TxVeto

TxVeto is an in-process budget and loop guard for AI agents. It is designed to stop runaway API spend and repeated tool-call loops before the next expensive call is made.

Why TxVeto

Agentic systems fail fast when guardrails are missing. A prompt injection, loop bug, or tool misuse can trigger expensive cascades in minutes.

TxVeto adds runtime safety primitives you can enforce locally:

• Hard budget ceilings per run
• Max-step loop circuit breakers
• Repeated tool-call detection
• MCP policy controls for guarded tool execution

What's included

• txveto.guard.VetoGuard for budget and loop enforcement
• txveto.errors for typed safety exceptions
• txveto.policy.BudgetPolicy for MCP policy configuration
• A browser playground in txveto.web_demo
• A JSON-RPC MCP server wrapper in txveto.mcp_server
• Pytest coverage for wallet-drain and infinite-loop scenarios

Install

Local development

pip install -e .[dev]
pytest

Package usage

# after publish
pip install txveto

Quick Start

python -m txveto.demo --mode both
python -m txveto.web_demo

Installed console scripts:

• txveto-demo --mode both
• txveto-web-demo
• txveto-mcp

Python API Example

from txveto.guard import VetoGuard

guard = VetoGuard(max_usd=0.50, max_steps=10)

with guard:
    guard.inspect_step(
        model="claude-3-5-sonnet",
        input_tokens=50_000,
        output_tokens=2_000,
        tool_name="fetch_web_data",
    )

MCP Tools

The MCP server exposes policy-aware tools for inspection and runtime configuration. budget.configure updates the active per-session policy, budget.status returns the current spend and policy snapshot, and policy.evaluate lets clients simulate a step before executing it.

Launch Artifacts

• Changelog: CHANGELOG.md
• Release checklist: RELEASE_CHECKLIST.md

License

MIT