By Kodelyth · updated · v2.30.0
Kodelyth ECC vs OpenHands
Short version. OpenHands (formerly OpenDevin) is an open-source autonomous software engineer — agent + sandbox + planner that drives a whole task end-to-end without a human-in-the-loop. Kodelyth ECC is a specialist toolkit that lives inside an interactive AI coding tool (Claude Code, Cursor, Windsurf, etc.). Different layers, different problems. This page is honest about whether they can be combined today.
What OpenHands is
OpenHands is a complete autonomous agent system. It:
- Runs in its own sandbox (Docker by default)
- Plans tool use (bash, browser, file edit)
- Executes actions, observes results, re-plans
- Aims to complete a whole task from a single prompt without further intervention
- Supports multiple model providers
- Has a web UI and a CLI
- MIT-licensed, community-driven
The design bet is autonomy. You hand OpenHands a task, it tries to solve it end-to-end, you look at the result.
What ECC is
ECC is not an autonomous agent. It's a catalog of agents, skills, commands, rules, and infrastructure that lives inside a human-in-the-loop AI coding tool. What it adds:
- 71 specialist agents — debug-detective, code-reviewer, security-reviewer, performance-optimizer, 67 more — routable by plain-language intent
- 196 skills — granular workflows the agent dispatches when a task matches
- 104 slash commands — including 8 parallel-fire multi-agent commands
- Universal MCP server — bridges the catalog to any MCP client
- RTK + Terse — 55-65% combined token reduction
- Memory Core — local, git-remote-identified, redacts secrets
- Super Agent — autonomous orchestrator inside your IDE for multi-step objectives (resumable, budget-governed, tokenless state machine)
The design bet is steer-ability: specialists on demand, honest numbers, nothing hidden, you drive.
How they stack (honestly)
| Layer | OpenHands | ECC |
|---|---|---|
| The agent loop itself (plan → act → observe → repeat) | ✓ | — |
| Own sandbox / own execution environment | ✓ | — |
| Unattended end-to-end task execution | ✓ | Partial (Super Agent, budget-capped) |
| Lives inside an interactive IDE | — | ✓ |
| Specialist agents (71), routable by plain-language intent | — | ✓ |
| Catalog of skills, slash commands, parallel-fire commands | — | ✓ |
| Universal MCP server exposing the catalog | — | ✓ |
| Local memory that recalls past fixes | — | ✓ |
| Shell-output token compression (RTK) | — | ✓ |
| AI-reply token compression (Terse) | — | ✓ |
ECC and OpenHands are adjacent, not competitive. You might run OpenHands to drive a long task autonomously and run ECC in your editor for the per-pull- request specialist review. Or you might prefer one entirely over the other — both are valid.
Can you install ECC into OpenHands today?
Not directly. OpenHands is not currently a target in ECC's installer because OpenHands runs in its own sandbox with its own catalog format. The practical paths today:
- Use ECC's catalog as reference material. Run
kodelythecc catalog --format=markdownto emit the full 71-agent + 196-skill + 104-command catalog as a single Markdown file. Point OpenHands at it as a persistent context document. That gives OpenHands the knowledge of what specialists would recommend for which task, even without routing. - Use ECC's MCP server from OpenHands when MCP client support lands.
The ECC MCP server is spec-compliant and exposes the catalog + local
memory. Any MCP-compatible client — including future OpenHands — can speak
to it. When OpenHands ships MCP client support, the integration becomes
kodelythecc mcprunning as a subprocess. kodelythecc explain <topic>prints the authoritative doc body for a topic. OpenHands (or any agent) can shell out to this at any time to get a verbatim, cite-able reference instead of inventing one.
If OpenHands exposes an MCP endpoint or a plugin API, drop an issue on sifxprime/kodelyth-ecc and we'll ship a first-class target for it.
When to pick which
Pick OpenHands when:
- You want an agent to complete a whole task unattended, in a sandbox, and you will inspect the result afterwards.
- You're building agent workflows, not writing code interactively.
- The task is well-specified enough that one autonomous run has a realistic chance of landing a usable result.
Pick an interactive IDE + ECC when:
- You want to steer every step and have specialists on demand.
- You want token compression layered under whatever you're already doing.
- You want the same catalog available across 12 IDEs.
- You want local Memory Core to carry past-fix knowledge into every session.
Use both when:
- OpenHands drives a long task overnight; your morning review in Claude Code + ECC catches what it missed.
Bottom line
OpenHands and ECC are at different layers, both open source, both MIT. No
direct integration exists today; the MCP server is the bridge when OpenHands
adds MCP client support, and the catalog --format=markdown + explain
subcommands already let OpenHands cite ECC content as reference.
# ECC — the specialist toolkit for your interactive IDE
npm i -g kodelyth-ecc && kodelythecc --target claude-code
# OpenHands — a separate autonomous agent harness
# https://github.com/All-Hands-AI/OpenHands