To choose an AI coding agent, weigh eight things: whether it acts or only advises, where your code and API keys go, which models it supports, what it can actually touch, how much control you keep, whether the model is filtered, the price, and your team's needs. This guide turns that into a checklist you can score any tool against.
In a year the field went from autocomplete to full agents — Cursor, GitHub Copilot, Windsurf, Claude Code, Cline and a dozen more. On a landing page they look alike; in daily use they behave very differently. The questions below are the trade-offs that actually bite later, roughly in the order they matter.
1. Agent or assistant?
The first fork. An assistant suggests — autocomplete and chat answers you copy-paste. An agent acts — it runs commands, edits files across your repo, and closes the loop without you shuttling text back and forth. If you want something that fixes the bug and runs the tests, you want an agent. If you just want smarter autocomplete, an assistant is enough and usually cheaper. Ask one question: does it execute, or only advise?
2. Where does your code go?
Most cloud tools send your prompts — and often the surrounding code — to someone else's servers, plus telemetry. For a hobby project that may be fine; for proprietary or regulated code it's often a dealbreaker. Check three things:
- Model location — does it run in the cloud or on your machine?
- What leaves the device — prompts, files, telemetry?
- Where your API keys live — on-device, or uploaded to a vendor?
A tool that keeps everything local with zero telemetry is a completely different risk profile from one that streams your repo to a server.
3. Which models can it use?
Lock-in is the quiet cost. Some agents only work with the vendor's own model. Better ones are model-agnostic: any OpenAI/Anthropic-compatible endpoint, or a local runtime like Ollama, LM Studio, llama.cpp or vLLM. Model choice means you can switch when a better or cheaper model ships, run fully offline, or mix a local model for private code with a hosted one for everything else.
4. What can it actually touch?
"Agent" is a spectrum. Map the real capabilities to your workflow:
- Shell / PowerShell and file edits across the whole repo
- SSH into servers, query databases
- Drive a browser, run web searches
- MCP connectors to plug in external tools and data
- Automation — scheduled/cron agents, memory, sub-agents
The more of your actual loop it can run, the less you babysit. A tool that only edits the open file is far narrower than one that can run migrations and deploy.
5. Safety and control
Power cuts both ways: an agent that runs shell commands can also run the wrong one. Look for explicit control — approval modes (auto-run vs ask-first vs plan-only), a visible log of every action and its result, and a scoped local API so nothing else on your machine can drive the agent. One subtle but important check: does it redact secrets (.env, tokens, private keys) from tool output before they reach the model? If not, your secrets ride along in every prompt.
6. Filtered or unfiltered?
Some models refuse legitimate work — security research, scraping, low-level code, "unsafe" shell — because of built-in guardrails. If you keep hitting refusals, that's friction you can't configure away on a filtered model. Decide up front whether you need an unfiltered model that works out of the box with no refusals, or whether the guardrails rarely touch what you build.
7. Price and licensing
Compare the real shape, not the headline number: is there a genuinely usable free tier; is it flat-rate or per-seat; and does the paid plan gate the things you actually need — speed, a specific model, concurrency? A flat local tool with a real free tier has a very different total cost than per-seat cloud pricing multiplied across a team.
8. Team needs (if any)
Solo? Skip this. For a team, check for a fleet/admin view (see installs, versions and runtimes in one place), a token-authenticated API to drive machines from scripts, and — critically — that team reporting never ships chat, prompts or keys to a central server.
A quick scorecard
Score each candidate 1–5 on: acts vs advises · local/private · model choice · capabilities · control · filtering · price (· team). The winner is rarely the loudest landing page — it's the one that fits where your code is allowed to go and what you need it to do.
Where NoTrack Coder fits
NoTrack Coder sits at the privacy-and-control end of this checklist: it runs locally on Windows with zero telemetry, acts (shell, files, SSH, databases, browser), works with any model — local runtimes or hosted, including a built-in uncensored one — redacts secrets before they reach the model, and adds approval modes plus a Fleet admin for teams. It's one honest option among several; see how it stacks up on the comparison pages, or read the docs.
Want to try the privacy-first end of the spectrum? Download for Windows or compare plans on the pricing page.
Frequently asked questions
What is the difference between an AI coding agent and an AI assistant?
An assistant suggests code — autocomplete and chat you copy-paste. An agent executes: it runs commands, edits files and completes tasks end to end. Agents save more time but need more control.
Are AI coding agents safe to let run commands?
They can be, with the right controls: approval modes (ask-first or plan-only), a visible log of every action, a scoped local API, and secret redaction. Avoid tools that auto-run everything with no audit trail.
Can I use an AI coding agent without sending my code to the cloud?
Yes — pick a model-agnostic agent that runs a local model (Ollama, LM Studio, llama.cpp, vLLM) and keeps your keys and context on your machine with no telemetry.
Do I need a paid subscription to use an AI coding agent?
Not always. Some agents have a genuinely usable free tier; paid plans usually unlock speed, specific models or concurrency. Check what the free tier actually allows before paying.
Which AI coding agent is best for privacy?
The one that runs the model locally, sends no telemetry, keeps API keys on-device, and redacts secrets before they reach the model. That is the profile to look for when privacy is the priority.