Agentic Aegis

Sister in the Agentra ecosystem | .aegis format | 20 Capabilities | 12 MCP Tools | 30+ CLI Commands

Why AgenticAegis

Every AI agent ships code it has never validated. The generation finishes, the user runs it, and errors appear. Syntax mistakes. Type mismatches. Security vulnerabilities. Prompt injection payloads embedded in output. The validation happens after the damage -- if it happens at all.

The current fixes do not work. Linters run after generation -- they cannot catch errors mid-stream. Static analysis requires a complete file -- it cannot validate a partial token sequence. Security scanners are batch tools -- they cannot block a dangerous payload while it is still being written.

Current AI: Generates code first, validates later (if ever). AgenticAegis: Validates every token as it is generated -- syntax, types, security, and correctness are checked in real time, not after the fact.

AgenticAegis provides streaming validation -- a live shield that analyzes code as it is being generated. Not "lint your output." Your agent has a guardian -- token-level syntax validation, type flow tracking, shadow compilation, prompt injection detection, and PII scanning -- all running in parallel with generation.

Quickstart

cargo install agentic-aegis-cli
aegis --help

Problems Solved (Read This First)

• Problem: generated code has syntax errors that are only discovered when the user tries to compile. Solved: streaming syntax validation catches malformed code while the LLM is still generating -- errors are flagged before the response completes.
• Problem: type mismatches propagate silently through generated code. Solved: type flow tracking follows types across function boundaries during generation, flagging incompatible assignments before they reach the user.
• Problem: LLM output can contain prompt injection payloads. Solved: prompt injection detection scans every token for injection patterns, blocking dangerous payloads during generation.
• Problem: generated code may leak PII or sensitive data. Solved: PII detection and content filtering run on the output stream, catching sensitive data before it leaves the generation pipeline.
• Problem: there is no way to undo a generation that went wrong. Solved: session management with rollback support lets the agent revert to a known-good state when validation fails.
• Problem: validation confidence is binary -- pass or fail with no nuance. Solved: confidence scoring provides granular assessment of code quality, security risk, and correctness probability.

# Validate as you generate, protect as you ship -- three commands
aegis session create --language rust
aegis validate streaming --session <id> --input "fn main() { ... }"
aegis scan security --session <id>

Three commands. Real-time protection. Works with Claude, GPT, Ollama, or any LLM you switch to next.

How It Works

Architecture Overview

+-------------------------------------------------------------+
|                     YOUR AI AGENT                           |
|           (Claude, Cursor, Windsurf, Cody)                  |
+----------------------------+--------------------------------+
                             |
                  +----------v----------+
                  |      MCP LAYER      |
                  |   12 Tools + stdio  |
                  +----------+----------+
                             |
+----------------------------v--------------------------------+
|                    AEGIS ENGINE                              |
+-----------+-----------+------------+-----------+------------+
| Validators| Shadow    | Protection | Session   | Token      |
| (4 types) | Execution | Layers     | Manager   | Conservation|
+-----------+-----------+------------+-----------+------------+
                             |
                  +----------v----------+
                  |     .aegis FILE     |
                  |(validation session) |
                  +---------------------+

20 Capabilities

Full capability documentation ->

MCP Tools

Benchmarks

Install

git clone https://github.com/agentralabs/agentic-aegis.git
cd agentic-aegis
cargo install --path crates/agentic-aegis-cli

curl -fsSL https://agentralabs.tech/install/aegis | bash

npm install @agenticamem/aegis

Standalone guarantee: AgenticAegis operates fully standalone. No other sister, external service, or orchestrator is required.

License

MIT -- see LICENSE.