The Shield That Validates While You Generate

Streaming validation during code generation -- syntax, types, security, and correctness checked token by token, not after the fact.

Quickstart · Problems Solved · How It Works · Capabilities · MCP Tools · Benchmarks · Install · API · Concepts

---

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

Tier	Capabilities	Focus
T1: Stream Validation	Token Validator, Syntax Accumulator, Type Flow Tracker, Error Predictor	Is the code well-formed?
T2: Shadow Execution	Shadow Compiler, Sandbox Executor, Effect Tracker, Resource Monitor	Does the code do what it claims?
T3: Input Protection	Prompt Injection Detector, Intent Verifier, Payload Scanner, Rate Limiter	Is the input safe?
T4: Output Protection	Content Filter, PII Detector, Code Safety Analyzer, Output Sanitizer	Is the output safe?
T5: Session Management	Validation Session Manager, Correction Hint Generator, Confidence Scorer, Rollback Engine	Can we recover from errors?

Full capability documentation ->

---

MCP Tools

Tool	Description
aegis_validate_streaming	Validate code tokens in real time
aegis_validate_complete	Validate complete code block
aegis_shadow_execute	Shadow-execute code in sandbox
aegis_check_input	Check input for injection risks
aegis_check_output	Check output for PII and safety
aegis_session_create	Create validation session
aegis_session_status	Get session validation status
aegis_session_end	End validation session
aegis_correction_hint	Generate correction suggestions
aegis_confidence_score	Score code confidence
aegis_rollback	Rollback to previous state
aegis_scan_security	Full security scan

---

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.