Devika Review Australia 2026: Open-Source AI Engineer With a Maintenance Problem

Verdict (TL;DR): Devika is the most privacy-friendly autonomous AI coding agent available to Australian developers — because it runs entirely on your own hardware. It can take a plain-English task description and write, research, and assemble code autonomously. But "experimental" is the operative word. Development activity has slowed, many features are incomplete or broken, and you will need Python and Node.js experience just to get it running. For production use, it is not ready. For curious developers who want to self-host an agentic coding tool without routing code to a US SaaS provider, it is worth understanding.

Score: 13/25 — ⭐⭐⭐ Good

Last reviewed: 26 June 2026 by Adam.

What Devika is

Devika is an open-source autonomous AI software engineer created by Stition AI, an Indian AI research company. The project launched on GitHub in early 2024 as an explicit open-source alternative to Devin — the USD $500/month autonomous coding agent from Cognition AI.

The concept: give Devika a high-level instruction in plain English ("build a REST API for a to-do list app") and it will break the task into steps, research what it needs using a web browser, write code, and assemble a working implementation — without you writing a single line.

Devika is not a plugin or SaaS product. It is software you clone from GitHub and run on your own machine or server. You supply the API keys to the AI models it uses — Claude 3, GPT-4, Gemini, Mistral, or a local model via Ollama. No data is sent to Stition AI. Your code stays on your hardware.

The project has accumulated approximately 19,500 GitHub stars, reflecting strong initial interest from the developer community when it launched.

What it costs in Australia

Devika itself is free and open-source under the MIT licence. There is no subscription, no hosted version, and no SaaS pricing.

However, you pay for the AI models Devika calls:

• Claude 3.5 Sonnet or Opus (Anthropic): Approximately USD $3–15 per million tokens depending on model. A typical coding task might use 50,000–200,000 tokens. (token consumption varies significantly by task complexity.)
• GPT-4o (OpenAI): Similar range.
• Local model via Ollama: Free, but performance is heavily dependent on your hardware. A capable local coding model (e.g. Qwen 2.5 Coder, DeepSeek Coder) requires a machine with significant RAM — 16GB minimum, 32GB+ for quality results.

In AUD terms, API costs for moderate use are roughly AUD $5–30 per complex task using frontier models. For lighter tasks, much less.

Running entirely via Ollama costs nothing in API fees but demands suitable hardware.

How it works in practice

Setting up Devika requires comfort with the terminal. You need Python 3.10–3.12, Node.js 18+, and the Bun JavaScript runtime. The setup involves cloning the repository, creating a Python virtual environment, installing dependencies, configuring API keys in a config file, and starting both a Python backend server and a Node frontend.

On a modern Mac or Linux machine, this takes approximately 20–30 minutes if you have the prerequisites. On Windows, expect more friction.

Once running, Devika presents a web-based interface at localhost. You type a task description, select an AI model, and click "Start." Devika then:

1. Breaks the task into sub-steps using the AI model's planning capabilities.
2. Searches the web for relevant documentation or examples using Playwright (a browser automation library).
3. Writes code file by file, creating a project structure.
4. Iterates if initial attempts produce errors.

Results vary significantly by model quality. Claude 3.5 Sonnet and GPT-4o produce the most coherent output. Smaller local models often struggle with complex multi-file tasks.

The honest assessment: Devika works for simple, well-defined tasks. A basic web scraper, a CRUD application, a data processing script. For complex real-world projects, it frequently stalls, produces broken code, or fails to handle dependencies correctly.

What it does well

Full local execution. This is Devika's most important feature for Australian developers. When you run Devika with Ollama, no code, no task descriptions, and no output leave your machine. For teams handling sensitive IP, client code under NDA, or data with Privacy Act implications, that is a genuine advantage no SaaS coding agent can match.

Model flexibility. You are not locked to one provider. Claude, GPT-4, Gemini, Mistral, and local models are all supported. If Anthropic's pricing changes or a new model overtakes the current leaders, you can switch.

Free to use. For developers who supply their own API keys, Devika adds zero cost on top of model usage. The MIT licence means you can modify it, extend it, or integrate it into internal tooling.

Agentic loop. Unlike simple code generators, Devika plans, researches, writes, and iterates. Watching it browse documentation to solve a problem it encounters mid-task is a useful demonstration of what agentic AI can do.

No vendor dependency. Devin, Claude Code, and GitHub Copilot Agent all require accounts with US companies and route your code through their infrastructure. Devika eliminates that dependency entirely.

What it doesn't do well

It's experimental, and showing its age. The most recent significant commits to the main GitHub repository were in late 2024. As of mid-2026, development activity has slowed considerably. Several GitHub issues have flagged unimplemented or broken features that have not been addressed. This is a risk for anyone relying on it for ongoing use.

Setup is not beginner-friendly. Installing Devika requires familiarity with Python virtual environments, package managers, and terminal workflows. If you have never set up a Python development environment, expect several hours of troubleshooting.

