aion-core 0.8.2

AION Core runtime security layer for AI agent tool-call control, receipts, scanning, and approvals.

AION Core

Runtime action control, receipt, and firewall layer for AI agents.

AION Core contains the open-source infrastructure pieces behind AION: Guard, Receipts, Scan, Team Policy, and the MCP Firewall.

AI Agent -> AION Guard / MCP Firewall -> Tool/API/System
                                  |
                                  +-> verified JSONL receipt log

One-Command Demo

From the repo root:

$env:PYTHONPATH='src'
python -m aion_core.demo

After local install or PyPI install:

python -m pip install aion-core
aion-demo

For editable development installs:

python -m pip install -e .

Expected result:

[PASS] scan detected unprotected MCP server
[PASS] guard blocked generic shell action
[PASS] guard allowed generic safe read
[PASS] team policy required approval
[PASS] blocked destructive shell command
[PASS] blocked secret exfiltration
[PASS] allowed safe read
Receipts written to: aion-demo-output/receipts.jsonl
Approvals written to: aion-demo-output/approvals.jsonl
Receipt verification: PASS (6 receipt(s), hash-verified)

This proves the infrastructure wedge:

• generic Guard actions can be allowed or blocked
• dangerous shell action is blocked before reaching the tool
• secret exfiltration attempt is blocked before reaching the tool
• safe file-read action is allowed
• approval-required team actions create approval records
• every decision gets a receipt

Proof Pack

AION Core includes deterministic agent-workflow proof tests that model LangChain, CrewAI, Groq function-calling, and raw MCP workflows.

$env:PYTHONPATH='src'
python examples\proof_pack\agent_workflow_proof.py

Expected result:

AION Core Agent Workflow Proof Pack
Scenarios: 6/6 passed
Receipts: 5 hash-verified

Real SDK integration tests are also included:

• LangChain 1.2.18: real StructuredTool guard test passed.
• CrewAI 1.14.4: real Agent, Task, and BaseTool guard test passed.
• Groq 1.2.0: real live function-calling test passed with llama-3.1-8b-instant.

Real-World Capacity Tests

AION Core includes end-to-end real-world capacity tests that exercise all 8 MVP layers together.

$env:PYTHONPATH='src'
python examples\real_world_capacity\release_ops_capacity_test.py --output-dir test-output\real-world-final

Expected result:

AION Real-World Capacity Test
Scenarios: 8/8 passed
Receipts: 5 hash-verified
Pending approvals: 1

Additional verified workflows:

• Hardcore single-agent customer support workflow: 9/9 passed, 6 hash-verified receipts, 1 pending approval.
• Hardcore multi-agent incident response workflow: 10/10 passed, 7 hash-verified receipts, 1 pending approval.

AION LaunchShield

LaunchShield is a browser-based first-pass scanner for AI agents and AI-built apps. Paste an agent prompt, tool list, MCP config, or launch notes to generate a risk score, risky-chain detection, launch blockers, covered security checks, an evidence log, and a downloadable report.

Public app:

https://sourabh1845.github.io/aion-core/launchshield.html

Local CLI:

aion-launchshield --project-name "My Agent" --workflow-file workflow.txt --tools-file tools.txt --mcp-config-file mcp.json --surface MCP --control receipts --output launchshield-report.md

JSON output:

aion-launchshield --workflow-file workflow.txt --tools-file tools.txt --mcp-config-file mcp.json --surface MCP --json

Pilot path:

• free launch scan for distribution
• feedback from real builders
• paid manual reviews after case studies and a clearer delivery workflow

Run Guard

Check a generic action:

$env:PYTHONPATH='src'
python -m aion_core.guard_cli check --policy examples\policies\stage6-default.json --receipt-log receipts\guard.jsonl --action-type shell.command --tool shell --arguments-file examples\actions\destructive_shell_args.json --agent-id demo --owner local

Run The Firewall

Run AION in front of any stdio MCP server:

aion-mcp-firewall --policy examples/policies/stage6-default.json --receipt-log receipts/aion.jsonl -- python path/to/mcp_server.py

For local development without installing:

$env:PYTHONPATH='src'
python -m aion_core.cli --policy examples/policies/stage6-default.json --receipt-log receipts/aion.jsonl -- python path/to/mcp_server.py

Manual Attack Demo

Blocked dangerous command:

$env:PYTHONPATH='src'
Get-Content examples/attacks/destructive_shell.json | python -m aion_core.cli --policy examples/policies/stage6-default.json --receipt-log receipts/demo.jsonl -- python examples/demo_mcp_server.py

Allowed safe call:

$env:PYTHONPATH='src'
Get-Content examples/attacks/safe_read.json | python -m aion_core.cli --policy examples/policies/stage6-default.json --receipt-log receipts/demo.jsonl -- python examples/demo_mcp_server.py

Policy Shape

Policies are JSON so the MVP has zero runtime dependencies.

{
  "default_action": "allow",
  "rules": [
    {
      "id": "block-shell-delete",
      "match": {
        "tool": ["shell", "run_command"],
        "argument_contains": ["rm -rf", "Remove-Item", "del /s"]
      },
      "action": "block",
      "reason": "Destructive shell command patterns require explicit approval."
    }
  ]
}

Supported rule matchers:

• tool: exact tool names or * wildcards.
• argument_contains: risky strings searched inside serialized arguments.
• argument_regex: risky regular expressions searched inside serialized arguments.
• owner: optional agent owner/team identity.

Supported actions:

• allow
• block

Receipt Example

Every MCP tools/call decision is logged as JSONL:

{"decision":"block","tool":"shell","rule_id":"block-shell-delete","reason":"Destructive shell command patterns require explicit approval."}

Development

Run tests:

$env:PYTHONPATH='src'
python -m unittest discover -s tests

Useful docs:

• Stage status
• AION Guard
• AION Receipts
• AION Scan
• AION Proof Pack
• Real agent test plan
• Real SDK integrations
• Real-world capacity test
• AION LaunchShield
• Team policy and approvals
• Stage 5 Cloud alignment
• AION Cloud control panel
• Stage 6 completion report
• Repo structure
• Stage 6 demo guide
• Install
• Real MCP integration
• Filesystem MCP example
• Architecture
• Launch checklist
• Launch outreach kit
• GitHub launch
• PyPI release
• Website copy
• Verification
• Demo video script
• Launch post draft
• Roadmap

Current Scope

Current core scope:

• generic Guard action checks
• MCP config and policy scanning
• stdio MCP firewall proxy
• runtime policy checks for tools/call
• MCP-compatible JSON-RPC block responses
• hash-verified JSONL audit receipts
• team approval-required policy decisions
• AION Cloud control panel summary and pending approval views
• dependency-free Python core

Next infrastructure layers:

• signed receipts
• agent identity
• cloud receipt vault
• tool risk registry
• real Slack/webhook approval delivery
• compliance exports