Safe Mcp Agent
A local LLM agent (LangGraph + MCP) with a red-teamed guardrail layer and an automated evaluation harness that benchmarks security and task performance across models.
Install
$ agentstack add mcp-jeneesh1014-safe-mcp-agent ✓ scanned · ✓ verified, works with Claude Code, Cursor, and more.
Security review
✓ PassedNo issues found. Passed automated security review. · v0.1.0 How review works →
- ✓ Prompt-injection patterns
- ✓ Secret / credential exfiltration
- ✓ Dangerous shell & filesystem operations
- ✓ Untrusted network calls
- ✓ Known-malicious package signatures
What it can access
- ✓ Network access No
- ✓ Filesystem access No
- ✓ Shell / process execution No
- ✓ Environment & secrets No
- ✓ Dynamic code execution No
From automated source analysis of v0.1.0. “Used” means the capability is present in the source — more access means more to trust, not that it’s unsafe.
Verified badge
Passed review? Show it. Paste this badge into your README, it links to the public security report.
Links to your public security report.[](https://agentstack.voostack.com/security/report/mcp-jeneesh1014-safe-mcp-agent) Reliability & compatibility
Declared compatibility
Compatibility is declared by the source manifest. End-to-end runtime verification is coming, see below.
We're building live execution health for every listing: tool-call success rate, median latency, uptime, and last-checked timestamps, measured, not self-reported. It isn't live yet, so we don't show numbers we can't stand behind.
How agent discovery & health will work →About
Safe MCP Agent
A local LLM agent (LangGraph + MCP) with a red-teamed guardrail layer and an automated evaluation harness that benchmarks security and task performance across models.
> 🚧 Work in progress — Week 0 scaffolding complete, implementation starting.
What this is
Four things that fit together:
| Layer | What it does | |---|---| | MCP server | Exposes 3 mock enterprise tools an agent can call | | LangGraph agent | Decides which tools to call and in what order | | Guardrail middleware | Intercepts every tool call, validates it against a threat model, logs what it blocks | | AgentEval harness | Automatically attacks the agent using real SAFE-MCP technique IDs and measures how often the shield stops them |
The end state is a pytest suite you can run that produces a report: which attacks got through, which got blocked, and how that changes when you swap the underlying model.
Quick start
> Prerequisites: Ollama installed and running, Poetry installed.
git clone https://github.com//safe-mcp-agent.git
cd safe-mcp-agent
poetry install
ollama pull llama3.2 # or whichever model you prefer
python scripts/seed_fixtures.py
pytest tests/ -m "not slow"Running the agent
poetry run python -m reference_system.agentRunning the security suite
pytest tests/test_security.py --agenteval -vProject structure
safe-mcp-agent/
├── reference_system/ the agent under test (MCP server + LangGraph + guardrail)
├── agenteval/ evaluation library — published separately as mcp-guardeval
├── attacks/ red-team scripts, one file per SAFE-MCP technique
├── tests/ pytest suite (behavior + security)
├── scripts/ setup helpers (seed_fixtures.py, etc.)
├── dashboard/ optional Next.js trace viewer
└── docs/ planning documentsSee [docs/PROJECT.md](docs/PROJECT.md) for the full project rationale, [docs/ARCHITECTURE.md](docs/ARCHITECTURE.md) for system design, and [docs/ROADMAP.md](docs/ROADMAP.md) for the week-by-week plan.
Hardware constraint
Ollama runs natively on the host (not in Docker) to access the Mac GPU via Metal.
License
MIT — see [LICENSE](LICENSE).
Source & license
This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.
- Author: Jeneesh1014
- Source: Jeneesh1014/safe-mcp-agent
- License: MIT
Install and usage instructions live in the source repository linked above.
Reviews
No reviews yet, be the first.
Write a review
Versions
- v0.1.0 Imported from the upstream source.