Build with Mithril

Integrate Mithril's governance pipeline into your stack — via MCP for agents, REST API for services, or the Python SDK for direct use. Every memory write is trust-scored, contradiction-checked, and audited.

Agent Integration

MCP Server

Expose Mithril over the Model Context Protocol. Claude Desktop, Cursor, and any MCP-aware agent can write memory through Mithril automatically.

Read guide

HTTP Endpoints

REST API

FastAPI endpoints for trust-scored admission, verified recall, audit trail, quarantine, and live source reputation. Powers the dashboard.

Read guide

Direct Integration

Python SDK

Import Mithril directly in your Python code. Set up the firewall, submit claims, query verified memory, and inspect reputation — all in a few lines.

Read guide

System Design

Architecture

The full governance pipeline — from incoming claim to admission decision. How source reputation, contradiction detection, and trust scoring work together.

Read guide

Try It Out

Run Demo

End-to-end attack simulation. Seeds legitimate policies, fires off poison attacks, and shows each admission decision with a full audit trail.

Read guide

Evaluation

Benchmark

A labeled evaluation suite measuring detection rate, false-positive rate, and precision across 30+ attack scenarios and legitimate updates.

Read guide

Quick Commands

make demoRun the attack simulation

make apiStart the REST API server

make mcpRun MCP server for agents

make benchmarkRun labeled benchmark

make devStart API + UI together

make testRun unit tests

Pipeline at a Glance

Incoming Claim  (MCP tool · Slack export · file · API · dashboard)
     │
     ▼
┌─────────────────────────────────────────────────────────┐
│  Credential Exfiltration Guard                          │
│  AWS keys, GitHub tokens, JWTs, PEM → redacted          │
└─────────────────────────────────────────────────────────┘
     │
     ▼
┌─────────────────────────────────────────────────────────┐
│  Adaptive Source Reputation (SQLite-backed, self-tuning) │
│  source_rep × 0.40                                      │
└─────────────────────────────────────────────────────────┘
     │
     ▼
┌─────────────────────────────────────────────────────────┐
│  Contradiction Detection (cognee.recall + LLM scoring)  │
│  contradiction × -0.40 · corroboration × 0.25           │
└─────────────────────────────────────────────────────────┘
     │
     ▼
┌─────────────────────────────────────────────────────────┐
│  Content Danger Check (LLM scoring, no memory needed)   │
│  dangerous content × -0.35                              │
└─────────────────────────────────────────────────────────┘
     │
     ▼
┌─────────────────────────────────────────────────────────┐
│  Trust Score = normalize(src + corroboration + fresh     │
│               - contradiction - content danger)          │
└─────────────────────────────────────────────────────────┘
     │
     ▼
┌─────────────────────────────────────────────────────────┐
│  Admission Gate                                         │
│  ≥0.85 Accept · ≥0.60 Warn · ≥0.40 Review              │
│  ≥0.20 Quarantine · <0.20 Reject                        │
└─────────────────────────────────────────────────────────┘
     │
     ├── ACCEPT / WARN  → cognee.remember(verified)
     ├── QUARANTINE / REVIEW / REJECT → SQLite quarantine
     ├── Source reputation updated from outcome
     └── All decisions → Audit log