Vijil SDK & CLI

Measure and improve AI agent trustworthiness from your terminal or Python code.

Vijil provides a unified trust layer for the entire agent lifecycle — from discovery through deployment. Evaluate your agent against known standards, probe it under adversarial pressure with adaptive redteaming, adapt it to fix weaknesses, and protect it with runtime Guardrails.

Install

pip install vijil-sdk

Installing alongside vijil-console? Versions of vijil-console up to and including 0.1.41 also install a console script named vijil, and whichever package is installed last wins. From vijil-console 0.1.47 its command is vijil-console, and the two packages coexist cleanly. If vijil is not the CLI you expect, run pip install -U vijil-console (or uninstall it).

Requires Python 3.11, 3.12, or 3.13. Installs both the vijil CLI and the vijil Python SDK.

Using Vijil from Claude Desktop, Cursor, or other MCP-aware agent frameworks? See vijil-mcp — a separate package that exposes the Vijil platform as MCP tools. The SDK and the MCP server are independent: vijil-sdk runs in your Python process; vijil-mcp runs as a stdio server attached to your agent framework.

Migrating from the legacy vijil PyPI package (pre-2026)? See docs/migration-from-legacy-vijil.md for the install change, auth flow change, and command-by-command map.

Authenticate

Create an API key in your Vijil Console deployment: open your Profile page (/profile — click your name at the bottom of the left nav), find the API Keys card, and press Create key. You get a client ID (vk_…) and a one-time client secret. Export the pair; the SDK exchanges it for a short-lived access token automatically:

export VIJIL_CLIENT_ID="vk_..."
export VIJIL_CLIENT_SECRET="..."   # shown once at creation

Already have a bearer access token? Use it directly instead — vijil auth login saves one, or set it in the environment:

vijil auth login                   # paste an access token; saved to ~/.vijil/credentials.json
export VIJIL_API_KEY="<access-token>"

Quick start: CLI

# Evaluate against standard trust Harnesses
vijil evaluate <agent-id> --baseline

# Probe under adversarial pressure (adaptive redteaming)
vijil evaluate <agent-id> --type redteam

# View trust scores
vijil scores show <agent-id>

# Configure runtime Guardrails
vijil protect <agent-id> --guards prompt_injection,pii --mode enforce

# Improve the agent through corrective evolution
vijil adapt <agent-id> --mode config

Quick start: SDK

from vijil import Vijil

client = Vijil()

# Evaluate and wait for results
evaluation = client.evaluate("my-agent", baseline=True)
print(f"Trust score: {evaluation.trust_score}")
print(f"Reliability: {evaluation.dimensions.reliability}")
print(f"Security:    {evaluation.dimensions.security}")
print(f"Safety:      {evaluation.dimensions.safety}")

# Probe under adversarial pressure (adaptive redteaming)
status = client.evaluate("my-agent", type="redteam")
print(f"Redteam status: {status.status}")

# Configure Guardrails
dome = client.protect("my-agent", guards=["prompt_injection", "pii"])

# Improve the agent
job = client.adapt("my-agent", mode="config")

How it works

The CLI and SDK are organized into two tiers, inspired by Git:

Porcelain commands (lifecycle verbs)

High-level commands that map to stages of the agent trust lifecycle:

Plumbing commands (resource nouns)

Low-level CRUD operations on platform resources:

Setup commands

Key concepts

Evaluate types: vijil evaluate --baseline measures against known standards (like a certification exam). vijil evaluate --type redteam explores for unknown weaknesses through adaptive adversarial pressure (iterative seed → attack → judge → reflect loop). Both produce trust insights.

Adapt vs Evolve: adapt improves an existing agent through corrective evolution, fixing weaknesses found by evaluation or testing. evolve creates an entirely new agent from a spec or natural-language description.

Trust dimensions: Every evaluation produces scores across three dimensions — reliability (does it work correctly?), security (can it resist exploitation?), and safety (does it behave responsibly?).

Global options

--output, -o    Output format: table, json, yaml, quiet (default: auto)
--gateway       Gateway URL override
--api-key       API key override (or set VIJIL_API_KEY)
--quiet, -q     Quiet mode — IDs only
--verbose, -v   Debug logging

Global options are declared on the root vijil command, so they must come before the subcommand:

vijil -o table agents list      # correct
vijil agents list -o table      # error: No such option: -o

Shell composability

The CLI supports shell scripting:

# JSON output for machine consumption
vijil -o json evaluate my-agent --baseline | jq '.trust_score'

# Quiet mode returns just the ID
EVAL_ID=$(vijil -q evaluate my-agent --baseline)
vijil reports download "$EVAL_ID"

# Pipe agent list to evaluation
vijil -q agents list | xargs -I{} vijil evaluate {} --baseline --no-wait

Exit codes follow Unix convention:

Agent aliases

Set short names for agent IDs you use often:

vijil config set aliases.travel-agent d0087de1-a032-49f7-9d28-3abf4a34404d

# Now use the alias anywhere
vijil evaluate travel-agent --baseline
vijil scores show travel-agent

Framework integrations

Coming soon — Guides for integrating Vijil into agent development frameworks:

• Google ADK — Evaluate and protect ADK agents
• LangGraph — Add trust checks to LangGraph workflows
• CrewAI — Monitor and adapt CrewAI agents

See docs/integrations/ for the roadmap.

Documentation

Requirements

• Python 3.11–3.13
• A Vijil account (access to your Vijil Console deployment)

License

Proprietary. Copyright 2024-2026 Vijil AI, Inc.