I Used AgentKit for 30 Days: Real Results After Paying $99 (2026 Case Study)

⚠️ Which AgentKit? (this is NOT OpenAI AgentKit)

Before we get into the case study, I have to be clear so you are not reading the wrong article. There are at least two things called "AgentKit":

• The AgentKit in this article = the kit for Claude Code at agentkit.best, used through the ak CLI. It is the successor to the old ClaudeKit (ck becomes ak), bundling 108+ skills and 45 agents to turn Claude Code / Codex / Copilot into an "AI dev and marketing team".
• OpenAI AgentKit = OpenAI's agent-building toolkit (Agent Builder, ChatKit), launched Oct 6, 2025 with a wave of enterprise case studies like Ramp and Canva (OpenAI source). Completely different from what I am reviewing here.

There is also agentkit.today (n8n workflows) and dev.to/agentkit - same name, different product. If you came looking for an OpenAI AgentKit case study, this article is not for you. But if you are on the fence about spending $99 on a kit for Claude Code, keep reading.

Context: me and my project before installing AgentKit

I am a solo dev in Vietnam doing full-stack work on a few small client projects plus one personal product. My comfortable stack: TypeScript/Next.js on the front, Node and a bit of Python on the back, Postgres, and odds-and-ends deploys to Vercel/VPS. Before this month I had used plain Claude Code (no kit) for about half a year - mostly opening a terminal, typing ad-hoc prompts, pasting in errors, asking for fixes, then stitching it together myself. I had a bare-bones CLAUDE.md, no subagents, no skills, no fixed process.

That way of working got the job done, but I kept hitting three recurring pain points: (1) every new task meant re-explaining the project context from scratch; (2) repetitive work like writing tests, reviewing diffs, and scaffolding CRUD still got hand-prompted every single time; (3) output quality swung wildly depending on whether I happened to prompt well that day.

Day-0 baseline (numbers I logged myself so I could compare later):

• Average time to finish one medium task (one CRUD feature + tests): 15 minutes
• Bugs that slipped past my own testing and surfaced at review/runtime: 0
• Process: ad-hoc, no fixed brainstorm/plan step.

If you are not sure how plain Claude Code differs once you add a kit, I have a separate comparison in AgentKit vs plain Claude Code.

How I measured (so you do not have to take my word for it)

A case study that does not explain its method is just an ad. Here is how I recorded numbers over the 30 days:

• Time per task: manual stopwatch from start to merge, jotted into a simple sheet. I only counted "medium" tasks so I was comparing like with like.
• Features shipped and bugs: counted from actual commits/PRs in the repo, where a bug is something I had to come back and fix after thinking I was done.
• Cost: the kit price ($99) plus the Claude Code I still pay for (Pro at $20/month). I do not use the token-metered API, so the token side is only a qualitative estimate.

A limitation I have to state plainly: this is n=1, self-reported. I have no control group, and the mere fact that I knew I was measuring may have made me more focused. Read the numbers below as one person's trend, not research.

Week 1 - Installing ak and getting oriented (setup under 5 minutes)

Install went exactly as the site claims: under 5 minutes. The ak CLI is a native binary for macOS/Linux/Windows, with no Node or Bun required. The installer auto-detects your OS/arch, downloads the stable build, verifies the SHA-256, and drops it into ~/.local/bin; you authenticate with an email or an API key (details at docs.agentkit.best). Since I was already on the old ClaudeKit, I ran an extra migrate step:

# preview the changes (read-only by default)
ak migrate --from=ck

# apply for real once you are happy
ak migrate --from=ck --dry-run=false --yes

# list the available skills
ak list skills

The truth about week 1: I slowed down. With 108+ skills and 45 agents in front of me, the first feeling was overwhelm, not an instant speed-up. I spent a few days just figuring out what each skill does, and I still reflexively hand-prompted out of old habit more than once. This is the thing no competitor tells you: the first-week learning curve is real. If you are completely new to Claude Code, read through the detailed ak CLI install guide first so you stumble less.

