DEV Community

Cover image for Why I Built an AI Coding Platform with BYOK (Bring Your Own Key) that uses Claude skills directly
Vandana
Vandana

Posted on

Why I Built an AI Coding Platform with BYOK (Bring Your Own Key) that uses Claude skills directly

#ai #productivity #showdev #devtools

The Problem That Wouldn't Go Away

I had a setup I loved: Claude API with custom skills I'd spent weeks building. Database query tools, file system automation, deployment scripts - everything tuned to my workflow.

Then I tried Cursor.

Here's what frustrated me:

  1. Paying twice: I was already paying $20/month for Claude API. Cursor wanted another $20/month to use my own API credits. I was literally paying twice to access the same AI.

  2. Lost my skills: Remember those Claude skills I built in the Developer Console? Completely unusable in Cursor. All that work, just sitting there.

  3. Stuck in real-time: Cursor is great for pair programming, but I couldn't kick off a refactor before bed and wake up to reviewed code. Everything required me at the keyboard.

  4. Siloed: My colleagues couldn't leverage my custom skills. No way to share the automation I'd built.

The final straw? I calculated I was spending $480/year on tools that used an API I was already paying for, while my custom skills collected dust.

The "Aha" Moment

I was working late, wanting to refactor a module before morning. I thought: "Why can't I just start this task, let AI agents handle it overnight, and review the results tomorrow?"

That's when it clicked:

What if you could:

  • Use your own Claude API key (no markup)
  • Bring your Claude Developer Console skills
  • Run workflows asynchronously (not just real-time assistance)
  • Share automation with your team
  • Trigger tasks via API without touching an IDE

The technical challenge was interesting: orchestrating multiple specialized agents (planning → coding → review) while maintaining access to custom skills.

But the core insight was simple: async workflows are fundamentally different from real-time assistance, and almost nobody was building for that.

How I Built Threadyx

Multi-Agent Architecture

Instead of one AI trying to do everything, I built specialized agents:

1. Planning Agent

  • Analyzes the task
  • Breaks it into steps
  • Identifies which Claude skills to use
  • Creates implementation plan
  • Provides human-readable summary
  • Recommends approve/reject

2. Coding Agent

  • Executes the plan
  • Generates code changes
  • References your custom skills
  • Works in cloned environment

3. Review Agent

  • Validates changes
  • Checks for issues

Each agent has a specific job. This reduces errors and improves consistency vs. a single agent context-switching between planning, coding, and reviewing.

Claude Skills Integration (The Hard Part)

Claude's API supports up to 5 skills per request. But I wanted users to reference unlimited skills.

The solution:

  • Users provide all their Claude skill IDs (no limit)
  • For each task, Threadyx analyzes which skills are relevant
  • Intelligently selects the top 5 most applicable skills
  • Passes those to the API

Example:

You have 15 Claude skills:

  • Salesforce Flows (3 skills)
  • Apex (4 skills)
  • API integrations (5 skills)
  • Development Expert (1 skill)

You ask: "Create a salesforce screen flow to update contact's address related to an account"

Threadyx automatically selects your 3 + 1 skills (to create flow + development), ignoring the irrelevant ones.

Here's what it looks like in practice:

Threadyx automatically selects the most relevant skills from your collection

API-First Design

Unlike IDE-based tools, everything in Threadyx is API-driven:

{
  "repository": {
    "id": "wds",
    "url": "https://github.com/GCS/wds.git",
    "baseBranch": "main"
  },
  "task": "Create a salesforce screen flow to update contact's address related to an account",
  "skills": "skill_011SyG9cLJYM6x1jgXkS6ryK,skill_01UoXqka4XKEBAgUzHZGmVwh,skill_013AErsWQ4DNzuTwjwBBA9Xa,skill_015PBk47MiyctTWykMob3eP1, skill_01M3yhcUNHpW676RphF3wyYr,skill_011NPA8LWNtviQsrVfZR8Pwj,skill_01GdgsfNcVhdNT4z2U8Y5U7k,skill_01MwFDgchx2nC96uMMCErfBp,skill_013UAXQAyMXwXEF6cSt6mgC4,skill_01C5a2Voh9YmVUGJ5BvNT9hF",
  "webhookUrl": "https://webhook.site/8f80a24b-e44a-4382-aca8-f963c7c86220",
  "apiKey": "your-anthropic-key",
  "userId": "your-user-id"
Enter fullscreen mode Exit fullscreen mode

