Agentguard

AgentGuard

AgentGuard is a self-hosted AI reliability platform for RAG and agentic applications.

It helps teams detect, evaluate, and prevent costly AI incidents before they become customer-visible.

[](LICENSE) [](https://github.com/glaborie/agentguard/actions/workflows/ci.yml) [](https://codecov.io/gh/glaborie/agentguard) [](https://www.python.org/downloads/) [](https://glaborie.github.io/agentguard/) [](docs/deployment/local.md) [](https://modelcontextprotocol.io) [](https://github.com/confident-ai/deepeval) [](https://opentelemetry.io/) [](https://github.com/glaborie/agentguard/commits/main)

Why AgentGuard

AI applications can fail in expensive ways:

• hallucinated pricing or policy answers
• unsafe or misleading outputs
• PII leakage
• regressions after prompt, model, retrieval, or tool changes

AgentGuard provides a control layer for observing, protecting, and evaluating those systems.

What it does

• Observability — traces, retrieval, latency, model behavior, and tool usage (Langfuse + Arize AX + Grafana + Jaeger)
• Protection — prompt injection blocking (regex + LLM-judge semantic pass), toxic content detection, PII masking, and agent tool-call guardrails
• Evaluation — DeepEval + RAGAS metrics, golden datasets, benchmarks, regression checks, and scoring
• Red teaming — automated adversarial probing across 4 attack types (prompt injection, jailbreak, PII extraction, system prompt leak) with CI-compatible exit codes
• Support for RAG and agents — works across both retrieval pipelines and agentic workflows

Why AgentGuard, not just observability or evals?

Production AI failures are costly.

When a RAG assistant hallucinates policy, an agent takes the wrong action, or a model response leaks sensitive data, the result is not just a bad answer — it is a business incident.

Most teams stitch together separate tools for tracing, evaluations, and guardrails. That creates gaps:

• observability shows what happened, but does not prevent it
• evaluations measure quality, but do not protect runtime behavior
• guardrails block narrow failure modes, but do not provide release confidence

AgentGuard brings these controls together in one platform so teams can:

• observe live AI behavior
• test high-risk scenarios before release
• block or reduce costly failures in production

It is built for teams that need more than model experimentation — they need operational control.

Architecture at a glance

flowchart LR
    APP[Your AI ApplicationRAG or Agent]
    AG[AgentGuard Control Layer]
    LLM[LLMs / Model Gateway]
    DATA[Knowledge, Tools, Actions]
    SAFE[Protection]
    OBS[Observability]
    EVAL[Evaluation]

    APP --> AG
    AG --> LLM
    AG --> DATA
    AG --> SAFE
    AG --> OBS
    AG --> EVAL

For a fast product-level summary, see [System overview](docs/system-overview.md).

For the full system view, see [Architecture](docs/architecture.md).

For Observability views, see [Screenshots](docs/screenshots.md).

Who it’s for

• AI engineers building RAG or agentic systems
• platform teams standardizing AI reliability
• technical product owners responsible for release confidence
• teams handling sensitive, regulated, or business-critical workflows

Use cases

• Customer support assistants — reduce hallucinated pricing, refund, and policy answers before they become customer-visible incidents.
• Internal knowledge copilots — monitor retrieval quality, evaluate answer faithfulness, and reduce sensitive data exposure.
• Agentic workflows — trace tool usage, benchmark outcomes, and catch costly action errors before production rollout.

Quick Start

cp .env.example .env
docker compose up -d
pip install -r requirements.txt
python -m app.main ingest
python -m app.main query "Does the Starter plan include SAML SSO?"

Open:

• Open WebUI: http://localhost:3100
• Langfuse: http://localhost:3200
• Arize AX: https://app.arize.com (project: agentguard)

GitHub MCP Integration

The agentguard-agent and agentguard-agent-claude-haiku models include GitHub tool access via a Model Context Protocol sidecar.

Available tools: search repositories, read file contents, list issues, list pull requests, create issues, and more (27 tools total).

Setup — add your GitHub token to .env:

GITHUB_MCP_URL=http://localhost:8091/mcp
GITHUB_PERSONAL_ACCESS_TOKEN=ghp_...   # repo + read:org scopes

The github-mcp container is profile-gated (mcp) in Compose. Start it with either:

docker compose --profile mcp up -d
# or only the MCP sidecar
docker compose --profile mcp up -d github-mcp

Try it in Open WebUI — select agentguard-agent-claude-haiku and ask:

Summarize the open issues in glaborie/agentguard

URL split: CLI uses localhost:8091; the rag-api container reaches the sidecar via http://github-mcp:8080/mcp (set in docker-compose.yml environment block, overriding .env).

Runtime Controls and Debugging

AgentGuard includes runtime controls and retrieval diagnostics exposed by the API:

• Control panel UI: http://localhost:8001/dashboard
• Read feature flags: GET http://localhost:8001/api/config
• Update feature flags: PATCH http://localhost:8001/api/config
• Reset feature flags: POST http://localhost:8001/api/config/reset
• Retrieval debug API: POST http://localhost:8001/api/retrieval/debug

CLI equivalent for retrieval diagnostics:

python -m app.main debug-retrieval "Does the Starter plan include SAML SSO?"
python -m app.main debug-retrieval "discount approval policy" --mode hybrid --json

Documentation

• [System overview](docs/system-overview.md)
• [Roadmap](docs/ROADMAP.md)
• [Architecture](docs/architecture.md)
• [Local deployment](docs/deployment/local.md)
• [Evaluation](docs/evaluation.md)
• [Repository conventions](docs/repository-conventions.md)
• [Recruiter documentation pack](docs/recruiter/README.md)
• [Agent workflow](docs/agent-workflow.md)
• [TODO / SOTA gaps](TODO.md)

Contributing

See [CONTRIBUTING.md](CONTRIBUTING.md).

Security

See [SECURITY.md](SECURITY.md).

License

Licensed under the Apache License, Version 2.0. 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: glaborie
• Source: glaborie/agentguard
• License: Apache-2.0
• Homepage: https://glaborie.github.io/agentguard/

Install and usage instructions live in the source repository linked above.