AgentGuard

Your coding agent just started looping through retries and shell calls. AgentGuard stops it before it burns budget.

Local-first runtime guardrails for coding agents. Stop loops, retry storms, and budget burn with a zero-dependency Python SDK, then expose traces and incident context through MCP when your tooling needs read access.

pip install agentguard47

Why this wedge

AgentGuard stays focused on coding-agent safety on purpose.

In an April 2026 report, a16z said that 29% of the Fortune 500 and about 19% of the Global 2000 were live paying customers of leading AI startups, with coding described as the dominant enterprise AI use case and support/search next behind it. The report also cited repeated claims of 10-20x productivity gains from AI coding tools. Source: AI Adoption by the Numbers.

That supports the public SDK strategy in this repo:

• stay narrow on coding-agent runtime safety
• make the first proof local, cheap, and easy to trust
• reuse the same runtime patterns for adjacent managed-agent workflows later, without turning the SDK into a generic observability platform

Verify your install

Before wiring a real agent, validate the local SDK path:

agentguard doctor

doctor makes no network calls. It verifies local trace writing, confirms the SDK can initialize in local-only mode, detects optional integrations already installed in your environment, and prints the smallest correct next-step snippet.

Generate a starter

When you know the stack you want to wire, print the exact starter snippet:

agentguard quickstart --framework raw
agentguard quickstart --framework openai
agentguard quickstart --framework langgraph --json

quickstart is designed for both humans and coding agents. It prints the install command, the smallest credible starter file, and the next commands to run after you validate the SDK locally.

If you want a real file instead of a printed snippet:

agentguard quickstart --framework raw --write
agentguard quickstart --framework openai --write --output agentguard_openai_quickstart.py

--write creates a local starter file you can run immediately. It refuses to overwrite an existing file unless you pass --force.

Coding-Agent Defaults

If you want humans and coding agents to share the same safe local defaults, add a tiny .agentguard.json file to the repo:

{
  "profile": "coding-agent",
  "service": "support-agent",
  "trace_file": ".agentguard/traces.jsonl",
  "budget_usd": 5.0
}

agentguard.init(local_only=True) and agentguard doctor will pick this up automatically. Keep it local and static: no secrets, no API keys, no dashboard settings.

Every agentguard quickstart --framework ... payload also has a matching runnable file under examples/starters/. Those starter files live in the repo for copy-paste onboarding and coding-agent setup; they are not shipped inside the PyPI wheel.

For the repo-first onboarding flow, see docs/guides/coding-agents.md.

For copy-paste setup snippets tailored to Codex, Claude Code, GitHub Copilot, Cursor, and MCP-capable agents, see docs/guides/coding-agent-safety-pack.md.

If you want AgentGuard to generate those repo-local instruction files for you:

agentguard skillpack --write
agentguard skillpack --target claude-code --write --output-dir .

skillpack writes a local .agentguard.json plus agent-specific instruction files for Codex, Claude Code, Copilot, or Cursor. By default it writes into agentguard_skillpack/ so you can review the files before copying them into a real repo.

MCP Server for Coding-Agent Workflows

If your coding agent already uses MCP, AgentGuard also ships a published read-only MCP server that exposes traces, decision events, alerts, usage, costs, and budget health from the AgentGuard read API:

npx -y @agentguard47/mcp-server

The MCP server is intentionally narrow. Use the SDK to enforce safety where the agent runs. Add MCP when you want Codex, Claude Code, Cursor, or another MCP-compatible client to inspect traces and incidents without bespoke glue.

Stateless Harnesses

If one managed-agent session can span multiple disposable harnesses or worker processes, pass a shared session_id to correlate those traces above the single-trace_id level:

from agentguard import JsonlFileSink, Tracer

tracer = Tracer(
    sink=JsonlFileSink(".agentguard/traces.jsonl"),
    service="managed-harness-a",
    session_id="support-session-001",
)

Each tracer instance still creates its own trace_id, but every emitted span and point event also carries the shared session_id. Guide: docs/guides/managed-agent-sessions.md

Try it in 60 seconds

No API keys. No dashboard. No network calls. Just run it:

pip install agentguard47
agentguard demo

AgentGuard offline demo
No API keys. No dashboard. No network calls.

