Over the past couple of years, AI agents inside the code editor have become a development standard: they write code, edit files, run commands, and explain unfamiliar projects. But almost every popular option is tied to cloud services: your code leaves for someone else's servers, a vendor account is required, and access to frontier models can be restricted for entire regions. For a team that cares about control over its code and the predictability of its tools, that's a risk rather than a convenience.
We built Spektr — an AI coding agent inside the VS Code editor that runs entirely on your side: on local models or on non-US APIs, and your code stays inside your perimeter. This article covers what it is, how it works, and why we built it this way.
What Spektr is
Spektr is an AI agent built directly into the VS Code editor. It doesn't just suggest the next line — it works like a developer: it reads the whole project, plans changes, creates and edits files (showing edits as a diff), runs terminal commands — with your approval at every step — uses the browser, and connects external tools and data through the MCP protocol.
Spektr is based on Cline (Apache-2.0), a mature open-source agent. We forked it and adapted it for sovereign operation: the agent is configured to use models that you control, and it works independently of any vendor cloud.
The name is a nod to the ZX Spectrum, the computer many of us started programming on. The rainbow "flash" in the logo is the same one that was on the Spectrum's case.
Why it matters
The problem with cloud AI assistants isn't the quality of the answers — it's the dependency on someone else's infrastructure.
Service availability. Access to frontier models from large foreign providers can be restricted for entire regions. A service your daily development depends on can stop responding overnight, and that's outside your control.
Data leaves for the cloud. Your code, including proprietary parts, is sent to third-party servers. For many companies this is a non-starter: classified projects, trade secrets, regulatory requirements.
Changing rules. Prices, limits, and terms of use are changed by the service owner without your involvement.
Spektr removes these risks differently: the agent runs on models you control, and your code stays inside your perimeter.
How it works
Spektr consists of three layers.
The agent in your editor
At its core is the agent loop: Spektr gathers project context, forms a plan, proposes changes as a clear diff, applies them once confirmed, runs the required commands and reads the results, and repeats the loop as needed. Anything that changes your project or executes commands requires your confirmation — you always see exactly what the agent is about to do.
The model layer
Spektr is not tied to a single model provider. It works with two sources, and the choice is yours:
- Local models via Ollama. Open models (Qwen, DeepSeek, and others) run directly on your hardware — a workstation or a GPU server. Code and requests never leave the machine. This is the mode for sensitive code and fully autonomous work.
- External non-US APIs. When you need a more capable model and local hardware isn't enough, Spektr calls external providers (for example, DeepSeek or Qwen from Alibaba). These services sit outside the export-restriction bottleneck.
Mix them per task: local for closed code, an external API for heavy jobs.
The sovereignty layer
What sets Spektr apart is that the entire workflow stays inside your perimeter. The agent needs only your models and your code; data stays in your infrastructure, and you fully control the tool.
What's inside
Spektr's stack is fully open:
- VS Code — the editor the agent is built into.
- Cline (Apache-2.0) — the open-source agent that Spektr is based on.
- Ollama — the local engine for running language models.
- Qwen, DeepSeek — open and non-US language models.
- MCP (Model Context Protocol) — the standard for connecting external tools and data sources to the agent.
All components are open-source. No license fees. Data never leaves your infrastructure.
How we distribute it to the team
Spektr is built into a single installable package (.vsix) that is cross-platform — the same file installs on macOS, Windows, and Linux. An employee installs the package in VS Code via "Install from VSIX," with no command line; updates are distributed as the same file.
We keep the built package and the source code in our own repository and file storage (links below); the team gets access from there.
Where it applies
- Closed development perimeters — where code must not physically leave the company boundary.
- Companies under regulatory requirements — finance, public sector, industries with a secrecy regime.
- Regions with restricted access to foreign cloud AI services.
- Cost control — your own hardware and open models instead of a per-seat cloud subscription.
- Fully autonomous (air-gapped) environments — a local perimeter on Ollama with no internet access.
Status and plans
Spektr is already used by our team as an everyday development tool — with its own brand, a sovereign perimeter, and distribution through our own storage.
Where we're taking it next:
- Adjacent domains. The agent has a rich extension layer (rules, skills, workflows, hooks, MCP), so Spektr naturally extends to DevOps, testing, and infrastructure work.
- Our own catalog of models and presets — for the team's typical tasks, so developers don't have to configure everything by hand.
- Updates tracking the open upstream — we take improvements from open-source Cline while preserving the fork's sovereignty.
Links
- Project repository: Source Code
- Build storage (files): Download
Want a sovereign AI perimeter for your team?
We build and deploy solutions like this for specific infrastructure — from model and hardware selection to a ready package your employees install in one click.
What we do:
- Select models and configuration for your tasks and available hardware.
- Set up the sovereign perimeter on your models and hardware.
- Build a branded, cross-platform package and set up its distribution inside your company.
- Hand over a working solution along with a method for maintaining and updating it.
The system runs entirely on your side. No license fees. All components are open-source under open licenses.