Prism

Prism is a local policy enforcement stack with:

• Management Plane (FastAPI)
• Data Plane (Rust gRPC)
• MCP server
• Web UI

Run Locally

git clone <repository-url>
cd prism

# Backend stack: data plane + MCP server + management plane
make run-all

In a second terminal:

cd ui
npm install
npm run dev

Tail Logs

tail -F \
data/logs/data-plane.log \
data/logs/mcp-server.log \
data/logs/management-plane.log \
| awk '
/data-plane\.log/       {tag="[DATA-PLANE]"; next}
/mcp-server\.log/       {tag="[MCP-SERVER]"; next}
/management-plane\.log/ {tag="[MGMT-PLANE]"; next}
{print tag " " $0}
'

Core Functionality

• Policies: create/update/toggle/delete tenant-scoped policies from the UI or API.
• Dry Run: send intent events to enforcement and inspect decision + per-slice evidence.
• Build Dataset: export reviewer feedback from dry runs as JSONL for model iteration.

APIs You Can Use

• Enforcement API: POST /api/v2/enforce
• Policies Install API: POST /api/v2/policies (creates and installs a policy into the data plane)
• Other GET APIs:
  • GET /health
  • GET /api/v2/policies
  • GET /api/v2/policies/{policy_id}
  • GET /api/v2/telemetry/sessions
  • GET /api/v2/telemetry/sessions/{agent_id}

Embeddings and Training

• Prism uses a SentenceTransformer embedding model in the management plane.
• The current encoder default in code is redis/langcache-embed-v3-small.
• Build Dataset downloads JSONL with labeled slice feedback (intent_text, anchor_text, similarity, threshold, feedback_score, rationale).
• You can use this JSONL to fine-tune or evaluate your own embedding model on Hugging Face or your own VPS training setup.

Configuration Notes

• For local backend runtime, environment values are loaded from management_plane/.env when present.
• Example deployment env template is at deployment/.env.example.