Home/AI Agent Audit Trails/AEVS
A
AEVS logo

AEVS

Beta

Updated: Nov 16, 2026

Sign every AI agent tool call with a tamper-evident ECDSA receipt that auditors verify independently, added in two lines of code with no changes to your tools.

AEVS

About AEVS

Overview

AEVS is a Fetch.ai SDK that writes a signed receipt for every tool call an AI agent makes. It sits between the agent and its tools and records the tool name, inputs, output, timing, and errors, then seals each record with an ECDSA P-256 signature and links it to the previous receipt in a hash chain. The result is tamper-evident proof of what an agent actually did, rather than what its model said it did.

Developers install the open-source SDK with pip, call aevs.configure and aevs.enable, and run their agent unchanged while AEVS auto-detects LangChain or MCP. Each invocation becomes a receipt carrying a reference_id that anyone can check through the public verify API or the explorer, with no account and no access to the original codebase.

Key Benefits

  • Every tool call produces a receipt covering tool name, inputs, output, timing, and errors.
  • ECDSA P-256 signatures and a linked hash chain expose any receipt that was altered or skipped.
  • Two setup calls instrument an existing agent while leaving its tools and business logic untouched.
  • Anyone holding a reference_id verifies a receipt through the public API or explorer without re-running the agent.
  • Public, private, and proof_only modes let a team decide exactly which payloads leave its backend.
  • The SDK is open source at fetchai/AEVS-sdk and runs on Python 3.10 through 3.13.

How It Works

You install the package with pip install aevs or pip install "aevs[langchain]", then call aevs.configure with your API key and agent id and aevs.enable to turn on auto-detection of LangChain or MCP. As the agent runs, AEVS catches each tool invocation, signs it with an ECDSA P-256 key held in KMS, and appends it to the chain; a flush call closes out the session. You then share the reference_id, and a reviewer opens the explorer or sends a GET request to the verify endpoint to confirm the signature and the chain.

Use Cases

  • AI engineers wrap an existing LangChain or MCP agent to get signed receipts without rewriting tool code.
  • Auditors and compliance reviewers confirm that a tool ran with the exact inputs and output by checking a reference_id in the verify API.
  • CI pipelines and scripts call the verify endpoint with a reference_id to assert that an agent run produced the expected tool calls.
  • Teams working with PCI or health data use proof_only mode to keep only input and output hashes, so raw payloads never persist.
  • Developer advocates and demo teams set receipt_visibility to public and share full tool payloads with reviewers in the explorer.
  • Platform teams building payment or refund agents prove that approve_payment fired with the amounts they expect.

Why Choose This Product

AEVS targets the gap between a model's chat reply and a verifiable execution record, so auditors check a signature instead of reconstructing behaviour from logs. Verification runs through a public endpoint or the explorer, which means a reviewer needs only a reference_id, and the strictest mode strips payloads before ingest so regulated data never lands in storage. The SDK is open source and adds no changes to your existing tools, though the beta label warns that APIs and the explorer may still change.

AEVS Pros & Cons

Strengths
  • Signs every tool call with ECDSA P-256 keys held in KMS
  • Public verify API needs no account and no codebase access
  • proof_only mode stores hashes only, never raw payloads
  • Auto-detects LangChain and MCP on instrumentation
  • Open source SDK with a two-line setup
Limitations
  • In beta โ€” APIs and the explorer may change
  • SDK targets Python 3.10โ€“3.13 only

Key Features

โœ๏ธ

ECDSA Signing

AEVS signs each tool call receipt with an ECDSA P-256 key held in KMS.

๐Ÿ”—

Hash Chain

Each receipt links to the previous one, so the chain reveals any record that was altered or skipped.

โšก

Two-Line Setup

Two calls, aevs.configure and aevs.enable, instrument an existing agent without changing its tools.

๐Ÿค–

Framework Auto-Detect

AEVS auto-detects LangChain and MCP and stamps the framework version onto each receipt.

๐Ÿ”

Independent Verify API

A GET request to /v1/receipts/verify/<reference_id> returns verified status without requiring an account.

๐Ÿ—‚๏ธ

Receipt Explorer

The explorer lets you search by reference_id or receipt_id and inspect verification, chain, and tool details.

๐Ÿ•ถ๏ธ

Visibility Controls

Per-agent settings choose public, private, or proof_only payload exposure plus whether the agent appears on the public explorer.

Frequently asked questions about AEVS

What is AEVS?

AEVS is a drop-in SDK from Fetch.ai that records and signs every AI agent tool call. It captures the tool name, inputs, output, timing, and errors, then seals each record with an ECDSA P-256 signature and links it into a hash chain. Developers share a reference_id so auditors verify the call independently.

How do I add AEVS to an existing agent?

Install the package with pip install aevs, or pip install "aevs[langchain]" for LangChain support. Then call aevs.configure with your API key and agent id, and aevs.enable, which auto-detects LangChain or MCP. Run your agent as usual and every tool call produces a signed receipt; call aevs.flush at the end.

Which frameworks and Python versions does AEVS support?

The page lists LangChain 0.2 and later and MCP 1.20 and later, running on Python 3.10 through 3.13.

Can I verify a receipt without an AEVS account?

Yes, and no account is required. Send a GET request to https://api.aevs.fetch.ai/v1/receipts/verify/<reference_id> and pass only the reference_id. The response reports verified true with the tool name and status. The explorer shows the same signed proof for anyone who prefers a UI.

Who can see my receipt payloads?

You decide. In public mode the explorer and verify API expose full tool payloads to anyone with the reference_id. In private mode inputs and outputs stay in the backend for your account while public views show REDACTED, yet tool name, status, timestamps, signatures, and chain remain visible.

What if I cannot store the inputs and outputs at all?

Use proof_only, the strictest setting. AEVS strips inputs and outputs before ingest and keeps only input_hash, output_hash, and chain data, so no payload exists to retrieve, even for the account owner. Fetch.ai suggests it for PCI, health, and other data that must not be persisted while still proving the tool ran.