Last updated · Santosh Kumar Radha

CodeAF vs OpenCode vs Gemini CLI: open terminal agents compared

CodeAF, OpenCode and Gemini CLI are all open-source terminal coding agents. How they differ on models, workflow, cost and benchmark results.

CodeAF, OpenCode and Gemini CLI are all open-source terminal coding agents: OpenCode is the broadest single-agent tool for switching providers, Gemini CLI is Google's agent (now aimed at enterprise Gemini Code Assist users), and CodeAF by AgentField runs a crew of models on every task and many tasks from one window. Claude Code, often compared with all three, is proprietary and built for Claude models.

Where you sit decides it. If you want one agent with every provider, a desktop app and by far the largest community of the three (about 228k GitHub stars on 2026-11-21), OpenCode is the natural pick. If your company already has Gemini Code Assist, Gemini CLI is the supported path. If you want handed-off work to run as parallel tasks with a planner and a checker, CodeAF is built for that.

At a glance

CodeAFOpenCodeGemini CLIClaude Code
LicenceApache 2.0 (repo)MIT (repo)Apache 2.0 (repo)Proprietary (LICENSE.md)
ModelsOpenRouter, DeepSeek, GLM, Kimi, MiniMax, Qwen, Codex plan, Ollama, any OpenAI-compatible endpoint (FAQ)The models.dev catalog plus custom OpenAI-compatible providers (providers)Gemini models; README says it is not tuned for Gemini 4 (README)Claude only (LLM gateway)
Local modelsOllama, llama.cpp, LM StudioOllama, LM Studio, vLLMNot documentedNot documented
CostFree; pay your providerOpen source; optional OpenCode Go subscription and pay-as-you-go Console (docs)Code Assist Standard or Enterprise, or paid Gemini API keys; no free tier (README)Claude plan or Console account (setup)
InterfaceTerminal; no IDE extension documentedTerminal, desktop app (beta), web app, IDE extension (IDE)Terminal, VS Code integrationCLI, desktop, IDEs, web
WorkflowTasks in isolated copies, crews of three seats, teams with managersbuild and plan agents, a @general subagentOne session with subagents (subagents)Subagents, background sessions
Headless / CIcodeaf do "..." --json (headless)opencode run "...", GitHub Action (CLI, GitHub)gemini -p "..." --output-format json, GitHub Actionclaude -p
MCP/connect to over a hundred servicesLocal and remote servers (MCP)Via ~/.gemini/settings.jsonMCP client
InstructionsCLAUDE.md, AGENTS.md, skills foldersAGENTS.md, falls back to CLAUDE.md (rules)GEMINI.mdCLAUDE.md, AGENTS.md
DeepSWE (same model)62 of 11330 of 113 (OpenCode 1.18.30)Not tested16 of 113

Versions checked 2026-11-21: CodeAF v0.8.3 (2026-11-19); OpenCode v2.3.2 (2026-11-18), with the 1.x line still maintained at v1.19.6; Gemini CLI v0.70.0 (2026-11-19).

OpenCode: one open agent, every provider

OpenCode is the open-source agent engines recommend first when someone asks for "any model", and with reason. Its provider list comes from the models.dev catalog, it discovers local models in Ollama, LM Studio and vLLM, and you can add any OpenAI-compatible API. It runs as a terminal interface, a desktop app or a web app, and has an extension for VS Code, Cursor, Windsurf and VSCodium. Tab switches between a build agent with full access and a plan agent that is read-only and asks before running bash. It reads AGENTS.md and falls back to CLAUDE.md, and opencode github install sets up its GitHub Action.

Gemini CLI, and what changed in June 2026

Google's transition post says that on June 18, 2026, Gemini CLI stopped serving Google AI Pro and Ultra subscribers and people using it free of charge through Gemini Code Assist for individuals, and that it "will remain accessible via paid Gemini and Gemini Enterprise Agent Platform API keys". The README now describes it as "Enterprise Code Assist only" and points free and Google AI plan users to Antigravity CLI (agy), which runs multiple agents in parallel. Gemini CLI itself is still Apache 2.0 and still has a clean headless mode: gemini -p with --output-format json or stream-json, and the google-github-actions/run-gemini-cli action. We did not verify Antigravity CLI's licence or pricing, so this page does not compare it.

