Skip to main content
Kodelyth ECC
NEW · v2.31.10Super Agent orchestrator · Memory Core · Kiro now supported · MCP auto-wires 6 IDEsRead the release notes →
Kodelyth ECC — Elite Code Crew
v2.31.10 · Elite Code Crew

AI coding toolkit that already knows your intent.

71 specialist AI agents · 196 skills · 104 slash commands · MCP server · RTK input compression · Terse output compression · codebase graph across 158 languages. All local. Zero telemetry.

npm i -g kodelyth-ecc
GitHub starsnpm weekly downloadsMIT License
The ECC Stack

Every layer local. Every layer real.

Most "AI agent kits" are folders of markdown files you have to remember the names of. ECC is infrastructure — a layered system where intent routing, compound memory, parallel orchestration, and quality hooks all reinforce each other.

71 specialists
Intent routing v2
Plain language → the right specialist agent automatically. Read the intent behind the words, not just keywords.
Multi-agent
Parallel Commands
Fire up to 8 agents simultaneously. E.g., /team-review triggers code-reviewer, security-reviewer, perf-optimizer, and api-guardian at once.
Adversarial
Devil Mode Crew
8 dedicated red-team agents that hunt for prompt injections, supply-chain vulns, hardcoded secrets, and backdoors.
Autonomous
Super Agent
One big objective → done. Decomposes into a dependency DAG, dispatches independent steps in parallel on their own branches, conflict-checks and integrates, then verifies. Resumable and budget-capped.
GOD vs EVIL
The Arena
Two crews fight over your code until the attacker gives up. GOD builds and hardens, EVIL attacks and proves, and every finding is verified before it counts.
-90% input
RTK compression
Auto-installed PreToolUse hook. Filters shell output like git status before it reaches the LLM.
-70% output
Terse mode
4-level output compression dial. Byte-preserves code and commands. Saves output tokens.
-99% structural
Codebase graph
AST intelligence across 158 languages via tree-sitter. Structural queries in <1ms without LLM cost.
6-IDE auto-wire
MCP Server
Auto-wires all 71 agents, 196 skills, and local memory into Claude Code, Cursor, Windsurf, Gemini, Codex, and Kiro — and speaks to Claude Desktop, LangGraph, AutoGen, and CrewAI.
Memory Core
Compound Memory
Local BM25 recall that survives a repo move (keyed by git remote), redacts secrets on capture, and fuses relevance with recency and outcome. Opt-in on-device semantics.
Zero-dep
Interactive CLI
Type `kodelythecc` for an arrow-key menu with update checks, background daemon, and uninstall tools.
How it works

Every layer, illustrated

One diagram per capability — with real numbers from the measured ledger, not marketing estimates.

It picks the specialist for you
Intent routing v2

It picks the specialist for you

Describe your problem in plain words. ECC reads the intent behind the message — the error, the frustration, even a pasted stack trace — and routes to the right one of 71 specialist agents, announcing which took over.

  • No agent names to memorize
  • 10 priority tiers, 8 confidence dimensions
  • Transparent — never routes silently
Read the docs
One objective, driven to done
Super Agent · autonomous

One objective, driven to done

Hand it a whole objective — too big for one pass — and it runs through a tokenless state machine: decompose into a dependency DAG, dispatch independent steps in parallel on their own branches, conflict-check and integrate, then verify. Deterministic, resumable, budget-governed.

  • Parallel dispatch, conflict-safe integration
  • Resumes after a crash — persists every step
  • Budget cap is a hard stop · nothing faked
Read the docs
Shell output shrunk before the model reads it
RTK · input savings

Shell output shrunk before the model reads it

RTK is an auto-installed PreToolUse hook. When your AI runs git status, cargo test, or docker ps, RTK filters the output first — stripping noise, grouping, deduplicating — so only signal reaches the LLM.

  • 60–90% fewer input tokens
  • 100+ commands supported
  • 5.1M tokens saved on the sample ledger (64%)
Read the docs
The AI says the same thing in fewer words
Terse · output savings

The AI says the same thing in fewer words

A 4-level dial that compresses what the model writes back — dropping filler while byte-preserving every code block, command, URL, and path. Turn it up on familiar work, off when the prose is the deliverable.

