Skip to main content
Kodelyth ECC
About

Who built this, and why

A free, MIT-licensed AI coding toolkit built so you keep the leverage without renting the agent.

What Kodelyth ECC is

Kodelyth ECC (Elite Code Crew) is a production-grade AI coding toolkit for Claude Code, Cursor, Windsurf, Codex, Antigravity, Kiro, and 6 more AI IDEs. 71 specialist agents selectable by plain-language intent routing; 196 skills; 104 slash commands; a universal MCP server; local Memory Core recall; two independent token-compression layers (RTK on input, Terse on output); a codebase AST graph across 158 languages; and a Super Agent orchestrator for objectives too big for one pass. Current release: v2.31.3.

Everything runs locally. Zero telemetry. MIT licensed. Every agent, skill, command, and rule ships as editable markdown — read it, change it, delete the ones you don't use.

Who built it

Kodelyth ECC is built and maintained by Shofiqul Islam (@sifxprime) under the Kodelyth umbrella. The toolkit is a single-maintainer project with contributions welcome — no VC, no SaaS, no account, no lock-in.

Why it exists

Most “AI agent kits” are folders of markdown you have to remember the names of. The names don't compose; the folders don't communicate; the whole system fragments the moment you leave one IDE. ECC is the opposite direction: a layered system where intent routing, memory recall, MCP bridging, and token compression reinforce each other, and the same catalog works in every IDE that speaks files or MCP.

The design bet is that you should keep the leverage. Your agents, your memory, your routing, your token savings — on your machine, in your git repo, under an MIT license. No one else owns the thing that makes your AI coding sessions better.

Design principles

  • Zero telemetry, zero network. Not a marketing claim — an architectural one. The only external traffic ECC introduces is the MCP server auto-registration step when you install, and that is a one-time local file write.
  • Zero runtime dependencies in the core. Pure-JS install. Three optional peer deps (MCP SDK, sql.js, on-device embeddings) gate opt-in features and degrade cleanly when absent.
  • Markdown first. Agents, skills, commands, rules are markdown. Read them, diff them, copy them, edit them, delete them. No binary formats, no SaaS console, no “please wait for a dashboard.”
  • Honest numbers. The RTK ledger is measured, not estimated — and ECC labels the headline figure (64.1%) as an example, not a promise, because savings depend on which commands you run. The Terse runtime ledger says “estimated” on every row because it is.
  • Nothing faked in autonomous runs. The Super Agent records failed steps as failed; a failed dependency stalls the run instead of letting the machine build on a broken foundation.
  • Composable with your tools, not around them. Cursor Rules,CLAUDE.md, your hooks, your MCP servers — untouched. ECC adds alongside; it never overwrites.

What's in the toolkit

  • Intent routing — 71 specialists selectable by plain language
  • Super Agent — autonomous orchestrator, resumable, budget-governed
  • MCP server — bridges the catalog to any MCP client (auto-wires 6 IDEs)
  • Memory Core — local BM25 recall, git-remote-identified, redacts secrets
  • RTK — shell-output input compression (60-90%)
  • Terse mode — AI-reply output compression (40-70%)
  • Codebase graph — AST intelligence across 158 languages
  • Dashboard — localhost observability, real data only

How to contribute

PRs welcome at sifxprime/kodelyth-ecc. A good first PR is a new agent, skill, or slash command — see CONTRIBUTING.md.

Found a bug, want a feature, built something cool on top? Open an issue. If this is useful, consider sponsoring on GitHub Sponsors. Free and MIT-licensed is the plan; sponsorship keeps it that way.

Using ECC in production? Share your story

A future version of this page will link real case studies — short, specific accounts of teams or individuals using ECC in actual projects. We refuse to make them up. If you ship with ECC and would be willing to be quoted (named or anonymous), open a thread at GitHub Discussions → Show & Tell. A paragraph is plenty: what you build with, which agents or commands you use most, what you'd change. Honest stories — including the parts where ECC isn't the answer — are the most useful.

Prefer to email privately? Open a case-study issue and mention you want to keep it off the main page — we'll follow up by email.

Stats (44 hooks, 15 rules, 71 agents, 196 skills)

Read the changelog for the full feature history, the docs for every subsystem, the blog for engineering postmortems and incident stories, or the install page to get started in two minutes.