Methodology
How We Test AI Coding Tools
Every tool we cover is tested the same way: real projects, the same tasks for all of them, our own repositories, and the actual command output — commands, version strings and the date we ran them. This site carries no affiliate links at all, so nothing commercial can move a score.
Last updated: September 4, 2026
01 / Scoring
What we score, and how much it counts
Each tool gets a weighted score across six dimensions. The weights reflect what matters when you are actually building, not what is easiest to demo.
| Dimension | Weight | What we look at |
|---|---|---|
| Code quality & correctness | 25 | Does the output compile, run and actually solve the task on a real repo? |
| Agent autonomy | 20 | How reliably it completes multi-step work — edits, runs, fixes — without hand-holding. |
| Context handling | 15 | How well it understands a large existing codebase, not just a single file. |
| Speed & flow | 15 | Latency and how little it interrupts your flow during real work. |
| Ease of use & setup | 15 | How quickly you get value without fighting configuration. |
| Price & value | 10 | What you actually get for the money, including free tiers and real limits. |
02 / Process
The testing process
- 01Use it for real
We set up each tool with a genuine account — paid where needed — and use it like a working developer, not a demo.
- 02Run the same tasks
Every tool runs the identical set of tasks on the same repositories, so the results are comparable side by side.
- 03Record the commands and the version
We keep the terminal output, and publish the exact commands plus the version string the tool printed, stamped with the date we ran them — so you can re-run the same command and compare.
- 04Score and write honestly
We fill the scoring sheet, list specific pros and cons, and say who each tool is not for.
- 05Publish only when earned
If testing is not finished, the page shows the facts and marks the score “in progress” — never a placeholder number.
03 / Independence
How we stay independent
Right now no third party pays us anything: there are no affiliate links, no sponsorships and
no paid placements. The only links here that are ours rather than a vendor's are the ones to
our own sibling sites, tagged so we can see which page sent the visit; they never appear in a
comparison or a ranking. Every vendor link is the vendor's own homepage, marked
rel="nofollow noopener". If we ever add an affiliate link it will be marked
rel="sponsored" and disclosed — and a tool with no affiliate program will still be
scored and ranked exactly like one that pays well. Full details are in our
affiliate disclosure.
Who writes these reviews
AI Coding Hub is a one-person site. Every review and guide is written and edited by DK, Editor — there is no wider editorial team, and no outsourced or ghost-written copy. We use these tools on our own projects every day, which keeps us honest about what actually helps versus what only demos well.
04 / FAQ
Questions
- Do affiliate links affect your scores?
- They cannot, because there are none. As of September 4, 2026 this site carries zero affiliate links: every outbound vendor link goes to the vendor’s own homepage with rel="nofollow noopener". If that ever changes, affiliate links will be marked rel="sponsored" and the policy in our affiliate disclosure applies — scores are filled in from the scoring sheet below during hands-on testing and a commercial relationship never moves one.
- What does “tested on real projects” mean?
- We install the tool, run it on our own repositories, and keep the terminal output. What ends up on the page is the text we can prove: the exact install and run commands, the version string the CLI printed, and the date we ran it. Every guide carries that version and date in its byline, so each claim is tied to a specific build at a specific time.
- Why are some scores still marked “in progress”?
- AI coding tools change fast, and a fair score takes a real test. Until that hands-on testing is finished for a given tool, we show the factual comparison only and mark the score “in progress” — we never publish a number we have not earned.
- How often do you re-test?
- We reinstall the CLIs, re-run their help and version commands, and check every command we have published against the build that is current that day. The most recent pass was September 4, 2026 (Claude Code v2.1.260, OpenCode v1.18.27, Codex CLI v0.153.2). Pages we did not re-check in that pass keep their older “last verified” date rather than borrowing a fresh one.