CodeAF: a crew per task, many tasks per window

CodeAF is an open-source coding agent built to run a software team of agents from one terminal. Work you hand off becomes a task in its own copy of your folder, on its own branch, so your checkout does not move while it runs. Each task gets a crew of three seats: a worker that does the work, a planner and a checker, each on the model that fits the task, and /crew models open limits every seat to open-weight models (README). On top of tasks sit three pieces that are CodeAF's own:

  • Specialists. /senior-dev for large coding changes, /sec for local security audits and /review for GitHub pull requests (specialists).
  • Teams. A named group of conversations with a manager, an agent whose job is the other agents (teams).
  • Standing orders. Schedules and watches that run from a 5-minute timer with no window open (keeps working).

CodeAF is one Go binary for macOS, Linux and Windows; the specialists are left out of Windows builds. It is an early preview. For model setup across all of these tools, see a coding agent that works with any model.

The DeepSWE row, and why it says OpenCode 1.x

On our DeepSWE comparison, every harness ran the full 113-task set on one model, deepseek/deepseek-v4-flash-0731 through OpenRouter, with official verifiers and 3 hours per task. CodeAF in senior-dev mode solved 62 of 113 (54.9%) at 22 cents a task, first of ten harnesses in that comparison. OpenCode 1.18.30 solved 30 (26.6%) at 50 cents a task, 4.8x CodeAF's 39.9 cents per solved issue (method page). Task by task, CodeAF solved 40 tasks OpenCode missed, and OpenCode solved 8 that CodeAF missed (head to head).

That was the latest production 1.x release in September 2026. The method page says plainly that the row does not measure OpenCode 2, which is now the current major version. It is our own comparison, one pass per harness, and CodeAF used the provider's default sampling while the others sent temperature 1.0 and top-p 0.95. Gemini CLI was not in it.

When to choose OpenCode

Choose OpenCode if you want one open-source agent that reaches the most providers, a desktop and web app as well as the terminal, an IDE extension, and a large community. Its plan and build agents make a simple, clear loop for interactive work.

When to choose Gemini CLI

Choose Gemini CLI if your organization has Gemini Code Assist Standard or Enterprise or Code Assist for GitHub on Google Cloud, or a paid Gemini API key, and wants Google's supported terminal agent. If you are an individual on a free or Google AI plan, Google now points you to Antigravity CLI instead.

When to choose Claude Code

Choose Claude Code if you want Claude models with Anthropic's first-party tool and its desktop, IDE and web surfaces. CodeAF vs Claude Code covers that pairing in depth.

When to choose CodeAF

Choose CodeAF if you want handed-off work to run as parallel tasks with a planner and a checker on each, a different model per seat, budgets at every level, and specialists for large changes, audits and reviews. CodeAF vs Codex CLI adds OpenAI's agent and Aider to the same comparison.

Questions people ask

OpenCode vs Claude Code vs Gemini CLI: which should I use?

Use Claude Code if you want Claude with Anthropic's own tool, OpenCode if you want one open-source agent across many providers and local models, and Gemini CLI if your company has Gemini Code Assist. CodeAF, also open source, fits when you want many tasks running at once, each on a crew of models.

Is Gemini CLI still free?

Not for individuals. Since June 18, 2026, Gemini CLI no longer serves free Gemini Code Assist for individuals or Google AI Pro and Ultra users; Google points them to Antigravity CLI. Gemini CLI remains open source and works with Code Assist Standard or Enterprise and paid API keys.

Is OpenCode open source, and which licence does it use?

Yes. OpenCode is MIT-licensed in the anomalyco/opencode repository. CodeAF and Gemini CLI are Apache 2.0, and Claude Code is proprietary.

Did the DeepSWE comparison test OpenCode 2?

No. It tested OpenCode 1.18.30, the latest production 1.x release when the runs started in September 2026, and the method page says the row does not measure OpenCode 2. OpenCode's current release is v2.3.2.

Does CodeAF have a desktop app like OpenCode?

No. CodeAF runs in the terminal, on your machine or on a remote one with codeaf chat --host devbox, and you can join the same conversation from a phone over ssh. OpenCode offers a desktop app (beta) and a web app.

Sources

Last updated 2026-11-21. Checked against CodeAF v0.8.3, OpenCode v2.3.2 and Gemini CLI v0.70.0; the DeepSWE OpenCode row is version 1.18.30.