Week 2 - The skills/subagents I used most (and the ones I dropped)

This is the part I find most valuable to share, because marketing loves to boast "108+ skills" when in reality one person only uses a handful. By week 2, once the overwhelm faded, I started seeing a few skills/subagents recur through my working day - while most of the rest I never touched. And that is fine: a kit is like a toolbox, not a checklist you must use in full.

An honest estimate: out of 108+ skills, I regularly use around 70; out of 45 agents, I call maybe 15~20 often. The table below is the ones that created the clearest value for me (your tasks differ, so your list will too):

If you want the full breakdown of what the Engineer Kit includes and whether it is worth it, I have a dedicated piece: what is in the Engineer Kit, and is it worth it, plus an overview of the 108+ skills and an overview of the 45 agents. The point I want you to remember: do not buy for the "108+" number, buy for the 8-12 skills that fit your work.

Week 3 - How the workflow changed (brainstorm to plan to cook to ship)

Week 3 was the biggest shift, not in typing speed but in how I think about a task. I used to prompt straight into "build me X." With the kit, I moved to a four-beat frame - I follow the brainstorm to plan to cook to ship workflow as the backbone:

1. Brainstorm: lock the desired outcome, constraints, and non-goals - before touching any code.
2. Plan: let a subagent break it into checkable steps, and I approve the plan.
3. Cook: execute with the context-appropriate skill (backend, frontend, review, and so on).
4. Ship: review the diff with the code-review skill, run tests, and only then merge.

The real difference: I re-explain context far less each time, and output swings less because it comes from a process rather than prompt luck. The trade-off is a slower, more disciplined rhythm - for tiny tasks the frame is overkill, so I still hand-prompt those for speed.

Week 4 - Steady results and the habits I kept

By week 4, things "settled." I no longer look up what each skill does, and I now know when not to use the kit (a task too small or too unusual for the patterns). The habits I kept:

• Always brainstorm + plan for anything from medium size up; skip it for small tasks.
• Always run the review skill before committing - the one I most regret not doing sooner.
• Keep only the handful of fitting skills in my head, and let the rest sit there.

Week 4 did not feel like "twice as fast" the way ads promise. It felt like less friction and more consistency. That is what keeps me around, not one shocking number.

30-day numbers: before vs after

This is the part you came to read. Reminder: n=1, self-reported, read as a trend.

Qualitative feel (more certain than the numbers): the clearest time drop was on repetitive tasks (CRUD, tests, review), and almost no change on research/hard tasks that need me to think a lot. Put another way, the kit shortens the "muscle" work, not the "brain" work.

One more thing worth noting when you measure yourself: your week-1 numbers will be worse than baseline because you are learning, and only improve from week 2-3 on. If you only measure the first 3 days and conclude "the kit slows me down," you will read it wrong. That is why I averaged across all 30 days and split week 1 out to see the trend correctly. It is also why I suggest giving the kit at least 2 weeks before judging.

What actually worked - and what I expected wrong

No competitor has this section, and it is also the most honest one.

Real wins (pleasant surprises):

• The code-review skill caught bugs I tend to miss - more valuable than I expected, and effectively "free" in effort because it runs automatically in the ship beat.
• Having a framed process made output far more consistent than ad-hoc prompting; it cut way down on the "the AI is dumb today" days.
• Less "re-explaining context" per task - a lot of small compounding micro-friction saved.

Not what I expected:

• I thought I would use most of the 108+ skills. In reality, only a handful. The big number is a selling point, not a usage point.
• Week 1 slowed me down rather than speeding me up right away - anyone expecting "plug in and fly" will be let down.
• The kit does not turn a Claude Code beginner into a pro. It amplifies existing skill; if you are not solid on Claude Code yet, learn that first.
• For very small tasks the workflow frame becomes a burden, and I go back to hand-prompting.

