gemini API (paid, pay-as-you-go) | Custom | Per-token Gemini API billing; still supported for Gemini CLI. |
| Gemini Code Assist Standard / Enterprise | Custom | Per-seat organisation licence; 1,500 / 2,000 requests per user per day per the repo's quota doc. Seat price not verified. |
Recent changes
- 23 Sept 2026
v0.61.0 stable: sandbox filesystem hardening and prompt-injection fixes. source
- 18 Jun 2026
Stopped serving free, Google AI Pro and Ultra users; they move to Antigravity CLI. source
- 19 May 2026
Google announced the transition from Gemini CLI to the new Go-based Antigravity CLI at I/O. source
- 24 Apr 2026
Critical advisory GHSA-wpqr-6v78-jr5g fixed; headless mode now requires explicit folder trust. source
Ranked alongside
Other ranked entries in Coding agents.
OpenCodeAnomaly75
Claude CodeAnthropic73
Zed (Agent Panel)Zed Industries73
GitHub CopilotGitHub (Microsoft)72
Codex (CLI and app)OpenAI71
Sources
Every figure above traces to one of these, observed up to 26 Sept 2026.
- An important update: Transitioning Gemini CLI to Antigravity CLI (Google Developers Blog, 2026-05-19, Dmitry Lyalin and Taylor Mullen)
- GitHub Discussion #27274 announcing the transition (82 comments)
- The Register: Bye-bye, Gemini CLI; Google nudges devs toward Antigravity (Brandon Vigliarolo, 2026-05-20)
- The New Stack: Google pushes Pro, Ultra, and free users from open-source Gemini CLI to closed-source Antigravity CLI (headline only; body not retrievable)
- Repo quota and pricing doc (still describes free tier; stale vs the announcement)