1. BudgetGuard: stopping runaway spend
  warning fired at $0.84
  stopped on call 9: cost $1.08 exceeded $1.00

2. LoopGuard: stopping repeated tool calls
  stopped on repeated tool call: Loop detected ...

3. RetryGuard: stopping retry storms
  stopped retry storm: Retry limit exceeded ...

Local proof complete.

Prefer the example script instead of the CLI? This does the same local demo:

python examples/try_it_now.py

Quickstart: Stop a Runaway Coding Agent in 4 Lines

from agentguard import Tracer, BudgetGuard, patch_openai

tracer = Tracer(guards=[BudgetGuard(max_cost_usd=5.00, warn_at_pct=0.8)])
patch_openai(tracer)  # auto-tracks every OpenAI call

# Use OpenAI normally - AgentGuard tracks cost and kills the agent at $5

That's it. Every ChatCompletion call is tracked. When accumulated cost hits $4 (80%), your warning fires. At $5, BudgetExceeded is raised and the agent stops.

No config files. No dashboard required. No dependencies.

For a deterministic local proof before wiring a real agent, run:

agentguard doctor
agentguard quickstart --framework raw
agentguard demo

agentguard doctor verifies the install path. agentguard quickstart prints the copy-paste starter for your stack. agentguard demo then proves SDK-only enforcement with a realistic local run. Keep the first integration local and only add hosted pieces after you need retained incidents or team-visible follow-through.

The Problem

Coding agents are cheap to start and expensive to leave unattended:

• Cost overruns average 340% on autonomous agent tasks (source)
• A single stuck retry or tool loop can burn through your budget in minutes
• Existing tracing tools show you what happened after the burn, not stop the run while it is still happening

AgentGuard is built to stop a runaway coding agent mid-run, not just explain the damage later.

	AgentGuard	LangSmith	Langfuse	Portkey
Hard budget enforcement	Yes	No	No	No
Kill agent mid-run	Yes	No	No	No
Loop detection	Yes	No	No	No
Cost tracking	Yes	Yes	Yes	Yes
Zero dependencies	Yes	No	No	No
Self-hosted option	Yes	No	Yes	No
Price	Free (MIT)	$2.50/1k traces	$59/mo	$49/mo

See also: AgentGuard vs Vercel AI Gateway -- in-process SDK vs gateway proxy, compared across 7 axes.

Guards

Guards are runtime checks that raise exceptions when limits are hit. The agent stops immediately.

Guard	What it stops	Example
BudgetGuard	Dollar/token/call overruns	BudgetGuard(max_cost_usd=5.00)
LoopGuard	Exact repeated tool calls	LoopGuard(max_repeats=3)
FuzzyLoopGuard	Similar tool calls, A-B-A-B patterns	FuzzyLoopGuard(max_tool_repeats=5)
TimeoutGuard	Wall-clock time limits	TimeoutGuard(max_seconds=300)
RateLimitGuard	Calls-per-minute throttling	RateLimitGuard(max_calls_per_minute=60)
RetryGuard	Retry storms on the same flaky tool	RetryGuard(max_retries=3)

from agentguard import BudgetGuard, BudgetExceeded

budget = BudgetGuard(
    max_cost_usd=10.00,
    warn_at_pct=0.8,
    on_warning=lambda msg: print(f"WARNING: {msg}"),
)

# In your agent loop:
budget.consume(tokens=1500, calls=1, cost_usd=0.03)
# At 80% → warning callback fires
# At 100% → BudgetExceeded raised, agent stops

from agentguard import RetryGuard, RetryLimitExceeded, Tracer

retry_guard = RetryGuard(max_retries=3)
tracer = Tracer(guards=[retry_guard])

with tracer.trace("agent.run") as span:
    try:
        span.event("tool.retry", data={"tool_name": "search", "attempt": 1})
        span.event("tool.retry", data={"tool_name": "search", "attempt": 2})
        span.event("tool.retry", data={"tool_name": "search", "attempt": 3})
        span.event("tool.retry", data={"tool_name": "search", "attempt": 4})
    except RetryLimitExceeded:
        # Retry storm stopped
        pass

Integrations

LangChain

pip install agentguard47[langchain]

from agentguard import Tracer, BudgetGuard
from agentguard.integrations.langchain import AgentGuardCallbackHandler

tracer = Tracer(guards=[BudgetGuard(max_cost_usd=5.00)])
handler = AgentGuardCallbackHandler(
    tracer=tracer,
    budget_guard=BudgetGuard(max_cost_usd=5.00),
)

