AgentGuard

Stop runaway agents before they burn money.

Zero-dependency Python kill switch for AI agents. Hard budget caps. Loop detection. Local traces. MIT.

pip install agentguard47

Getting started

1. Install and verify

pip install agentguard47
agentguard doctor   # package ok?
agentguard demo     # offline proof (no API keys)

2. Guard an OpenAI client

from agentguard import BudgetGuard, LoopGuard, Tracer, patch_openai

budget = BudgetGuard(max_cost_usd=5.00, warn_at_pct=0.8)
loop = LoopGuard(max_repeats=3)
tracer = Tracer(service="my-agent", guards=[loop])

patch_openai(tracer, budget_guard=budget)
# every OpenAI call is now traced + budget-enforced

When spend crosses the hard limit, BudgetExceeded is raised and the run stops.

3. Cap a single task

Session budget can still have headroom. One goal can still be killed:

with budget.goal("refund", max_cost_usd=0.50, warn_at_pct=0.8) as g:
    g.attempt()
    budget.consume(cost_usd=0.12)
    # BudgetExceeded names the goal when it crosses

4. Read the local proof

agentguard report .agentguard/traces.jsonl
agentguard incident .agentguard/traces.jsonl

Or scaffold a starter file:

agentguard quickstart --framework raw --write
python agentguard_raw_quickstart.py

What it stops

Not a dashboard. Not a model router. An in-process exception that kills the bad run mid-flight.

Cap your agent's x402 wallet spend

Agents that pay per-call via x402 (USDC micropayments) can drain a wallet in a silent loop. X402SpendGuard wraps the payment step and refuses before paying:

from agentguard import X402SpendGuard

guard = X402SpendGuard(
    max_total_usd=5.00,        # wallet cap, add period="day" for a daily reset
    max_per_endpoint_usd=1.00, # cap per resource URL
    max_per_call_usd=0.10,     # refuse any single payment above this
)
guard.charge(0.001, "https://api.example.com/search", my_x402_pay_step)

AgentGuard meters and refuses; it never signs or settles. Amounts come from your x402 client. No crypto dependencies.

Features

• Hard stops — exceptions inside your process, not after-the-fact alerts
• Task-level budgets — BudgetGuard.goal(...) for sub-task caps + warn hooks
• Local traces — JSONL by default; no network unless you opt in
• Zero deps — stdlib only; Python 3.9+
• Provider patches — patch_openai / patch_anthropic
• Framework hooks — LangChain, LangGraph, CrewAI (optional extras)

Local by default

• No API key required for local proof
• No network unless you configure HttpSink
• MIT licensed

The SDK is the free local proof path. Start local. Add hosted ingest later only if you want retained history, alerts, team visibility, spend trends, hosted decision history, or dashboard-managed remote kill signals. Local guards remain authoritative. HttpSink mirrors trace and decision events; it does not execute remote kill signals by itself.

Integrations

OpenAI · Anthropic · LangChain · LangGraph · CrewAI · raw agent loops

pip install "agentguard47[langchain]"   # optional extras as needed

Security

The base install declares zero runtime dependencies. pip install agentguard47 pulls nothing, so a default install adds no third-party exposure.

Extras install third-party packages and need a separate audit. The LangChain and LangGraph extras now require Python 3.10+ and raise their minimum versions to the tested September 2026 releases. OpenTelemetry requires 1.44.0 or newer. The base SDK remains compatible with Python 3.9+.

The optional [crewai] extra requires CrewAI 1.15.21 or newer. Its current dependency tree still installs ChromaDB 1.1.1, with four distinct unresolved advisories: CVE-2026-45829, CVE-2026-45830, CVE-2026-45831, and CVE-2026-45833. The audit database provides no fixed version. Avoid this extra unless you have reviewed that upstream exposure. Installing AgentGuard alone does not install ChromaDB or start a server. See the upstream advisory and the release audit.

HttpSink validates the address it actually connects to, retains TLS hostname verification, and refuses cross-origin redirects. It connects directly and does not use environment proxy settings. A local guard stops instrumented work in your Python process; it does not cancel an agent loop running on a provider's server. Cost estimates are not invoices; supply provider-reported cost or use strict cost resolution when an estimate is insufficient.

Docs

• Getting started guide
• Examples
• MCP server — npx -y @agentguard47/mcp-server

Links

• PyPI: https://pypi.org/project/agentguard47/
• Issues: https://github.com/bmdhodl/agent47/issues
• AgentGuard on the web (hosted history, alerts, and MCP visibility for Claude Code, Cursor, and Codex): https://bmdpat.com/tools/agentguard?utm_source=agentguard47&utm_medium=readme&utm_campaign=touchpoints

The hosted page is an optional next step, not a requirement. The SDK stays free, local, and MIT, and the local guards stay authoritative. Nothing in this package phones home. The only network egress is a sink or exporter you configure yourself, such as HttpSink or an OpenTelemetry exporter.

