kg-tools

Give your AI agent real-time access to the Ascend knowledge graph (KG) and 600+ official Skills — operator development, model migration, performance tuning, and debugging, without guessing from model memory.

Agents KG Edges

An AI model's training data can go stale. Ascend API parameters, configuration defaults, and error-code meanings — answering them "from memory" is guessing. kg-tools connects your AI agent to the Ascend knowledge graph, a real-time authoritative data source: 6.4M+ nodes, 20M+ relationship edges, 600+ executable Skills, and 23 Plugins, covering 14 Ascend source groups + 8 Kunpeng source groups, with cross-repo links, config mappings, and error-code→source maps — real engineering knowledge beyond your training data.

Built for Ascend / Kunpeng developers (operator development, model migration, performance tuning, inference deployment, debugging) and teams that want to equip their AI agents with Ascend capabilities.

✨ Features

Type Use case What the agent does
🔍 Retrieval · look up facts API params/defaults, error-code→source map, precise code location, cross-repo impact Answers in real time, with the KG node source
⚙️ Execution · get work done Operator dev, GPU→NPU migration, perf tuning, deployment, quantization Runs 600+ official Skill workflows
🤖 Orchestration engine · auto-tune vLLM-Ascend / SGLang-NPU / MindIE latency & throughput optimization Ten-stage closed loop, no manual step-by-step tuning

🚀 Quick start (3 steps, ~60 seconds)

① Install into your AI agent

Method 1 (recommended, simplest) — send this line to your agent (Claude Code / Cursor / Codex / Cline, etc.); it will clone the repo and install itself:

Clone https://gitcode.com/agent0/kg-tools and install the ascend-kg skill. If gitcode is not reachable, ask me for an offline zip of kg-tools, extract it, and continue the install. Tell me which agent to restart when done.

Method 2 (manual):

git clone https://gitcode.com/agent0/kg-tools.git && cd kg-tools
bash install.sh ascend-kg        # install ascend-kg only; no arg = install all

Supports 9 agents. See Supported agents for details.

② Configure the API key

export ASCEND_KG_API_KEY="your-key"

No key yet? Apply free at ascend.wiki/register, or try the built-in test key: kg-test-1489df447809de55e30251bf59bf2d6c. If the env var is unset, the agent will ask which key to use on first activation.

③ Ask a question

/ascend-kg help me look up the parameters of the ReduceSum interface
/ascend-kg help me write a LayerNorm operator

Supported agents

Agent Install method Includes subagent
Claude Code skill + subagent + CLAUDE.md rule block (~/.claude/skills/ascend-kg + ~/.claude/agents/ + ~/.claude/CLAUDE.md)
Cursor rules file (.cursor/rules/ascend-kg.mdc, alwaysApply: true)
Trae rules file (.trae/rules/ascend-kg.md, alwaysApply: true)
Codex (OpenAI) rules file (AGENTS.md)
Hermes global skill (~/.hermes/skills/ascend-kg/) + HERMES.md trigger pointer
Cline rules file (.clinerules)
OpenCode global skill (~/.config/opencode/skills/ascend-kg/)
OpenClaw global skill (~/.openclaw/skills/ascend-kg/)
GitHub Copilot rules file (.github/copilot-instructions.md, requires --copilot / -Copilot to enable)
Install details: manual scripts / Windows / update / uninstall

Linux / macOS / Windows (Git Bash):

# 1. Clone and enter the repo
git clone https://gitcode.com/agent0/kg-tools.git
cd kg-tools

# 2. Run the installer
bash install.sh                   # install all tools
bash install.sh ascend-kg          # install ascend-kg only

# Uninstall / update
bash install.sh ascend-kg --uninstall   # uninstall ascend-kg
bash install.sh --update                # update all tools (git pull + reinstall)

# 3. Restart the target agent as prompted

Windows (CMD / PowerShell): double-click install.bat, or invoke it from the command line — no permissions to set.

install.bat
install.bat ascend-kg
install.bat ascend-kg -Claude
install.bat ascend-kg -Copilot
install.bat ascend-kg -Uninstall
install.bat -Update

No argument installs everything; ascend-kg installs only that plugin; -Claude / -Copilot restrict the target agent; -Uninstall removes, -Update updates. Restart the target agent after installing.

💡 Except GitHub Copilot, all agents are auto-detected: Cursor/Trae/Codex/Cline detect their directory/file in the repo, OpenCode detects the global config ~/.config/opencode, OpenClaw detects ~/.openclaw, Hermes detects ~/.hermes. Copilot is not auto-detected because .github is a shared directory — pass --copilot (bash) or -Copilot (Windows) explicitly.

💡 Claude Code users: after install, verify ~/.claude/skills/ascend-kg/SKILL.md, ~/.claude/agents/ascend-kg-worker.md, and ~/.claude/CLAUDE.md (with the ascend-kg rule block) exist; after restart, /agents should list ascend-kg-worker, and a new session raising an Ascend task should auto-trigger KG retrieval (no manual /ascend-kg needed). To update later: cd kg-tools && git pull && bash install.sh --force. For Claude Code, run /reload-plugins after changing agents/rules.

Usage example

After install, the skill triggers automatically and the agent searches the KG first. For a retrieval-type question:

You: /ascend-kg help me look up the DataCopy repeatTimes parameter

Agent: searched KG (score 0.94) →
       returns the parameter signature, default value, and notes, with the source document id.

Sample output above; actual results depend on what the search hits. The agent's answer includes the KG node id / source for verification.

Other agents (Cursor / Codex / Cline / web): have an agent with the skill installed run the "export rules" command, then paste the prompt into the target agent.

FAQ

Do I need a local NPU/GPU? No. The KG is a cloud service; you only need a network connection.

Test key vs. dedicated key? The test key is shared at 5 RPS, good for trying things out. A dedicated key has its own quota, free to apply for, and recommended for real use.

Does it work offline? No.

Project structure

kg-tools/
├── README.md
├── README.en.md
└── skills/
    └── ascend-kg/                       # ascend-kg tool directory (plain skill + subagent, not a plugin)
        ├── SKILL.md                     # skill body (rule + decision paths + API cheat sheet + Skill routing + §7 anti-forgetting + §8 engine)
        ├── claude-memory.md             # rule block injected into ~/.claude/CLAUDE.md (always-loaded decision layer)
        ├── agents/
        │   └── ascend-kg-worker.md      # subagent: carries the rule in its system prompt, resilient to context compression
        ├── engine/                      # orchestration engine (vLLM-Ascend latency-optimization loop)
        ├── references/                  # reference material (API cheat sheet / code retrieval / export template / first access)
        ├── experience/                  # accumulated experience (operator/inference/cases)
        └── evals/evals.json

Resources

Contributing

Issues and PRs welcome.