AEVS SDK

Agent Execution Verification System — transparent audit SDK for AI agents

Documentation · Explorer · Examples · Get Credentials

---

Intercepts tool calls from supported frameworks, builds tamper-evident receipts (HMAC-signed, hash-chained), and sends them to the AEVS backend. Zero changes to your agent code.

Installation

pip install aevs

With framework extras:

pip install aevs[langchain]   # LangChain / LangGraph
pip install aevs[mcp]         # Model Context Protocol

Framework	Extra	Min version
LangChain / LangGraph	aevs[langchain]	langchain-core >= 0.2
MCP	aevs[mcp]	mcp >= 1.20

Quick Start

import aevs
from langchain_core.tools import tool

@tool
def search(query: str) -> str:
    """Search the web."""
    return f"Results for: {query}"

aevs.configure(
    api_key="aevs_sk_<key_id>_<hex_secret>",
    agent_id="<your-agent-uuid>",
)
aevs.enable()

result = search.invoke({"query": "AI news"})

refs = aevs.get_reference_ids(clear=True)
print(refs)
# [{"seq": 1, "tool_name": "search", "reference_id": "abc-123-...", ...}]

aevs.flush()
aevs.disable()

Credentials can also be set via AEVS_API_KEY / AEVS_AGENT_ID environment variables. If missing, the SDK logs a warning and runs in no-op mode — your agent keeps working, receipts just aren't recorded.

How It Works

Agent (LangChain / MCP)
  │
  ▼  tool call intercepted
ReceiptBuilder  ──▶  HMAC sign + hash chain
  │
  ▼
LocalBuffer (SQLite, encrypted at rest)
  │
  ▼  background drainer
AEVSClient  ──▶  POST /v1/receipts  ──▶  AEVS Backend

1. aevs.enable() patches your framework's tool dispatch
2. Every tool call is intercepted and a signed receipt is created
3. Receipts are buffered locally (encrypted, crash-safe)
4. A background thread flushes receipts to the AEVS backend
5. Verify any receipt using its reference_id

API Overview

aevs.configure(api_key=..., **options)   # set configuration
aevs.enable()                            # start intercepting tool calls
aevs.disable()                           # stop and restore originals
aevs.flush()                             # send buffered receipts now
aevs.get_session_id()                    # current session UUID
aevs.get_reference_ids(clear=True)       # all captured reference IDs
aevs.get_reference_id(tool_call_id)      # lookup single reference ID
aevs.is_healthy()                        # buffer write health check

See the full API reference for details.

Receipt Visibility

Control what data is included in each receipt:

Mode	Inputs & outputs	Use case
"public"	Included	Full audit — verifiers can inspect everything
"private"	Included	Signed and submitted, but restricted access (default)
"proof_only"	Stripped	Prove a tool call happened without revealing data

aevs.configure(api_key=..., agent_id=..., receipt_visibility="proof_only")

Examples

Script	What it teaches	Requirements
01_local_quickstart.py	Minimal SDK loop — invoke a tool, see AEVS capture it	AEVS credentials only
02_openai_agent.py	LangChain agent with OpenAI	OPENAI_API_KEY + AEVS
03_asi_agent.py	Same agent with ASI:One — provider-agnostic	ASI_API_KEY + AEVS

See examples/README.md for setup instructions.

Documentation

Page	Description
Getting Started	Install, configure, and capture your first receipt
Core Concepts	Receipts, hash chains, sessions, invocation tracking
Configuration	All configuration options with defaults
LangChain Integration	LangChain / LangGraph guide
MCP Integration	Model Context Protocol guide
Receipt Verification	Visibility modes and verification
Security & Privacy	Threat model and data handling
API Reference	Complete function reference
Troubleshooting	Common issues and fixes

Data & Privacy

• Receipts are buffered locally in an encrypted SQLite database
• Submitted to the AEVS backend over HTTPS
• Use receipt_visibility="proof_only" to prevent inputs/outputs from leaving the host
• AEVS is tamper-evident, not tamper-proof — it detects modification after the fact

Development

git clone https://github.com/fetchai/AEVS-sdk.git && cd AEVS-sdk
make install        # poetry install --all-extras
make check          # lint + typecheck + tests (the CI gate)

make test           # run tests
make test-cov       # tests with coverage
make lint           # ruff check
make format         # ruff format + auto-fix
make typecheck      # mypy --strict
make build          # build sdist + wheel

Contributing

See CONTRIBUTING.md for the full guide.

Security

Please do not open a public issue for security problems. See SECURITY.md for the disclosure process.

License

Apache 2.0