  • 40–70% fewer output tokens
  • lite / full / ultra / off
  • Code stays byte-exact
Read the docs
Ask structural questions, skip the grep
Codebase graph

Ask structural questions, skip the grep

An AST knowledge graph across 158 languages via tree-sitter. “Who calls this?” and “what breaks if I change it?” resolve from the graph in one query instead of dozens of file reads.

  • ~99% fewer tokens than file-by-file grep
  • 158 languages, <1ms queries
  • Runs locally, no LLM cost
Read the docs
Every agent, anywhere
MCP server

Every agent, anywhere

One universal adapter exposes all 71 agents, 196 skills, and your local memory over the Model Context Protocol. It auto-wires into Claude Code, Cursor, Windsurf, Gemini, Codex, and Kiro, and speaks to Claude Desktop, LangGraph, AutoGen, CrewAI, and the OpenAI Agents SDK natively.

  • 16 tools · 6 prompts · 389 resources
  • Auto-wires 6 IDEs; merge never clobbers
  • stdio JSON-RPC · zero telemetry
Read the docs
One command, a whole control panel
Interactive CLI

One command, a whole control panel

Type kodelythecc alone for an arrow-key menu: open the dashboard, install for another IDE, check RTK / Terse / Codebase status, run a background daemon, or uninstall — with a live update check baked in.

  • Zero dependencies, raw-mode TUI
  • 24h-cached npm update check
  • Health check + full uninstall built in
Read the docs
8 agents that attack your code first
Devil mode

8 agents that attack your code first

A dedicated red-team crew that thinks like an attacker. One command fires all eight in parallel to hunt prompt injections, supply-chain vulns, leaked secrets, and backdoors before you ship — not after.

  • Prompt injection · jailbreaks · RAG security
  • Typosquats · malicious postinstall · lockfile drift
  • Leaked keys · backdoors · obfuscated payloads
Read the docs
It remembers what worked last time
Memory Core

It remembers what worked last time

A local BM25 memory that captures solutions at session end and auto-recalls the relevant one on your next prompt. Memory Core keys it to your git remote so it survives a repo move, redacts secrets before they're ever written, and fuses relevance with recency and outcome when it ranks.

  • Survives a repo move — keyed by git remote
  • Secret redaction on capture · fused recall
  • 100% local · opt-in on-device semantics
Read the docs
Compare

Figuring out if ECC fits?

Honest side-by-sides against the tools people ask about most. No strawmen, no "ECC wins everything" — real tradeoffs.

Compound token savings

55-65% total token reduction on real coding sessions.

RTK compresses shell command output before it reaches the LLM. Terse mode compresses what the LLM writes back. Two axes, they stack.

How it works
RTK input savings64.1%
Terse output savings40-70%
Codebase structural queries99%
Live ledger from a real user Mac (1,285 commands sampled).
Estimate your savings

What would ECC save you?

$200
Est. monthly savings
$110–$140
Est. yearly savings
$1,320–$1,680

Blended 55–70% token reduction: RTK cuts input ~64% (measured), Terse cuts output ~50%, the codebase graph removes grep/read passes. Actual savings depend on your workload — explain-heavy sessions save more, terse debugging saves less. Numbers are estimates, not a guarantee.

12 AI IDEs, 14 install targets

One install command, every IDE wired up.

FAQ

Common questions

Is Kodelyth ECC free?

Yes. Kodelyth ECC is MIT-licensed and free forever. Install it with npm i -g kodelyth-ecc. No account, no subscription, no telemetry.

Does it work with Cursor, Windsurf, and other IDEs?

Yes. ECC installs into 12 AI coding platforms including Claude Code, Cursor, Windsurf, Codex CLI, Google Antigravity, OpenCode, Cline, Roo Code, Aider, Kimi, Gemini CLI, and Kiro — 14 install targets total.

Is my code sent anywhere?

No. Everything runs locally on your machine — agents, skills, memory store, RTK compression, and the codebase graph. Zero telemetry, no cloud calls.

How much does it reduce token usage?

RTK compresses shell-command input by 60-90%, Terse mode compresses AI output by 40-70%, and the codebase graph answers structural queries with ~99% fewer tokens than file-by-file grep. Combined, a typical coding session drops 55-65% in total tokens.

How do I install it?

Run npm i -g kodelyth-ecc then kodelythecc --target claude-code --codebase-graph, and restart your AI tool. One command wires up all 71 agents, the MCP server, RTK, Terse mode, and the codebase graph.

Ready in one command.

MIT · Free forever · No account required · Runs locally on your machine.