approvekit 0.1.2

Human-in-the-loop approval middleware for risky AI agent actions

ApproveKit

Human approval gates for risky AI agent actions.

ApproveKit is a lightweight, framework-agnostic Python package that wraps tool calls, evaluates simple policy rules, pauses risky actions for human review, and records every outcome in SQLite audit history.

It is built for agent builders, platform engineers, and teams who need a practical local-first approval layer before giving agents access to tools like email, production writes, record deletion, or PII access.

Reviewers

Browser reviewer — approve or reject pending requests from a local web UI:

Terminal reviewer — inspect, approve, and reject from the command line:

What It Does

• Wrap any Python callable with @kit.guard.
• Auto-approve low-risk tools while still writing audit entries.
• Hold risky tool calls until a reviewer approves or rejects them.
• Default-deny on timeout.
• Redact configured payload fields before persistence.
• Review pending requests in a browser with approvekit-web.
• Keep a durable SQLite audit trail.

Install

pip install approvekit

For local development:

python3 -m pip install -e ".[dev]"

Quick Start

from approvekit import ApproveKit, Policy, Storage

policy = Policy.from_dict({
    "default_timeout": 60,
    "rules": [
        {
            "tool": "send_email",
            "require_approval": True,
            "risk_level": "high",
            "redact_fields": ["body"],
        },
        {
            "tool": "read_record",
            "require_approval": False,
            "auto_approve": True,
            "risk_level": "low",
        },
    ],
})

storage = Storage(db_path="/tmp/approvekit.db")
kit = ApproveKit(policy=policy, storage=storage)

@kit.guard
def send_email(to: str, subject: str, body: str) -> dict:
    return {"status": "sent", "to": to, "subject": subject}

send_email(
    to="[e-mail]",
    subject="Quarterly report",
    body="Sensitive content that will be redacted in storage.",
)

In a second terminal:

approvekit-web --db /tmp/approvekit.db --port 8765

Open http://127.0.0.1:8765 to approve or reject the pending request.

Guided Demo

Terminal 1:

python3 demo/agent.py --db /tmp/approvekit_demo.db --reset

Terminal 2:

approvekit-web --db /tmp/approvekit_demo.db --port 8765

The demo walks through:

• auto-approved read
• approval-required email
• PII access with redacted fields
• rejected delete
• production write that times out unless reviewed

The terminal reviewer is still available:

approvekit-review --db /tmp/approvekit_demo.db

Policy Reference

default_timeout: 60

rules:
  - tool: send_email
    require_approval: true
    timeout: 45
    risk_level: high
    redact_fields: [body]

  - tool: read_record
    require_approval: false
    auto_approve: true
    risk_level: low

  - tool: "*"
    require_approval: true
    risk_level: medium

Field	Type	Description
tool	string	Exact tool name, or * for fallback.
require_approval	bool	Whether a human decision is required before execution.
timeout	int	Seconds to wait before default-deny timeout.
auto_approve	bool	Execute immediately and write an approved audit entry.
risk_level	string	Informational label shown in reviewer UI and audit metadata.
redact_fields	list	Dict field names to mask before request/audit persistence.

Architecture

Agent tool call
  -> ApproveKit guard
  -> Policy evaluation
  -> Auto-approved path OR pending request in SQLite
  -> Web/CLI reviewer decision
  -> Approved tool execution OR rejected/timeout block
  -> Audit entry

Only approved risky requests execute the wrapped tool body. Rejected and timed-out requests are persisted and audited without executing the action.

Project Structure

approvekit/
  core.py       # guard decorator and approval wait loop
  policy.py     # policy rules, YAML/JSON loading, redaction settings
  storage.py    # SQLite request and audit persistence
  reviewer.py   # terminal reviewer
  web.py        # local browser reviewer
demo/
  agent.py      # guided two-terminal demo
tests/
  test_*.py     # approve, reject, timeout, storage, policy, web tests
docs/
  MVP_IMPLEMENTATION_PLAN.md

Development

python3 -m pip install -e ".[dev]"
python3 -m pytest -q

Maintainer Release Flow

The repository supports one-command patch/minor/major releases via release.py.

Prerequisites:

• gh CLI is installed and authenticated (gh auth login).
• GitHub repository secrets are set for Actions publishing:
  • TEST_PYPI_API_TOKEN
  • PYPI_API_TOKEN

One-command patch release:

python3 release.py patch

What it does:

• bumps version in pyproject.toml and approvekit/__init__.py
• scaffolds a new CHANGELOG.md release section
• runs tests + build + twine checks
• commits, tags, and pushes (vX.Y.Z)
• creates a GitHub Release (gh release create)

Publishing is then handled automatically by .github/workflows/publish-pypi.yml when the release is published.

Documentation

• Landing page: http://approvekit.theajaykumar.com
• Developer docs: http://approvekit.theajaykumar.com/docs.html
• MVP tracker: docs/MVP_IMPLEMENTATION_PLAN.md
• Changelog: CHANGELOG.md

License

MIT