Cost and ROI after 30 days

Prices as of when I wrote this (verified on the site, you should verify yourself since prices can change):

• Engineer Kit - $99: 60+ skills, 30+ workflows, 17 engineer agents. The site does not state a recurring fee for the kit (the $99 price shows with no "renews" wording).
• Marketing Kit - $99; Bundle of both - $149.
• The Desktop App is the yearly one: $19/year early access (first year, renews later at standard price), or App + 2 kits at $99/year for the first 100 people. That is the subscription part, separate from the kit.
• There is a money-back guarantee (the site does not list a specific day count) and lifetime updates for the kit.

How I compute ROI myself (plug in your own numbers):

Kit cost = $99 (one-time)
Hours saved/month = 40
Value of one hour = $50
Value saved = 40 x $50
Month-1 ROI = $2000 - $99

For me the key point is that the kit is a one-time $99 while the time savings recur every month - so if you code regularly, the break-even threshold is fairly low. But I will not hand you a fantasy ROI figure: fill in the real hours you measure. For a deeper analysis, see how much AgentKit costs and whether it is worth buying and my detailed AgentKit ROI calculation.

Want to check the prices and plans yourself? Look at the pricing page directly - kit $99, bundle $149, with a money-back guarantee and lifetime updates: see AgentKit pricing today (20% off via link).

Pros and cons after 30 days

After 30 days, is AgentKit worth buying? (who should / should not)

My verdict by type of person:

• Solo / indie devs already comfortable with Claude Code: Yes. This is the group that benefits most clearly - you have the base skill for the kit to amplify, and a one-time $99 is easy to recoup if you code often.
• Busy devs with lots of repetitive tasks: Yes, especially since the review skill + workflow cut daily friction.
• Complete beginners to Claude Code: Hold off. Learn plain Claude Code first; the kit cannot replace the foundation. See the concrete comparison in AgentKit vs plain Claude Code.
• Small teams: there is potential, but this article is a solo angle - I have not measured at team scale, so I will not conclude for you.

If you want a general product review before deciding, see what AgentKit is - a general review. If you are already sure you are in the "yes" group, you can grab AgentKit for $79.20 (20% off via link) and use the money-back guarantee to try it for yourself over the first few days.

Frequently asked questions (FAQ)

Is 30 days enough to draw a conclusion?

Enough to see a trend and decide whether to buy, but not research. These are n=1, self-reported numbers. Treat it as one person's experience, and re-measure on your own tasks.

How is this AgentKit different from OpenAI AgentKit?

Completely different. This article is about AgentKit for Claude Code (agentkit.best, the ak CLI). OpenAI AgentKit is OpenAI's agent-building suite (Agent Builder/ChatKit), launched Oct 6, 2025 - a different product from a different company.

Does the kit charge a yearly fee?

The kit shows a $99 (Engineer) / $149 (Bundle) price and the site does not state a recurring fee for the kit, along with lifetime updates. The yearly part is the Desktop App ($19/year early access) - that is a separate product, not the kit.

Is there a refund if it does not fit?

The site has a money-back guarantee. It does not spell out a specific day count or conditions, so read the policy carefully before buying.

Do I need Claude Code Pro?

You need Claude Code to use the kit. I pay for the Pro plan at $20/month. There is a limited free tier, but for steady work a paid plan is more comfortable.

Which link should I buy through?

I put the link in the box at the end of the article and in the Cost and ROI section. Buying through it does not change your price. I keep the review balanced on both pros and cons, so I do not lose credibility with you.

Conclusion and buy link

After 30 days, the honest verdict: AgentKit is no magic wand, but it is a reasonable $99 for devs already comfortable with Claude Code - the value comes from a consistent process and the review skill, not from the "108+ skills" number. A slow week 1 is normal; get past it and the friction drops noticeably. If you are in the "should buy" group above and want to verify for yourself, use the money-back guarantee to try it.