agent-onboarding

Turn any repo into a self-improving agent environment. Guided 7-phase onboarding produces CLAUDE.md, INTENT.md, PROJECT_BRIEF.md, SPEC_INVENTORY.md, and a full SPECS/ directory. Installs a self-improving runtime that tracks friction, logs errors, and generates structured improvement proposals.

agent-onboarding

Build a harness for any Claude agent — single expert or multi-expert orchestrated system.

Quick Install

claude plugin marketplace add namelesstherebel/agent-onboarding && claude plugin install agent-onboarding

Then open Claude Code in your project and type *onboard.

What It Does

The plugin classifies your agent and builds a harness — the infrastructure that gives it durable memory, grounded retrieval, and a built-in improvement loop.

Two architectures:

Simple — one Claude instance with a LightRAG corpus (what it knows cold) and a MemPalace memory store (what it learns). For chatbots, assistants, and single-domain agents.
Complex — an orchestrator Claude plus N expert Claude instances, each with their own isolated corpus and memory. For agents that require genuinely different domain expertise working together.

Self-improvement is the gate pattern. Every turn: pre-gate retrieves relevant knowledge and memory → Claude responds → post-gate persists new facts and insights. The agent improves by doing its job.

How It Works

*onboard — Classify and build. Five questions determine the architecture. Then targeted questions fill in the domain knowledge (LightRAG corpus), memory categories (MemPalace write taxonomy), and retrieval rules. Output is a directory of harness files the agent loads at runtime.

*reflect — Review MemPalace health: surface provisional insights, stale derivations, corpus gaps.

*review — Process provisional MemPalace items: promote confirmed insights to active, deprecate stale ones.

*status — Harness health report: memory stats, LightRAG status, gate health.

What Gets Generated

[agent-name]/
├── CLAUDE.md                     ← harness entry point: persona + gate protocol + memory pointers
├── core/
│   ├── GATES.md                  ← pre/post gate rules
│   └── INVARIANTS.md             ← the five invariants
├── memory/
│   ├── LIGHTRAG.md               ← corpus sources + ingestion checklist
│   ├── MEMPALACE.md              ← what gets saved, when, confidence rules
│   └── WRITE-TAXONOMY.md         ← 5–8 MemPalace categories
├── retrieval/
│   ├── QUERY-FORMATION.md        ← how prompts become retrieval queries
│   └── GROUNDING.md              ← groundedness check rules
└── observability/
    └── HEALTH.md                 ← health metrics and alert thresholds

Complex agents also get:

└── subagents/
    ├── ORCHESTRATOR.md           ← routing rules, authority boundaries
    ├── TOPOLOGY.md               ← expert list + connection pattern
    └── experts/
        └── [expert-name]/        ← same structure as above, per expert

Install & Update

Install:

claude plugin marketplace add namelesstherebel/agent-onboarding && claude plugin install agent-onboarding

Update:

claude plugin marketplace remove agent-onboarding
claude plugin marketplace add namelesstherebel/agent-onboarding
claude plugin install agent-onboarding

Verify:

claude plugin list

agent-onboarding should appear with status installed.

Memory Architecture

The harness uses two stores:

Store	Purpose	Characteristics
LightRAG	Domain knowledge the agent knows cold	Static, ingested once, queried via direct chunk retrieval
MemPalace	What the agent learns about users and context	Dynamic, grows every turn, scoped per agent

Every MemPalace item carries metadata: timestamp, source, confidence, status, and provenance for derived insights. Superseded items stay in the store for audit — they're just filtered from default retrieval.

agent-onboarding

Popularity

What's Inside

README