Agent Detector

agent-detector is a small, dependency-free Python package for detecting which AI coding agent is driving the current process, and for parsing that identity back out of a User-Agent header on the receiving side.

It returns evidence rather than only a boolean, so callers can distinguish an explicit identity from a broad environmental hint.

Installation

pip install agent-detector

Usage

from agent_detector import detect_agent

detection = detect_agent()

if detection:
    print(detection.agent)       # "codex"
    print(detection.confidence)  # "high"
    print(detection.signal)      # "CODEX_THREAD_ID"

The returned DetectionResult contains:

• agent: an AgentName literal containing a supported agent name
• confidence: high, medium, or low
• source: environment, path, or user-agent
• signal: the name of the matched signal, never its value

Pass a mapping to make detection deterministic in tests:

assert detect_agent({"OPENCODE": "1"}).agent == "opencode"

Require a minimum confidence when broad environmental hints are not useful:

detection = detect_agent(minimum_confidence="high")

minimum_confidence is typed as Literal["high", "medium", "low"] and defaults to "low".

User-Agent propagation

A client can propagate a detected identity to a server as a second User-Agent product. The value must be exactly two whitespace-separated products:

<product>/<version> AI-Agent/<agent>

For example:

example-cli/1.2.3 AI-Agent/codex

The rules are strict and case-sensitive:

• product and version are both required and must be non-empty RFC 9110 tokens (letters, digits, and !#$%&'*+-.^_`|~; no /, spaces, or parentheses)
• the marker must be exactly AI-Agent (mirroring the AI_AGENT environment variable)
• agent must be one of the supported agent names below, in lowercase
• comments such as (darwin) and additional products are rejected
• leading and trailing whitespace is ignored

A client can build the value like this:

from agent_detector import detect_agent

user_agent = "example-cli/1.2.3"
if detection := detect_agent(minimum_confidence="high"):
    user_agent += f" AI-Agent/{detection.agent}"

Use parse_invoking_agent on the receiving side. It returns a DetectionResult with confidence="high", source="user-agent", and signal="User-Agent", or None when the value does not match:

from agent_detector import parse_invoking_agent

detection = parse_invoking_agent(
    request.headers.get("user-agent"),
    expected_product="example-cli",
)

if detection:
    print(detection.agent)  # "codex"

expected_product is optional. When given, it must be a non-empty string and only values whose product matches it exactly are accepted; anything else returns None. Use it to ignore unrelated clients that happen to send a similar header. It is a filter, not authentication: as with environment detection, the identity is informational and any HTTP client can send it.

Supported agents

AI_AGENT takes precedence over inferred signals when its value is one of the supported agent names. Unknown values are ignored.

The detector is deliberately ordered. For example, Amp is checked before Claude Code because Amp also sets CLAUDECODE.

Important limitations

Detection is best-effort. None means unattributed, not "human". Some signals can also be present in an integrated terminal where a person typed the command manually.

This package detects the execution harness. It cannot determine whether a particular skill, plugin, prompt, or model caused the command. Use a separate explicit marker when that attribution matters.

Neither environment detection nor the User-Agent format verifies the caller. Do not use a detected identity to grant access or bypass checks.

Privacy

Environment values such as thread IDs are never returned. A result contains only a normalized agent name and the name and category of the matched signal. parse_invoking_agent likewise returns only the agent name, never the client product, version, or the raw header.

Development

uv sync --all-groups
uv run ruff check .
uv run ruff format --check .
uv run mypy
uv run pytest
uv build