The business operations harness for AI agents — persistent memory, reusable skills, SaaS protocols, approval gates, and full audit — built on Lark.
The biggest shift in AI agents had nothing to do with making models smarter. It was about making the environment around them smarter.
Lark Harness is that environment — for business operations.
LLM (Claude / GPT / ...)
↓
Lark Harness
memory · skills · protocols · gates · audit
↓
SaaS APIs — freee / Shopify / HubSpot / SmartHR / ...
↓ lark-cli
Lark — Drive / Base / IM / Approval / Wiki
Same model. Completely different reliability.
AI agents fail in business operations for three reasons:
| Problem | Lark Harness answer |
|---|---|
| No memory across sessions | Lark Base as persistent memory |
| No reusable domain knowledge | SKILL.md — 25 skills for Lark + SaaS operations |
| No governance | Gate policy (none / preview / approval) + full audit log |
| Capability | What it does |
|---|---|
larc bootstrap |
Loads agent context (SOUL → USER → MEMORY → HEARTBEAT) from Lark Drive |
larc auth suggest |
Infers minimum required scopes from a natural-language task description |
larc approve gate |
Enforces execution gate policy before any action |
larc ingress enqueue |
Queues a task into Lark Base with source tagging |
larc ingress run-once |
Picks up and executes the next queued task |
larc ingress done / fail |
Records outcome + notifies via Lark IM |
larc agent register |
Registers agents in Lark Base with declared scopes |
larc memory |
Syncs daily memory with Lark Base |
larc kg build / query |
Indexes Lark Wiki node graph; answers concept queries |
| Claude Code skills | 25 Lark-oriented skills in .claude/skills/ |
Lark Harness wraps best-in-class MCP servers, adding HITL gates and Lark routing on top:
| SaaS | MCP | Coverage |
|---|---|---|
| freee | freee-mcp (OSS, Apache-2.0) |
会計 / HR / 請求 / 工数 / 電子契約 |
| Shopify | Shopify AI Toolkit (OSS, MIT) |
注文 / 商品 / 顧客 / 在庫 |
| HubSpot | HubSpot MCP | CRM / マーケ |
| Google Workspace | Google MCP | Drive / Docs / Sheets / Calendar |
| Notion | Notion MCP | ドキュメント管理 |
| SmartHR | API wrapper | HR / 労務 |
# Install
curl -fsSL https://raw.githubusercontent.com/ShunsukeHayashi/lark-harness/main/scripts/install.sh | bash
# Configure
lark-cli config init --app-id <App ID> --app-secret-stdin --brand lark
lark-cli auth login
# One-command workspace setup
larc quickstart
# Verify
larc status→ Full guide: docs/quickstart-ja.md
bin/larc # CLI entrypoint
lib/
bootstrap.sh # Context loading from Lark Drive
memory.sh # Memory sync ↔ Lark Base
ingress.sh # Task queue + execution
auth.sh # Scope inference + authorization
approve.sh # Gate policy enforcement
agent.sh # Agent registration
config/
scope-map.json # 50 task types × scopes × authority
gate-policy.json # 50 task types × risk × gate
.claude/skills/ # 25 Claude Code skills
docs/ # Guides and references
MIT