Key takeaways
- An open-source agent is an open harness around a model: you can audit what it executes, but the model itself is usually a hosted, paid service.
- For iOS work the agent must run on macOS; container-based agents running Linux cannot use xcodebuild.
- Project instructions, MCP support and sandboxing differ more between tools than the models do – compare those first.
- Open source helps with auditability and vendor independence; it does not replace version pinning, token scoping and reviews.
Open agent, hosted model: what is actually open
A coding agent has three layers, and "open source" usually describes only one of them:
- The harness – the program that reads your files, plans, runs commands, asks for approval and talks to the model. This is what Codex CLI, Gemini CLI, Cline, Aider, OpenHands and goose publish as open source. You can read how commands are sandboxed, what telemetry is sent and how context is built.
- The model – usually a hosted model from Anthropic, OpenAI, Google or another provider, reached through an API key or a subscription. Some tools also work with open-weight models running on your own hardware.
- The protocols – MCP for connecting tools and data, AGENTS.md for project instructions. Both, together with goose, were contributed to the Agentic AI Foundation under the Linux Foundation in December 2025, which makes them a safer long-term bet than any single vendor's format.
So the honest summary is: open-source agents give you transparency and choice, not a free model. The model bill and the data processing terms of your provider still apply.
The main open-source agents at a glance
| Tool | Form | Models | Project instructions | Worth knowing |
|---|---|---|---|---|
| Codex CLI (OpenAI) | Terminal agent | OpenAI models via ChatGPT sign-in or API key | AGENTS.md | Built-in sandbox, approval modes, MCP, non-interactive codex exec |
| Gemini CLI (Google) | Terminal agent | Gemini models | GEMINI.md, file name configurable | MCP support, extensions |
| Cline | VS Code extension | Bring your own key; many providers and local models | .clinerules | Plan and Act modes, approval per action, MCP |
| Aider | Terminal pair programmer | Many providers and local models | Conventions file loaded with --read | Git-native: each change becomes a commit |
| OpenHands | Agent platform, local or self-hosted | Many providers | Repository instructions | Agents work in a sandboxed container |
| goose (Block) | Desktop app and CLI | Many providers | .goosehints | Built around MCP extensions; part of the Agentic AI Foundation |
The list keeps growing – OpenCode, Roo Code, Kilo Code, Continue and others have active communities. The criteria below matter more than the names.
Talk to an AI Code Engineer about shipping your app to the App Store and Google Play.
In a free call we look at what you built, tell you honestly what stands between it and an approved release – and whether you need us at all.
What matters for iOS and Android work
Where it runs. iOS builds need Xcode, so the agent must execute on macOS. Terminal agents and editor extensions on your Mac qualify. Platforms that run agents in Linux containers – OpenHands' default sandbox, for example – can write Swift but cannot build or test an iOS app; for Android, Flutter and React Native code they can do much more. See Codex for SwiftUI for how to work around the gap.
Command handling. Mobile builds are long and noisy. Check how a tool handles commands that run for minutes, whether it truncates logs sensibly, and how its sandbox treats writes outside the project – Xcode's DerivedData and Gradle's cache both live in your home folder by default.
Approvals and sandboxing. Codex CLI ships OS-level sandboxing; Cline asks per action; Aider relies on git to undo; OpenHands isolates in containers. Pick the model that matches how much autonomy you want to grant.
MCP. Most of these tools speak MCP, so the same servers – Xcode's bridge, device automation, the Dart and Flutter server – work across them. Our overview of MCP servers for mobile development lists the useful ones.
Instructions. Every tool has its own memory file, which leads to drift. Keeping one AGENTS.md and pointing the other tools at it is described in CLAUDE.md and AGENTS.md for mobile projects.
Local and open-weight models: when they make sense
Tools like Cline, Aider, OpenHands and goose can talk to models running on your own machine or server. That is attractive when code must not leave the company network, when you work offline, or when you want predictable costs.
The trade-offs are real. Large models need serious hardware, and smaller models that fit on a laptop are noticeably weaker at multi-step tasks – planning a feature across Swift files, keeping Gradle builds green, following a long memory file. A sensible approach is to test before you commit: take five real tasks from your backlog, run them with a hosted model and with the local candidate, and compare results and review effort. Many teams end up with a split – local models for small, well-defined tasks such as tests and documentation, a hosted model for feature work.
Governing open-source agents in a company
Open source makes governance possible; it does not do it for you.
- Pin versions and install from official sources. Agents are themselves npm, pip or Homebrew packages, and a compromised release would run with your developer's permissions.
- Review defaults such as telemetry, auto-approval settings and which commands run without asking.
- Scope keys. Bring-your-own-key tools need API keys; issue them per developer or per project, with usage limits and a rotation plan.
- Check the data path. With a hosted model, your code goes to the provider under its API terms. Read the data retention and training settings of the provider you choose.
- Check licences. Most of these projects use permissive licences, but read each repository's licence before you fork or embed it.
Our article on the security risks of agentic coding goes deeper into prompt injection, secrets and permissions.
How to choose: four common team situations
| Situation | A good starting point |
|---|---|
| Solo developer on a Mac, shipping iOS and Android | A terminal agent such as Codex CLI or Gemini CLI, with one AGENTS.md and the Xcode MCP bridge |
| Team in VS Code that wants model choice and tight control | Cline or one of its forks, with approval per action and shared rules files |
| Developers who think in small commits | Aider, where every change lands in git history |
| Platform team that wants self-hosted agents working on tickets | OpenHands for Android, Flutter and backend work; a macOS runner for anything that needs Xcode |
Mixing open-source and commercial tools is normal. Plenty of teams run Claude Code or Cursor day to day and keep an open-source agent for scripted jobs. The comparison of agentic coding tools for mobile apps covers the commercial side.
If you want a neutral view on what fits your team, our AI Code Engineers run tool-agnostic workshops – remote or on-site, in English or German – and the enterprise programme adds governance and rollout planning.
Evaluating an open-source coding agent
- It runs where your builds run – on macOS if you ship iOS.
- It supports your model provider and its data terms are acceptable.
- It reads a project instruction file you can share with other tools.
- It supports MCP for the servers your team needs.
- Its sandbox or approval model fits the autonomy you want to grant.
- You have tested it on real tasks from your backlog, not on demos.
- Versions are pinned, keys are scoped and telemetry settings are reviewed.
Frequently asked questions
Are open-source coding agents free to use?
The tools are free to download and run. The model behind them usually is not: you pay for API usage or need a paid plan with the provider – unless you run an open-weight model on your own hardware.
Is an open-source agent safer than a commercial one?
It is more transparent, not automatically safer. You can audit what it does, but risks like prompt injection, leaked secrets and over-broad permissions apply to every agent. Sandboxing, approvals and reviews matter more than the licence.
Can I use Claude models with open-source agents?
Yes, many open-source agents – including Cline, Aider, OpenHands and goose – support Anthropic's API alongside other providers. Check each tool's documentation for the current list.
Can you help us pick and roll out a tool?
Yes. Book a free call with an AI Code Engineer. We look at your stack, your security requirements and your release process and recommend a setup – or run a hands-on workshop on your own codebase.
Free, 15 minutes, no obligation
Get your app from agent output to approved release.
Our AI Code Engineers take apps built with Claude Code, Codex, Cursor, Lovable, Base44 and co. through signing, CI/CD, store listings and review – in your own developer accounts.
AI tools, store policies and requirements change quickly – always check the current documentation of the tools and stores you use. This article is independent advice and not affiliated with or endorsed by Apple, Google, Anthropic, OpenAI or any other company or product mentioned; all trademarks belong to their owners.