ToolShield: A Package to Guard Your Agent

Supports: 

---

Quickstart | Pre-Generated Safety Experiences | Generate Your Own Safety Experiences |
Extend to New Tools | Safety Benchmark | Citation

ToolShield is a training-free, tool-agnostic defense for AI agents that use MCP tools. Just pip install toolshield and a single command guards your coding agent with safety experiences — no API keys, no sandbox setup, no fine-tuning. Reduces attack success rate by 30% on average.

Quickstart

pip install toolshield

Use Pre-generated Experiences

We ship safety experiences for 6 models across 5 tools, with plug-and-play support for 5 coding agents:

Inject them in one command — no need to know where files are installed:

# For Claude Code (filesystem example)
toolshield import --exp-file filesystem-mcp.json --agent claude_code

# For Codex (postgres example)
toolshield import --exp-file postgres-mcp.json --agent codex

# For OpenClaw (terminal example)
toolshield import --exp-file terminal-mcp.json --agent openclaw

# For Cursor (playwright example)
toolshield import --exp-file playwright-mcp.json --agent cursor

# For OpenHands (notion example)
toolshield import --exp-file notion-mcp.json --agent openhands

Use experiences from a different model with --model:

toolshield import --exp-file filesystem-mcp.json --model gpt-5.2 --agent claude_code

Or import all bundled experiences (all 5 tools) in one shot:

toolshield import --all --agent claude_code

You can also import multiple experience files individually:

toolshield import --exp-file filesystem-mcp.json --agent claude_code
toolshield import --exp-file terminal-mcp.json --agent claude_code
toolshield import --exp-file postgres-mcp.json --agent claude_code

See all available bundled experiences:

toolshield list

This appends safety guidelines to your agent's context file (~/.claude/CLAUDE.md, ~/.codex/AGENTS.md, ~/.openclaw/workspace/AGENTS.md, Cursor's global user rules, or ~/.openhands/microagents/toolshield.md). To remove them:

toolshield unload --agent claude_code

Available bundled experiences (run toolshield list to see all):

Model
claude-sonnet-4.5	✅	✅	✅	✅	✅
gpt-5.2	✅	✅	✅	✅	✅
deepseek-v3.2	✅	✅	✅	✅	✅
gemini-3-flash-preview	✅	✅	✅	✅	✅
qwen3-coder-plus	✅	✅	✅	✅	✅
seed-1.6	✅	✅	✅	✅	✅

More plug-and-play experiences for additional tools coming soon (including Toolathlon support)! Have a tool you'd like covered? Open an issue.

Generate Your Own Safety Experiences

Point ToolShield at any running MCP server to generate custom safety experiences:

export TOOLSHIELD_MODEL_NAME="anthropic/claude-sonnet-4.5"
export OPENROUTER_API_KEY="your-key"

# Full pipeline: inspect → generate safety tree → test → distill → inject
toolshield \
  --mcp_name postgres \
  --mcp_server http://localhost:9091 \
  --output_path output/postgres \
  --agent codex

Or generate without injecting (useful for review):

toolshield generate \
  --mcp_name postgres \
  --mcp_server http://localhost:9091 \
  --output_path output/postgres

Auto-discover Local MCP Servers

Automatically scan localhost for running MCP servers, run the full pipeline for each, and inject the results:

toolshield auto --agent codex

This scans ports 8000-10000 by default (configurable with --start-port / --end-port).

Extend to New Tools

ToolShield works with any MCP server that has an SSE endpoint:

toolshield generate \
  --mcp_name your_custom_tool \
  --mcp_server http://localhost:PORT \
  --output_path output/your_custom_tool

MT-AgentRisk Benchmark

We also release MT-AgentRisk, a benchmark of 365 harmful tasks across 5 MCP tools, transformed into multi-turn attack sequences. See agentrisk/README.md for full evaluation setup.

Quick evaluation:

# 1. Download benchmark tasks
git clone https://huggingface.co/datasets/CHATS-Lab/MT-AgentRisk
cp -r MT-AgentRisk/workspaces/* workspaces/

# 2. Run a single task (requires OpenHands setup — see agentrisk/README.md)
python agentrisk/run_eval.py \
  --task-path workspaces/terminal/multi_turn_tasks/multi-turn_root-remove \
  --agent-llm-config agent \
  --env-llm-config env \
  --outputs-path output/eval \
  --server-hostname localhost

Add --use-experience <path> to evaluate with ToolShield defense.

Repository Layout

ToolShield/
├── toolshield/              # pip-installable defense package
│   └── experiences/         # bundled safety experiences (6 models × 5 tools)
├── agentrisk/               # evaluation framework (see agentrisk/README.md)
├── workspaces/              # MT-AgentRisk task data (from HuggingFace)
├── docker/                  # Dockerfiles and compose
└── scripts/                 # experiment reproduction guides

Acknowledgments

We thank the authors of the following projects for their contributions:

• OpenAgentSafety
• SafeArena
• MCPMark

Citation

@misc{li2026unsaferturnsbenchmarkingdefending,
      title={Unsafer in Many Turns: Benchmarking and Defending Multi-Turn Safety Risks in Tool-Using Agents},
      author={Xu Li and Simon Yu and Minzhou Pan and Yiyou Sun and Bo Li and Dawn Song and Xue Lin and Weiyan Shi},
      year={2026},
      eprint={2602.13379},
      archivePrefix={arXiv},
      primaryClass={cs.CR},
      url={https://arxiv.org/abs/2602.13379},
}

License

MIT