This means you can:

  • Integrate with CI/CD pipelines
  • Schedule recurring tasks
  • Build custom automations
  • Trigger workflows from anywhere (not just your IDE)

Safety First: Cloned Environments

Every workflow runs in a cloned environment. Your main codebase is never touched until you explicitly approve and merge changes.

Workflow:

  1. Clone your repo/environment
  2. AI agents work in the clone
  3. You review the changes
  4. You decide: merge or discard

No surprises. No accidental commits. Full control.

What I Learned Building This

1. BYOK Resonates More Than I Expected

When I posted about Threadyx, the #1 response was: "Finally! I'm already paying for Claude API, why would I pay again?"

Cost-conscious developers (indie hackers, small teams) feel the pain of subscription fatigue. They want control over their AI costs.

2. Async Workflows Are Severely Underserved

Everyone's building real-time coding assistants. But there's a huge gap for:

  • Overnight refactors
  • Scheduled maintenance tasks
  • Large codebase migrations
  • Tasks that don't need human babysitting

The developer doesn't need to be at the keyboard 24/7. Let agents work while you sleep.

3. Claude Skills Are Powerful But Underutilized

Most developers don't even know Claude Developer Console skills exist. Those who do build them find they're trapped in the console.

Making skills portable and reusable across tools is a game-changer.

4. Direct Outreach > Big Launches

I tried launching on Hacker News. It went nowhere (2 upvotes, buried in 34 minutes).

What actually worked? Sending 20 personal messages to developers who tweet about Claude API. 10% response rate. Real conversations. Better feedback than any launch.

Lesson: Find 10 people who love your product > getting 1000 people to glance at it.

Try Threadyx Yourself

I'm looking for early users who:

  • Already use Claude API
  • Have built custom Claude skills
  • Want to automate development workflows
  • Need async task execution

Get started:

What you'll need:

  • Claude API key (get one at console.anthropic.com)
  • GITHUB PAT key
  • Your Claude Developer Console skill IDs (optional but recommended)
  • 5 minutes for setup

Early user benefits:

  • Extended trial period
  • Direct access to me (the founder) for support
  • Input on what features to build next
  • Priority bug fixes

Questions for You

I'm genuinely curious:

  1. Would you use a BYOK model, or do you prefer all-in-one tools even if they cost more?

  2. What would you automate first if you had async AI agents?

    • Nightly refactors?
    • Test generation?
    • Documentation updates?
    • Code reviews?
    • Something else?

  3. Do you have Claude skills you wish you could use elsewhere?

  4. What's stopping you from trying it? (Be honest - this helps me improve)

Drop your thoughts in the comments. I read and respond to everything.

About Threadyx:
Threadyx is a multi-agent AI coding platform where you bring your own Claude API key and skills. Built for developers who want control, automation, and async workflows without subscription lock-in.

About me:
I'm an indie developer frustrated by AI tool pricing. You can reach me directly in the comments or at vandana.threadyx@gmail.com.

Thanks for reading! 🚀
Enter fullscreen mode Exit fullscreen mode

Top comments (1)

Collapse
 
vandana_5cca941ab74e21680 profile image
Vandana •

Hey everyone!

Author here. Happy to answer any questions about:

  • The multi-agent architecture
  • How Claude skills selection works
  • BYOK security model
  • Technical implementation details
  • Why I chose Railway for deployment

Also offering to jump on quick calls with anyone interested in trying it out. Just want to understand your workflow and help you get value from Threadyx.

What would you automate first?