v1.2.6

Hosted Ingest Compatibility

• HttpSink now drops local-only kind="meta" watermark records before posting to the hosted ingest API, preventing first-batch 400s from validators that only accept trace spans and point events.
• HttpSink now mirrors supported trace kinds into both kind and type on outbound payloads so the SDK remains compatible across hosted validators while preserving local SDK semantics.

v1.2.5

1.2.5

Distribution and Registry Hygiene

• Added official MCP Registry metadata plus package-local Docker and Smithery config for @agentguard47/mcp-server.
• Added sdk/tests/test_mcp_registry_metadata.py to keep MCP registry metadata, packaging files, and environment-variable contracts aligned.
• Refreshed README, SDK README, PyPI README, and package metadata around coding-agent safety and local-first onboarding.

Public Repo Hygiene

• Removed stale tracked context/ files that carried business-sensitive planning data not meant for the public SDK repo.
• Retired the obsolete inbox/INBOX_PROTOCOL.md workflow in favor of the current memory/ plus inbox/log.md contract.

v1.2.4

Coding-Agent Onboarding

• Added repo-local .agentguard.json support so humans and coding agents can share static SDK defaults without dashboard coupling.
• Added the built-in coding-agent profile with tighter loop and retry defaults for repo automation and coding workflows.
• Added executable starter files under examples/starters/ and aligned agentguard doctor / agentguard quickstart around .agentguard/traces.jsonl.
• Added the docs/guides/coding-agents.md onboarding guide plus doc updates across the README, SDK README, examples, architecture doc, roadmap, and generated PyPI README.

SDK Hardening

• JsonlFileSink now creates parent directories automatically so repo-local trace paths like .agentguard/traces.jsonl work out of the box.
• Repo-config parsing now rejects boolean values in numeric fields to keep local defaults deterministic and auditable.
• init() now still honors repo-level profile defaults when service, budget, or trace path are passed explicitly but guard-profile values are left implicit.
• Invalid AGENTGUARD_BUDGET_USD values now fall back to a valid repo-local budget_usd instead of silently dropping budget enforcement.

v1.2.3

AgentGuard 1.2.3

Install or upgrade:

pip install --upgrade agentguard47

What changed

• release hardening for the next SDK cut: docs, roadmap, changelog, and release criteria are now aligned
• fixed the current CodeQL findings in examples/cost_guardrail.py and sdk/tests/e2e_v110.py
• pinned GitHub Actions by commit SHA across CI, publish, CodeQL, Scorecard, and maintenance workflows
• refreshed stale docs and examples that still referenced v1.2.1 or the removed agentguard view command

Cleanup tracked separately

• #278 refresh vulnerable mcp-server dependencies
• #279 resolve remaining repo-level Scorecard governance findings

v1.2.1 — CostTracker Improvements

What's New in v1.2.1

CostTracker Improvements

• Per-span cost events: CostTracker now includes running total in span end events (cost_usd field)
• Model pricing accuracy: Updated model pricing for latest OpenAI, Anthropic, and Google models
• Warning on unknown models: estimate_cost() logs a warning when encountering unrecognized model names

Bug Fixes

• Fixed span end events not including accumulated cost data
• Improved CostTracker thread safety

Full Changelog

Full Changelog: v1.2.0...v1.2.1

---

Installation

pip install agentguard47==1.2.1

Quick Start

from agentguard import Tracer, BudgetGuard

guard = BudgetGuard(max_budget_usd=5.00)
tracer = Tracer(service="my-agent", guards=[guard])

with tracer.span("agent-run"):
    # Your agent code here
    pass  # BudgetExceeded raised if cost > $5

v1.0.0 — Production GA

AgentGuard v1.0.0 — Production GA

Runtime guardrails for AI agents. Stop loops, enforce budgets, trace everything — zero dependencies.

pip install agentguard47

Highlights

• 5 guards: LoopGuard, FuzzyLoopGuard, BudgetGuard, TimeoutGuard, RateLimitGuard
• 4 integrations: LangChain, LangGraph, CrewAI, OpenTelemetry
• Auto-instrumentation: patch_openai() / patch_anthropic() for automatic tracing
• Full async API: AsyncTracer, async decorators, async monkey-patches
• Cost tracking: Built-in pricing for OpenAI, Anthropic, Google, Mistral, Meta
• Evaluation: Assert properties of traces in tests or CI
• Production-ready: Gzip compression, retry with backoff, sampling, SSRF protection
• CLI: agentguard report, view, summarize, eval

Quality

• 317 tests, 86% coverage
• Python 3.9-3.12
• Zero dependencies (pure Python stdlib)
• Development Status: Production/Stable

Links

• PyPI: https://pypi.org/project/agentguard47/1.0.0/
• Docs: https://github.com/bmdhodl/agent47#readme
• Changelog: https://github.com/bmdhodl/agent47/blob/main/CHANGELOG.md

v0.5.0 — Cost Tracking & Dollar Budget Guards

New Features

• Cost tracking: Dollar estimates per LLM call (OpenAI, Anthropic, Gemini, Mistral)
• Dollar budget guards: BudgetGuard(max_cost_usd=5.00) — stop at a dollar threshold
• Cost in LangChain handler: Auto-estimates cost from token usage and model name
• Dashboard UI overhaul: Improved trace viewer, cost breakdowns, responsive layout
• 106 tests passing across all modules

from agentguard import BudgetGuard

guard = BudgetGuard(max_cost_usd=5.00)
guard.consume(cost_usd=0.12)  # raises BudgetExceeded when over budget

pip install agentguard47

Full changelog: v0.4.0...v0.5.0

v0.4.0 — Hosted Dashboard & Billing

New Features

• Hosted dashboard: Next.js 14 app with auth, Gantt viewer, usage tracking at app.agentguard47.com
• Billing: Stripe integration — Free / Pro ($39/mo) / Team ($149/mo)
• HttpSink: Send traces to the hosted dashboard
• API key management: Generate and revoke ag_-prefixed keys
• Landing page: agentguard47.com with pricing, FAQ, trust signals
• Rate limiting: 100 req/min per IP on ingest endpoint

pip install agentguard47

v0.3.0 — Evaluation, Auto-Instrumentation, Gantt Viewer

New Features

• Evaluation as Code: EvalSuite — chainable assertion-based trace analysis for CI
• Auto-instrumentation: @trace_agent / @trace_tool decorators, patch_openai() / patch_anthropic()
• Gantt trace viewer: agentguard view traces.jsonl opens a timeline visualization in your browser

pip install agentguard47

v0.2.0 — Initial Release

AgentGuard v0.2.0 — Initial Release

The first public release of AgentGuard: zero-dependency runtime guardrails for AI agents.

Features

• Tracing: Structured JSONL traces with span/event correlation
• Loop detection: LoopGuard catches repeated tool calls in a sliding window
• Budget enforcement: BudgetGuard with token, call count, and dollar limits
• Timeout protection: TimeoutGuard for wall-clock limits
• Recording & replay: Deterministic replay for regression testing
• CLI: agentguard report, agentguard summarize
• LangChain integration: AgentGuardCallbackHandler callback handler

Install

pip install agentguard47

Pure Python stdlib. Zero dependencies. MIT licensed.