agenthint

agenthint

Detect AI agent runtimes and adapt CLI output.

agenthint is a small detection spec, CLI, and multi-language library for developer tools that want to know when they are probably being run by an AI agent such as Codex, Claude Code, Cursor, Gemini CLI, Aider, or another automated coding environment.

Use it to choose better defaults for agent-driven runs: structured output, quiet logs, no spinners, no pagers, no interactive prompts, and clearer diagnostics.

Detection is advisory. agenthint is for user experience decisions, not authentication, authorization, sandboxing, or policy enforcement.

Quick Start

Install the CLI:

npm install -g agenthint
# or
cargo install agenthint
# or
python3 -m pip install agenthint

Use its exit code in scripts:

if agenthint >/dev/null; then
  exec my-tool --json --no-progress --no-pager "$@"
else
  exec my-tool "$@"
fi

Prefer the explicit convention when you control the agent or wrapper:

AI_AGENT=codex my-tool
AI_AGENT=claude-code my-tool
AI_AGENT=my-custom-agent my-tool

Why

Humans and agents often need different CLI behavior.

agenthint gives tools a shared, explainable way to switch modes without each project inventing its own agent detection logic.

CLI

agenthint             # exit 0 if an agent is likely detected, otherwise 1
agenthint --json      # print the structured detection result
agenthint --explain   # print a short human-readable explanation
agenthint doctor      # print detection details and setup advice
agenthint doctor --json
agenthint init codex  # print the recommended AI_AGENT value
agenthint --version   # print the version

Example JSON:

{
  "isAgent": true,
  "agent": "codex",
  "confidence": 0.92,
  "signals": ["env:CODEX_CI", "env:CODEX_THREAD_ID"]
}

Exit codes:

Setup-only commands such as agenthint init <agent> exit 0.

Libraries

TypeScript

import { detectAgent } from "agenthint";

const result = detectAgent();

if (result.isAgent) {
  // Prefer structured, quiet, non-interactive output.
}

Rust

use agenthint::detect_agent;

let result = detect_agent();

if result.is_agent {
    // Prefer structured, quiet, non-interactive output.
}

Python

from agenthint import detect_agent

result = detect_agent()

if result.is_agent:
    # Prefer structured, quiet, non-interactive output.
    pass

Install

npm

npm install -g agenthint
agenthint --json

crates.io

cargo install agenthint
agenthint --json

PyPI

python3 -m pip install agenthint
agenthint --json

Native binary

curl -fsSL https://raw.githubusercontent.com/forjd/agenthint/main/install.sh | sh

The install script downloads the latest agenthint-v* GitHub Release asset for your platform and verifies it against SHA256SUMS.

Override the install directory or version:

AGENTHINT_INSTALL_DIR=/usr/local/bin sh install.sh
AGENTHINT_VERSION=agenthint-vX.Y.Z sh install.sh
AGENTHINT_ALLOW_MISSING_CHECKSUM=1 sh install.sh

Detection Model

Every detection result includes:

Detection priority:

1. AGENTHINT_DISABLE
2. AGENTHINT_FORCE (optionally named via AGENTHINT_AGENT)
3. Explicit AI_AGENT
4. Known environment signals
5. Documented filesystem signals
6. Low-confidence parent process signals
7. Low-confidence stdio hints

Known agents include Codex, Claude Code, Cursor, Gemini CLI, Aider, Augment CLI, AMP, OpenCode, OpenClaw, GitHub Copilot, Replit, Devin, Google Antigravity, Pi, Kiro CLI, Windsurf, Cline, Roo Code, Kilo Code, Mistral Vibe, v0, and Cowork.

Some signals are weak. For example REPL_ID is present in every Replit workspace, including human-driven sessions, so it reports a low confidence. Exit-code consumers that want to avoid false positives can read confidence from agenthint --json and apply their own threshold.

Custom agents are supported through any non-empty AI_AGENT value.

Docs

• Agent integration notes: recommended AI_AGENT values
• Integration snippets: Bash, Zsh, Fish, Node.js, Rust, and Python examples
• Signal registry: known signals and confidence levels
• Detection spec: the portable detection contract

Principles

• Prefer explicit AI_AGENT support over heuristics.
• Return confidence, not false certainty.
• Print signal names, not environment variable values.
• Keep filesystem probes documented and configurable.
• Keep requested machine-readable output quiet and stable.
• Treat detection as a hint, never as a security boundary.

Development

mise install
mise exec -- npm run check

Useful scripts:

npm run build
npm run format
npm run lint
npm run test
npm run check
cargo test --workspace

Contributions are welcome. Please keep detection results explainable, avoid printing secret-bearing environment values, and update the docs when adding or changing signals.