Output quality is inconsistent. For tasks beyond "simple script," Devika frequently produces code with dependency errors, incorrect logic, or incomplete implementations. You need to be a developer to evaluate and fix what it produces.

No persistent project context. Each task starts fresh. Devika does not maintain memory of previous sessions or build up understanding of your specific codebase over time.

Web research is fragile. Devika's web browsing relies on Playwright. If a documentation site blocks automated browsers, or if search results are unhelpful, the research phase fails and code quality degrades.

Australian fit

Devika's data story is the best of any AI coding agent in this category. When run locally with Ollama, nothing leaves your machine. When run with external API keys, your code and prompts go to Anthropic or OpenAI — US companies — but not to Stition AI. Stition AI is an Indian company and has no stated data processing terms for Australian customers, but since Devika never contacts Stition AI's servers during operation, that is largely irrelevant.

For Australian developers and businesses:

• Self-hosted with Ollama: No offshore data transfer. Maximum control. Fully compliant with any internal data handling policy. Best option for health, finance, legal, or government-adjacent work.
• Self-hosted with Anthropic/OpenAI API keys: Code travels to US servers. Same considerations as using Claude or ChatGPT directly — no Australian data residency, but Anthropic and OpenAI both have enterprise data processing agreements available if required.

There is no AUD pricing (free is free), no Australian support, and no APAC infrastructure. The self-host model makes most of those gaps less relevant.

Australian Fit: 4/5 — The local deployment path makes Devika uniquely suitable for Australian teams with data sovereignty requirements. Docked one point for difficult setup and the absence of any AU-specific support.

Alternatives

• Devin — The commercial product Devika was designed to replace. More capable and actively maintained, but USD $500/month and routes all code through US cloud infrastructure. No self-host option.
• Claude Code — Anthropic's terminal-based coding agent. More reliable than Devika, actively maintained, and supports bring-your-own-key usage. Code goes to Anthropic's US servers. No self-host, but a much better production option.
• Aider — Another open-source terminal coding agent that supports local models via Ollama. More actively maintained than Devika and better at integrating with existing codebases. Worth comparing directly.

Bottom line

Devika proves the concept of an open-source autonomous AI engineer. It is free, model-agnostic, and the only option in this category that can run entirely on your own hardware with zero data leaving your network.

But the honest assessment for Australian businesses is this: Devika is a research project and learning tool, not a production tool. Development has slowed, features are incomplete, and output quality is inconsistent. If your priority is data sovereignty, pair Aider or a local LLM setup with a capable model instead — you will get better results with less frustration.

If you are a developer curious about agentic AI, have the technical skills to set it up, and want to experiment without a SaaS subscription, Devika is worth an afternoon. For anything more than that, look at the alternatives above.

Recommendation: Experiment only. Not production-ready. Use Aider or Claude Code for serious work.

Score

Frequently asked questions

Is Devika free to use in Australia?

Yes. Devika itself is free and open-source under the MIT licence. You do pay for the AI model API calls it makes — typically to Anthropic or OpenAI — or you can run it with a free local model via Ollama, in which case there are no ongoing costs beyond your own hardware.

Does Devika send my code to external servers?

Only to the AI model you choose. Devika never contacts Stition AI's servers during operation. If you run a local model via Ollama, nothing leaves your machine at all. If you use an external API key (Anthropic, OpenAI, etc.), your code and prompts are sent to that provider's servers in the US.

How does Devika compare to Devin for Australian developers?

Devika is free and self-hosted; Devin costs USD $500/month and is fully cloud-based with no self-host option. Devika is more suitable for Australian teams with data sovereignty needs, but Devin is more capable and actively maintained. For production use, Devin or Claude Code are more reliable choices.

Sources

• Vendor page, re-read live 24 September 2026: github.com/stitionai/devika — 200 — “Devika is the first open-source implementation…”. For an open-source tool the repository is the vendor page. (no Wayback capture — Save Page Now still refusing.)
• Devika GitHub repository (primary): https://github.com/stitionai/devika — Checked 26 June 2026
• Devika architecture documentation (primary): https://github.com/stitionai/devika/blob/main/ARCHITECTURE.md — Checked 26 June 2026
• Stition AI GitHub organisation (primary): https://github.com/stitionai — Checked 26 June 2026
• GitHub issue on active development status: https://github.com/stitionai/devika/issues/662 — Checked 26 June 2026 (no change in status had been reported as at June 2026)

AUD API cost conversions: USD × 1.40 (excl. GST). Model API pricing is subject to change; verify current rates at Anthropic and OpenAI pricing pages before use. GST on digital services from foreign suppliers applies to AU businesses — consult your accountant.

The AI Directory Australia is an independent editorial publication. We don't accept payment for positive reviews. Read our Methodology and Affiliate Disclosure.