# Pass to any LangChain component
llm = ChatOpenAI(callbacks=[handler])

LangGraph

pip install agentguard47[langgraph]

from agentguard.integrations.langgraph import guarded_node

@guarded_node(tracer=tracer, budget_guard=BudgetGuard(max_cost_usd=5.00))
def research_node(state):
    return {"messages": state["messages"] + [result]}

CrewAI

pip install agentguard47[crewai]

from agentguard.integrations.crewai import AgentGuardCrewHandler

handler = AgentGuardCrewHandler(
    tracer=tracer,
    budget_guard=BudgetGuard(max_cost_usd=5.00),
)

agent = Agent(role="researcher", step_callback=handler.step_callback)

OpenAI / Anthropic Auto-Instrumentation

from agentguard import Tracer, BudgetGuard, patch_openai, patch_anthropic

tracer = Tracer(guards=[BudgetGuard(max_cost_usd=5.00)])
patch_openai(tracer)      # auto-tracks all ChatCompletion calls
patch_anthropic(tracer)   # auto-tracks all Messages calls

Cost Tracking

Built-in pricing for OpenAI, Anthropic, Google, Mistral, and Meta models. Updated monthly.

from agentguard import estimate_cost

# Single call estimate
cost = estimate_cost("gpt-4o", input_tokens=1000, output_tokens=500)
# → $0.00625

# Track across a trace — cost is auto-accumulated per span
with tracer.trace("agent.run") as span:
    span.cost.add("gpt-4o", input_tokens=1200, output_tokens=450)
    span.cost.add("claude-sonnet-4-5-20250929", input_tokens=800, output_tokens=300)
    # cost_usd included in trace end event

Tracing

Full structured tracing with zero dependencies — JSONL output, spans, events, and cost data.

from agentguard import Tracer, JsonlFileSink, BudgetGuard

tracer = Tracer(
    sink=JsonlFileSink("traces.jsonl"),
    guards=[BudgetGuard(max_cost_usd=5.00)],
)

with tracer.trace("agent.run") as span:
    span.event("reasoning", data={"thought": "search docs"})
    with span.span("tool.search", data={"query": "quantum computing"}):
        pass  # your tool logic
    span.cost.add("gpt-4o", input_tokens=1200, output_tokens=450)

$ agentguard report traces.jsonl

AgentGuard report
  Total events: 9
  Spans: 6  Events: 3
  Estimated cost: $0.01
  Savings ledger: exact 800 tokens / $0.0010, estimated 1500 tokens / $0.0075

When a run trips a guard or needs escalation, render a shareable incident report:

agentguard incident traces.jsonl
agentguard incident traces.jsonl --format html > incident.html

The incident report summarizes guard triggers, exact-vs-estimated savings, and the dashboard upgrade path for retained alerts and remote kill switch.

Decision Tracing

Capture agent proposals, human edits, overrides, approvals, and binding outcomes through the normal AgentGuard event path.

from agentguard import JsonlFileSink, Tracer, decision_flow

tracer = Tracer(
    sink=JsonlFileSink(".agentguard/traces.jsonl"),
    service="approval-flow",
)

with tracer.trace("agent.run") as run:
    with decision_flow(
        run,
        workflow_id="deploy-approval",
        object_type="deployment",
        object_id="deploy-042",
        actor_type="agent",
        actor_id="release-bot",
    ) as decision:
        decision.proposed({"action": "deploy", "environment": "staging"})
        decision.edited(
            {"action": "deploy", "environment": "production"},
            actor_type="human",
            actor_id="reviewer-123",
            reason="Customer approved direct rollout",
        )
        decision.approved(actor_type="human", actor_id="reviewer-123")
        decision.bound(
            actor_type="system",
            actor_id="deploy-api",
            binding_state="applied",
            outcome="success",
        )

Every decision event includes a stable schema in event.data:

• decision_id
• workflow_id
• trace_id
• object_type
• object_id
• actor_type
• actor_id
• event_type
• proposal
• final
• diff
• reason
• comment
• timestamp
• binding_state
• outcome

Guide: docs/guides/decision-tracing.md

For local JSONL traces, you can extract the normalized decision events without writing your own parser:

agentguard decisions .agentguard/traces.jsonl
agentguard decisions .agentguard/traces.jsonl --workflow-id deploy-approval --json

For retained traces exposed through MCP, use the get_trace_decisions tool to pull the same normalized decision payloads from a hosted trace by trace_id.

