Local Code Review 2026: How to Run a Private AI Agent on Your Mac Without an API Key
From the makers of Draw Things, Local Code bundles a model engine, a sandboxed toolbox and an agent into one open-source app. Here is what that really takes, and what it does not prove yet.
This Local Code review scores the private Mac agent 7.75 out of 10: clever on-device engineering, held back by hardware needs and thin independent testing. Draw Things, the team behind the on-device image-generation app of the same name, opened the source and released it publicly on October 28, 2026, after a public beta that began on September 12. It is not just a code editor with a local model attached. It packages an inference engine, a coding harness and a sandboxed set of tools such as Python, Node.js and a shell, so an agent can work through whole tasks on folders you choose.
The pitch is easy to repeat and harder to verify: serious AI work that never leaves your Mac. We checked it against the official site, the project's GitHub README and independent coverage, and the picture is more nuanced than the tagline. If you are comparing it with other open-source agents we have covered, such as OpenCode or Kilo Code, the biggest difference is where the model runs and what you must own to run it.
What Local Code Bundles
An inference engine for Apple Silicon
The app downloads its own quantized checkpoints and runs them on your Mac's GPU. For the big mixture-of-experts models it keeps the most-used experts in memory and streams the rest from your SSD, which is how a model larger than your RAM can run at all.
A toolbox compiled into the app
A shell, Python 3.14 with common data and document libraries, Node.js 24, Lean 4 and Z3, LibreOffice, FFmpeg, git and curl, all running inside the app's own process rather than as separate programs on your computer.
Office and data work
Built-in skills create and edit Word, PowerPoint and Excel files, draw charts with Matplotlib, and export to PDF or images. The launch demo asks it to chart a spreadsheet and turn the result into a slide deck.
Images and video
Through the Draw Things command-line tool it generates and edits images and videos, and an HTML-to-video skill makes motion graphics, captions and slideshows.
A browser and your personal data
A WebKit browser the agent can drive, including sites that need a login, plus web search. On a Mac it can edit Calendar and Photos once you grant access, while Notes, Mail and Messages are read-only.
Work that survives interruptions
Threads are saved, interrupted replies resume where they stopped, you can steer the agent mid-task, and the context cache is stored on disk so long sessions are not re-read from scratch.
Which Model Will Your Mac Actually Run?
| Model (quantization) | Input | Download size | Memory, min / recommended (site) | M5 Max 128 GB, 8K context (prefill / decode, tok/s) |
|---|---|---|---|---|
| Qwen 3.8 27B (4-bit S) | Text and images | 21.3 GB (2-bit 9.6, 8-bit 29.2) | 16 / 32 GB | 1,450 / 40.4 |
| DeepSeek 4 Flash 0731 (2-bit S) | Text | 89.4 GB | 32 / 64 GB | 1,361 / 42.3 |
| DeepSeek 4.1 Flash 1017 (2-bit S) | Text and images | 376 GB (3-bit: 412 GB) | 32 / 96 GB | 346 / 18.7 |
The speeds come from Draw Things' own measurements, so treat them as vendor numbers. The pattern is still instructive: the largest model, the one the project is built around, is also the slowest to answer on a top-end laptop. The M5 Ultra figures are far higher for it (4,600 prefill and 58.9 decode on a 256 GB machine), but that reading is a second pass without the usual cool-down, which the README notes openly.
Don't confuse download size with memory. The 376 GB package includes a 203 GB "engram" file, and the weights themselves are about 173 GB at 2-bit. The app caps memory use at your total minus 32 GB, so a 64 GB Mac keeps far less in RAM and leans harder on the SSD. At setup the app suggests Qwen below 32 GB, a 4-bit Qwen from 32 GB, DeepSeek 4 Flash from 48 GB and DeepSeek 4.1 from 64 GB. The developer's own launch post says the big model can run with about 36 GiB resident, while the site recommends 96 GB, so the honest range is wide. Plan on a fast SSD with several hundred gigabytes free, and test before you commit a work week to it.
One more detail from the project's own quantization table: at similar sizes, other community quants of the same model show lower drift from the original than Local Code's 2-bit build. Its edge is speed, not fidelity.
What It Costs
| Option | Price | Notes |
|---|---|---|
| On-device models | $0 | No per-token fees. You pay with hardware, storage and electricity. |
| Your own API keys | Provider's price | The site says there is no markup on keys you bring yourself. |
| Cloud credit | From $9.99 | Prepaid top-ups for hosted models, with no subscription. |
| Teams | Contact the team | Setup, custom skills and on-prem options; no public price. |
"Free" is accurate and still incomplete. The expensive part is a Mac with enough memory and storage for the model you actually want. For a sense of how quickly self-hosting costs add up elsewhere, our notes on running agents fully locally and on the real costs behind free agent tools make the same point from different angles.
How Hard Is It to Get Started?
Installation is the easy part: download the app, let it suggest a model for your memory, and start a task. There are no separate installs, sign-ins, API keys or external servers to wire up, according to the developers.
