RFC: Cord
Description
Status: Draft Author: June Kim Created: 2026-02-18 Version: 0.4.0
Installation
This entry records only its repository, not the path inside it, so there is no
exact command to give. Open the source below and copy the folder into
~/.claude/skills/, or the file into ~/.claude/agents/.
Repository README
This is the README for kimjune01/cord, shared by 2 entries
in this directory. It describes the repository, not this entry specifically.
RFC: Cord
Status: Draft
Author: June Kim
Created: 2026-02-18
Version: 0.4.0
Abstract
This document specifies Cord, a coordination protocol for agent trees. Agents — implemented as Claude Code CLI processes — execute tasks, create subtasks, and return typed results. A SQLite database is the shared coordination state. Agents interact with it through MCP tools.
The protocol introduces four primitives (`goal`, `task`, `serial`, `ask`) and a single context-flow mechanism: `needs`. When creating a child task, the parent lists which completed nodes' results should be injected into the child's prompt. This replaces the earlier `spawn`/`fork` distinction with explicit context selection.
1. Motivation
Existing agent orchestration systems fall into two categories:
**Imperative frameworks** (LangGraph, AutoGen, OpenAI Swarm) — coordination logic is embedded in application code. No single artifact describes the work. Hard to inspect, modify, or hand off mid-execution.
**Declarative workflow DSLs** (Serverless Workflow, Argo, Airflow) — designed for deterministic microservice orchestration, not for intelligent agents that may need to restructure their own work, ask clarifying questions, or operate with varying degrees of shared context.
No existing system provides:
- Explicit context selection — the parent decides which
results flow to each child via a single
needsparameter. - Native elicitation — agents asking humans, parents, or children as part of the coordination structure.
- Runtime self-modification — agents adding subtasks and restructuring their own subtree during execution.
- Human-in-the-loop through the same interface as agents.
Cord addresses these gaps with a minimal set of primitives designed for coordinating LLM-based agents on a single machine.
2. Overview
2.1 System Architecture
┌─────────────────────────────────────────────┐
│ SQLite Database │
│ (coordination state, WAL mode) │
└──────────────────┬──────────────────────────┘
│ read/write
▼
┌─────────────────────────────────────────────┐
│ Engine │
│ (scheduler, process manager, TUI) │
└──────────────────┬──────────────────────────┘
│ launches
┌────────┴────────┐
│ Claude Code │
│ CLI processes │
│ (one per node) │
└────────┬────────┘
│ MCP (stdio)
┌────────┴────────┐
│ MCP Servers │
│ (one per agent,│
│ stdio transport│
│ → SQLite) │
└─────────────────┘
2.2 Design Constraints
- Single machine. The SQLite database and agent processes are local. Distributed execution is out of scope.
- **Ag
Related Agents
Openai Codex CLI
(55.8k ⭐) - Lightweight coding agent that runs in your terminal.
Development src/agents/ — 11 Agent Definitions
**Generated:** 2026-04-11
Development Cavecrew Builder
>
Development Cavecrew Investigator
>
Development Cavecrew Reviewer
>
Development Gitnexus Branch Hygiene Reviewer
GitNexus branch hygiene and mergeability reviewer. Use to classify merge state, conflicts, stale branches, mer
Development