deleg8
Enables Claude Code to delegate work to persistent oh-my-pi (omp) subagents via an MCP server with 7 tools (spawn, send, output, status, list, stop, prune), wrapping omp RPC mode with enforcement hooks for descriptive agent naming, model configuration, denylists, write-scope ownership, and cooperative locking.
README
deleg8 — Persistent omp Subagents Plugin
Makes oh-my-pi persistent subagents the default delegation mechanism in Claude Code.
Bundles the deleg8 MCP server, an enforcement hook, and a skill that teaches the
orchestrator when and how to use persistent subagents instead of bare task() calls.
Components
| Component | Purpose |
|---|---|
| MCP server | deleg8 — wraps omp --mode rpc as named, persistent subagents (7 tools: spawn, send, output, status, list, stop, prune) |
| Hook | PreToolUse on task — blocks bare/unnamed task() spawns; forces descriptive agent_id via deleg8 |
| Skill | deleg8 — teaches the orchestrator when deleg8 beats native subagents, with workflow patterns and jq output discipline |
Install
cc --plugin-dir ~/projects/deleg8
Or install from the project:
cd ~/projects/deleg8
cc --plugin-dir .
Prerequisites
- Bun ≥ 1.1.0 (
brew install bunorcurl -fsSL https://bun.sh/install | bash) - omp on PATH (or set
OMP_BINin settings) - Dependencies installed:
bun install
Configuration
Create .claude/deleg8.local.md in your project (or ~/.claude/ for global):
---
omp_bin: /opt/homebrew/bin/omp
default_model:
provider: deepseek
modelId: deepseek-v4-pro
---
| Field | Default | Description |
|---|---|---|
omp_bin |
omp (PATH lookup) |
Path to omp binary |
default_model.provider |
deepseek |
Model provider for deleg8 agents |
default_model.modelId |
deepseek-v4-pro |
Model ID for deleg8 agents |
Usage
Once installed, the skill loads on phrases like "fan out", "parallelize", "waves",
"delegate", "subagents", or when the orchestrator considers spawning a bare task().
Basic fan-out:
spawn(agent_id="compiler", initial_prompt="Build the auth module")
spawn(agent_id="tester", initial_prompt="Write tests for auth", background=true)
Multi-turn refinement:
spawn(agent_id="impl", initial_prompt="Implement payment processor")
→ review output
send(agent_id="impl", message="Add retry logic for timeout failures")
Full workflow patterns: see the deleg8 skill output or references/when-to-use.md
in the skill directory.
MCP Tools
| Tool | Purpose |
|---|---|
spawn |
Launch a new omp agent with optional initial prompt |
send |
Send follow-up to existing agent (auto-resumes idle agents) |
output |
Read buffered frames — digest/summary/raw with jq projection |
status |
Check one agent's state (running/idle/dead) |
list |
Snapshot of all registered agents |
stop |
Abort, terminate, remove from registry |
prune |
Drop dead/idle agents from registry |
Spawn Parameters (New in v0.2)
The spawn tool now supports these enforcement parameters:
| Parameter | Type | Description |
|---|---|---|
denylist |
string[] |
Regex patterns over tool commands. Prompt-level constraints + real-time violation monitoring. Note: omp runs Bash/Read/Edit internally — deleg8 cannot intercept before execution. |
own |
string[] |
Glob patterns limiting write scope. Same injection + monitoring pattern. Same architectural constraint as denylist. |
preamble |
string |
Shared context block every spawned agent receives before its prompt. |
fallback_model |
{provider, modelId} |
If the primary model fails to apply (e.g. budget exceeded), try this one. |
idle_ttl |
number (ms) |
Auto-reap idle agents after N ms of inactivity (registry-level, 0 = off). |
dead_ttl |
number (ms) |
Auto-reap dead agents after N ms (registry-level, 0 = off). |
exclusive |
{pattern, wait}[] |
Cooperative agent protocol — agents acquire/release locks via host tools. Voluntary: agents must call exclusive_acquire/exclusive_release. |
Usage example:
spawn(
agent_id="builder",
initial_prompt="Compile the .csproj",
denylist=["git (stash|checkout)", "dotnet test"],
own=["src/**", "*.csproj"],
preamble="Friflo: Tags getter returns a struct copy — use AddTag/RemoveTag",
idle_ttl=300_000,
exclusive=[{pattern: "dotnet build", wait: true}]
)
Host tools available to agents:
exclusive_acquire(pattern)— acquire an exclusive command lockexclusive_release(pattern)— release an exclusive command lockmsg(text)— send IRC message back to orchestratortask_create(label, note?)— create sub-task visible to orchestratortask_update(task_id, status?, note?)— update sub-task status
Enforcement Hook
The PreToolUse hook on task evaluates every subagent spawn:
- Allows:
exploresubagents (read-only scouting), tasks with descriptiverolefields - Blocks: Bare
task("do X")calls, unnamed/generic subagents - Blocked message: "Bare task() is banned. Use deleg8: spawn(agent_id='descriptive-name', initial_prompt='...')"
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。