The learning curve is in everything after that. You choose which folders to share, decide how much to trust a result before it touches a real project, and learn to steer an agent that may work for a long time on a big model that decodes at under 20 tokens a second on a laptop. For coding specifically, expect to review its edits as carefully as you would a junior colleague's. Our comparison of cloud coding agents shows the workflow you are trading against.
Strengths and Weak Spots
💡 Early Reactions
That commenter called a private, subscription-free option a huge win if the claims hold, which captures the mood well. Other readers in the same thread asked a fair question: with so many ways to run models locally, what makes this one different? The developer answered in the thread, pointing to the bundled engine, the in-app execution environment and the Apple-enforced sandbox. He also noted he runs it on a 128 GB M5 Max, so results will differ on smaller machines.
Treat that reply as the vendor's view, not independent confirmation. We found no hands-on review from an outside tester at the time of writing. Coverage from MacRumors confirmed the basics: a free local mode, optional cloud access, and a direct download with an App Store release described as coming soon.
✅ Choose Local Code if...
You have a Mac with plenty of memory and SSD space, you care about keeping prompts and files on your own machine, and you want one app for code, documents and media rather than a code-only editor.
❌ Look elsewhere if...
You use Windows or Linux, have a 16 GB laptop and need top coding quality, or want a mature, widely tested agent whose reliability on large codebases is already well documented.
Expert Editorial Opinion
What makes Local Code interesting is where the trust sits. Most agent tools ask for broad access and ask you to trust the tool. Local Code instead lives inside Apple's App Sandbox, so macOS enforces the boundary and the app has to ask for each folder. That is a smarter design than a promise on a landing page, and it is open to inspection because the code is public.
The evidence for how well it works, though, is thin. The benchmark table shows the quantized DeepSeek model scoring 30 out of 30 on AIME 2026 and 93.4% on GPQA Diamond, but those test the model, not an agent finishing a multi-step refactor. They are also compared with DeepSeek's reported scores for an earlier version, so the comparison is not like for like. We would not take them as proof that it can replace a cloud coding agent.
This is a young product and it shows in the edges. Two stars on the repository, a Mac build that outsiders cannot yet reproduce, and a store listing described as both live and coming soon. None of that is damning for a week-old release. It just means the usual signals of maturity are absent, and the sensible response is a trial on a non-critical project, not a migration.
Now the pricing gap. The app costs nothing, but the experience scales with hardware: the flagship setup wants a high-memory Apple machine and hundreds of gigabytes of fast storage, and the speeds quoted in the README come from top-end chips. If the app charged a subscription, it would be a harder sell than a $20 cloud plan for most people. Because it is free, the right question is whether you already own the Mac or would buy one for this, and the second case is hard to justify on this evidence alone.
If you do own the machine, it is one of the most thoughtful local-agent designs we have seen, and the optional cloud routes give you a safety valve when the local model falls short. Just keep the privacy settings in mind each time you switch models, because that one dropdown decides whether your files stay home.
Final Verdict
At 7.75 it sits in the 7.0–7.9 band: competent but compromised, a good fit for a specific user rather than the majority. That user owns a high-memory Apple Silicon Mac and values keeping work on the device. The compromises are a week-old general release, almost no independent testing, Apple-only support and a flagship model that demands serious hardware. We did not round up for the technical ambition.
| Dimension | Weight | Score /10 | Why |
|---|---|---|---|
| Technical quality | 30% | 8.0 | Strong sandbox design and impressive engine speeds, but agent reliability on real projects is unmeasured. |
| Price-to-value | 25% | 8.4 | Free on-device use and no key markup, offset by the hardware and storage you need to bring. |
| Maturity & documentation | 20% | 6.5 | Detailed README, but a new release with a non-reproducible Mac build and inconsistent store messaging. |
| Ceiling & flexibility | 15% | 8.0 | Open source, local and cloud models, custom skills and MCP support, limited by Apple-only reach and the sandbox. |
| Honesty of positioning | 10% | 7.5 | The README is candid about cloud privacy, though "frontier" and "up to 2x faster" claims lean on self-run comparisons. |
Calculation: (8.0 × 0.30) + (8.4 × 0.25) + (6.5 × 0.20) + (8.0 × 0.15) + (7.5 × 0.10) = 2.400 + 2.100 + 1.300 + 1.200 + 0.750 = 7.75.
Would you still trust it on a 16 GB Mac?
Everything impressive here is measured on machines with 128 GB or more. Before you judge Local Code, check what the smallest model does on the Mac you actually own, on a real task, and decide whether that is a private assistant or a slow demo.
❓ Frequently Asked Questions
🔗 Related ToolRadar Reviews
Official sources: the Local Code website, the GitHub repository and the Draw Things release announcement. Model sizes, speeds and store availability are changing quickly, so confirm them there before you download hundreds of gigabytes.
Comments
Post a Comment