Evaluation

Assert properties of your traces in tests or CI.

from agentguard import EvalSuite

result = (
    EvalSuite("traces.jsonl")
    .assert_no_loops()
    .assert_budget_under(tokens=50_000)
    .assert_completes_within(seconds=30)
    .assert_total_events_under(500)
    .assert_no_budget_exceeded()
    .assert_no_errors()
    .run()
)

agentguard eval traces.jsonl --ci   # exits non-zero on failure

CI Cost Gates

Fail your CI pipeline if an agent run exceeds a cost budget. No competitor offers this.

# .github/workflows/cost-gate.yml (simplified)
- name: Run agent with budget guard
  run: |
    python3 -c "
    from agentguard import Tracer, BudgetGuard, JsonlFileSink
    tracer = Tracer(
        sink=JsonlFileSink('ci_traces.jsonl'),
        guards=[BudgetGuard(max_cost_usd=5.00)],
    )
    # ... your agent run here ...
    "

- name: Evaluate traces
  uses: bmdhodl/agent47/.github/actions/agentguard-eval@main
  with:
    trace-file: ci_traces.jsonl
    assertions: "no_errors,max_cost:5.00"

Full workflow: docs/ci/cost-gate-workflow.yml

Incident Reports

Turn a trace into a postmortem-style incident summary:

agentguard incident traces.jsonl --format markdown
agentguard incident traces.jsonl --format html > incident.html

Use this when a run hits guard.budget_warning, guard.budget_exceeded, guard.loop_detected, or a fatal error. AgentGuard will summarize the run, separate exact and estimated savings, and suggest the next control-plane step.

Async Support

Full async API mirrors the sync API.

from agentguard import AsyncTracer, BudgetGuard, patch_openai_async

tracer = AsyncTracer(guards=[BudgetGuard(max_cost_usd=5.00)])
patch_openai_async(tracer)

# All async OpenAI calls are now tracked and budget-enforced

Optional Hosted Dashboard

For teams that need retained history, alerts, and remote controls, the SDK can mirror traces to the hosted dashboard:

from agentguard import Tracer, HttpSink, BudgetGuard

tracer = Tracer(
    sink=HttpSink(
        url="https://app.agentguard47.com/api/ingest",
        api_key="ag_...",
        batch_size=20,
        flush_interval=10.0,
        compress=True,
    ),
    guards=[BudgetGuard(max_cost_usd=50.00)],
    metadata={"env": "prod"},
    sampling_rate=0.1,  # 10% of traces
)

Keep the first integration local. Add HttpSink only when you need retained incidents, alerts, or hosted follow-through.

Architecture

Your Agent Code
    │
    ▼
┌─────────────────────────────────────┐
│           Tracer / AsyncTracer       │  ← trace(), span(), event()
│  ┌───────────┐  ┌────────────────┐  │
│  │  Guards    │  │  CostTracker   │  │  ← runtime intervention
│  └───────────┘  └────────────────┘  │
└──────────┬──────────────────────────┘
           │ emit(event)
    ┌──────┼──────────┬───────────┐
    ▼      ▼          ▼           ▼
 JsonlFile  HttpSink  OtelTrace  Stdout
  Sink      (gzip,    Sink       Sink
            retry)

What's in this repo

Directory	Description	License
sdk/	Python SDK — guards, tracing, evaluation, integrations	MIT
mcp-server/	Read-only MCP surface for traces, alerts, usage, costs, and budget health	MIT
site/	Landing page	MIT

Dashboard is in a separate private repo (agent47-dashboard).

Security

• Zero runtime dependencies — one package, nothing to audit, no supply chain risk
• OpenSSF Scorecard — automated security analysis on every push
• CodeQL scanning — GitHub's semantic code analysis on every PR
• Bandit security linting — Python-specific security checks in CI

Contributing

See CONTRIBUTING.md for dev setup, test commands, and PR guidelines.

Commercial Support

Need help rolling out coding-agent safety in production? BMD Pat LLC offers:

• $500 Async Azure Audit -- cost, reliability, and governance review. No meetings. Results in 5 business days.
• Custom agent guardrails -- production-grade cost controls, compliance tooling, kill switches.

Start a project | See the research

License

MIT (BMD PAT LLC)