Your code never leaves: local LLMs as private coding assistants

Every cloud coding assistant carries the same asterisk: your source code is transmitted to third-party servers. For open-source hobby projects, fine. For proprietary algorithms, unreleased features, or client work under NDA, that asterisk is a dealbreaker - and increasingly, a policy violation.

What local coding assistance looks like

• Code review and refactoring with DeepSeek V3 - within a few points of frontier models on coding benchmarks, running on your GPU.
• A knowledge base over your repos - ask where rate limiting is implemented and get answers grounded in your actual code.
• Shell and file tools with approval gates - the model can run tests and read files, but only what you allow.
• GitHub via MCP - issues, PRs, and reviews handled by agents, with credentials that never leave your OS keychain.

Latency is a feature

Local inference has no network round trip. On Apple Silicon, responses start in under half a second - which changes how you use the assistant. It stops being a search engine you consult and becomes a pair programmer you interrupt freely.

For companies with strict IP policies, local isn't a nice-to-have. It's the only assistant legal will sign off on.

Alongside your editor, not instead of it

Plenty of developers pair a cloud AI editor for open code with Gotchi for everything sensitive: internal repos, incident postmortems, architecture docs. Different trust levels, different tools. Gotchi handles the work that must stay on-machine.

A concrete workflow: reviewing a pull request locally

Here's what the loop looks like in practice. The GitHub MCP server surfaces the open PR; the model pulls the diff and the linked issue. It queries the knowledge base for the module's design doc and prior review comments on the same files. It runs the test suite through the shell tool - gated behind your approval - and reads the failures. Then it drafts a review: two genuine bugs, one style inconsistency against your own conventions, and a suggested refactor with the reasoning spelled out. Total wall time: about ninety seconds. Total data transmitted to third parties: zero bytes.

Fine-tuning on your own conventions

General models write generic code. The next level is a LoRA adapter trained on your codebase's conventions - naming patterns, error-handling idioms, preferred libraries, comment style. On-device fine-tuning in Gotchi makes this a weekend project rather than an ML initiative: export representative code as a dataset, train an adapter on your GPU overnight, and load it into any conversation. The result reads like code your team wrote, because statistically, it is. And the training data - your proprietary source - never touched a cloud GPU.

• Train on merged PRs to teach the model what 'approved' code looks like in your org.
• Adapters are small (tens of megabytes), swappable per project, and reversible instantly.
• Combine a convention adapter with knowledge-base retrieval for both style and factual grounding.

The security review you don't have to have

Adopting a cloud coding assistant at a company with IP sensitivity triggers a predictable review: where is code transmitted, what is retained, is it training data, who can subpoena it, what happens on breach. Every answer is a document someone has to write and defend. The local alternative collapses the review to one verifiable claim: source code is processed on the developer's machine and transmitted nowhere. Gotchi's core is open source, so 'verifiable' means auditable - your security team can read the code path that proves there is no code path.

Honest limits

A local 32B–70B coding model is excellent at review, refactoring, test generation, and explaining unfamiliar code. It is not yet the equal of frontier cloud models at architecting large novel systems from scratch or navigating million-line monorepos in a single context window. Teams that need both keep both - cloud for greenfield exploration on public-safe code, local for everything touching the crown jewels. The boundary is a policy decision your legal team will actually enjoy writing for once.