MIT · Built for people who ship agents and hate surprise bills.

Latest Release Notes (1.3.0)

(2026-09-12)

This release includes the accumulated, unpublished 1.2.14 candidate work below.

Security and enforcement fixes

• LangChain now propagates guard exceptions through its real callback manager. A zero-call budget stops the tool before its body runs; previously LangChain could log the exception and continue. Sync and async dispatch run inline.
• Budget and timeout caps reject invalid, negative, boolean, and non-finite values. Corrupt stored budget counters fail closed without rewriting state. Warning callbacks run outside budget locks and zero limits do not divide by zero.
• Failed x402 payment callbacks refund only their original budget generation, so a reset or day rollover cannot reduce a later period's spending.
• HTTP trace delivery rejects credential-bearing URLs, cross-origin redirects, mapped private IPv6 addresses, and private/reserved DNS answers at connection time. Connections use the validated address while TLS retains hostname checks. This transport deliberately does not use environment proxies.
• Retry-After delays are finite, non-negative, and capped at 30 seconds.
• MCP dependency updates resolve the npm audit findings in the committed lockfile.

Optional dependency compatibility

• LangChain requires 1.6.3+, LangGraph 1.2.11+ with checkpoint 4.2.0+ and SDK 0.4.4+, OpenTelemetry 1.44.0+, and CrewAI 1.15.21+.
• LangChain and LangGraph extras require Python 3.10+. The dependency-free base package remains compatible with Python 3.9+.
• The optional CrewAI tree still installs ChromaDB with four distinct unresolved advisories (CVE-2026-45829, CVE-2026-45830, CVE-2026-45831, CVE-2026-45833). No fixed upstream version was available in the audit. Avoid this extra unless its exposure has been reviewed. Base installs do not include ChromaDB.
• Audit scope, regression results, dependency resolutions, and limitations: September audit.

Reliability

• Added the file-backed JsonFileStateStore integration for BudgetGuard(store=...), so configured budget usage can persist across processes and scheduled tasks. This is local persistence, not distributed coordination or a fairness guarantee.
• Hardened the cross-process state lock (JsonFileStateStore, used by BudgetGuard(store=...)) against two Windows races that crashed concurrent processes under contention: an exclusive lock create that fails with PermissionError instead of FileExistsError during a concurrent release ("delete pending"), and an os.replace that transiently fails with access-denied when an antivirus/indexer holds the destination. Both now retry safely, so cross-process budget enforcement holds on Windows scheduled tasks.

Budget Goals

• Added BudgetGuard.goal(...) for scoped per-goal caps on tokens, calls, and cost, with an optional warn_at_pct threshold and on_warning callback. Goal warnings are emitted once per goal while hard caps still refuse excess spend.

Payment Guardrails

• Added X402SpendGuard for local caps on total, per-endpoint, and per-call x402/USDC spend. It checks and reserves configured spend before payment and rolls the reservation back if the payment callback raises. It does not settle x402 payments or add a crypto dependency.

Cost Accounting

• Added maximum-precision billable-cost resolution with explicit source labels for provider-reported values, caller prices, estimates, zero-cost tool/local work, and unknown cost. Unknown usage stays conservative or fails in strict mode; the result is not a provider invoice.

Usage Accounting

• Anthropic usage normalization now preserves thinking/reasoning tokens and separates them from answer tokens when the provider payload exposes that detail, alongside cache-read and cache-write fields.

Hardening

• Rejected NaN, infinite, and negative budget inputs before state mutation so non-finite values cannot bypass a cost ceiling.
• Made LoopGuard argument fingerprinting tolerate non-JSON-serializable tool arguments instead of crashing the guard while it checks for repeats.

Public Docs

• Made the reader-facing surface fully model-agnostic to match the already-vendor-neutral code path: the README/PyPI "As a skill" heading now leads with Codex alongside Claude Code, and the budget-aware escalation example notes the escalate target can be any provider's model, not just Claude.

Onboarding

• Bare agentguard now prints a friendly first-run welcome with the 60-second local path and the star call to action instead of an argparse help dump.
• Added python -m agentguard as an entry point so the CLI works even when the agentguard script is not on PATH.
• Added agentguard welcome and agentguard badge. badge prints a paste-able "Guarded by AgentGuard" README badge (markdown, rST, or HTML) so adopters can advertise the SDK and drive new installs.

Distribution

• Added an opt-in bridge to the hosted AgentGuard page (bmdpat.com/tools/agentguard) from the README/PyPI page, the agentguard --help footer, and the first-run welcome. These are static links only: the SDK still makes no network calls unless you configure HttpSink, and nothing in the package phones home. The links carry UTM parameters so the site can measure click-through; no identifier is sent from your machine.

Full changelog: CHANGELOG.md