waxell-observe 0.0.30

Lightweight observability & governance for any AI agent framework

waxell-observe

Lightweight observability and governance for any AI agent framework. Add production-grade tracing, cost tracking, and policy enforcement to your agents with a single line of code.

Installation

# Core (HTTP telemetry only)
pip install waxell-observe

# With OpenTelemetry tracing (recommended)
pip install "waxell-observe[otel]"

# With infrastructure auto-instrumentation (HTTP, databases, caches, queues)
pip install "waxell-observe[infra]"

# With LangChain integration
pip install "waxell-observe[otel,langchain]"

# Everything
pip install "waxell-observe[all]"

# Development (includes test dependencies)
pip install "waxell-observe[dev,otel]"

Quick Start

import waxell_observe

# One-line init — configures tracing, AI/ML instrumentation, and infrastructure instrumentation
waxell_observe.init(api_key="wax_sk_...")

# That's it. All LLM calls, HTTP requests, database queries, and cache operations
# are now traced automatically.

After calling init(), every LLM call, HTTP request, database query, and cache operation made by your agent is automatically captured — model names, token counts, latency, SQL statements, HTTP endpoints, Redis commands — all nested in a trace tree visible in the Waxell dashboard.

Auto-Instrumentation

AI/ML Libraries

init() automatically detects and instruments installed AI/ML libraries:

Library	What's Captured
OpenAI SDK	All chat.completions.create() calls — model, tokens, latency, streaming
Anthropic SDK	All messages.create() calls — model, tokens, latency
LangChain	Chain executions, LLM calls, tool invocations via callback handler
+ 145 more	Bedrock, Gemini, Mistral, Cohere, Groq, Pinecone, Chroma, etc.

To instrument only specific AI/ML libraries:

waxell_observe.init(api_key="wax_sk_...", instrument=["openai"])

Infrastructure Libraries

When installed with the [infra] extra, init() also instruments infrastructure libraries to capture everything your agent touches:

HTTP Clients:

Library	Span Examples
httpx	POST api.openai.com, GET api.example.com
requests	POST api.stripe.com, GET api.weather.com
urllib3	Lower-level HTTP spans
aiohttp	Async HTTP client spans

Databases:

Library	Span Examples
psycopg2 / psycopg	pg SELECT, pg INSERT
asyncpg / aiopg	pg SELECT (async)
SQLAlchemy	pg SELECT users, mysql INSERT orders
PyMongo	mongo FIND, mongo AGGREGATE
PyMySQL / mysqlclient	mysql SELECT, mysql INSERT
pymssql	mssql SELECT, mssql INSERT
sqlite3	sqlite SELECT, sqlite CREATE
Elasticsearch	es SEARCH, es INDEX
Cassandra	cassandra SELECT

Caching:

Library	Span Examples
redis	redis GET, redis SET, redis HGET
pymemcache	memcache GET, memcache SET

Message Queues & Task Brokers:

Library	Span Examples
Celery	apply_async/task, run/task
kafka-python / confluent-kafka	kafka send events, kafka receive events
pika / aio-pika	RabbitMQ publish/consume
aiokafka	Async Kafka spans

Cloud & RPC:

Library	Span Examples
botocore	AWS SDK calls (aws s3.PutObject, aws dynamodb.GetItem)
boto3 (SQS)	aws sqs.SendMessage
gRPC	grpc UserService.GetUser

Example Trace Tree

With both AI/ML and infrastructure instrumentation enabled:

agent: rag-demo.document-qa (3200ms)
├── chain: analyze_query
│   ├── llm: chat gpt-4o (800ms)
│   │   └── tool: POST api.openai.com (790ms)
│   └── tool: redis GET session:abc (5ms)
├── chain: retrieve_documents
│   ├── tool: pg SELECT * FROM documents WHERE ... (15ms)
│   └── tool: POST pinecone.io/query (200ms)
├── chain: synthesize_answer
│   └── llm: chat gpt-4o (1000ms)
│       └── tool: POST api.openai.com (990ms)
└── [governance summary]

Controlling Infrastructure Instrumentation

# Default: auto-detect and instrument everything available
waxell_observe.init()

# Disable all infrastructure instrumentation
waxell_observe.init(instrument_infra=False)

# Only instrument specific libraries (allowlist)
waxell_observe.init(infra_libraries=["redis", "httpx", "psycopg2"])

# Instrument everything except specific libraries (blocklist)
waxell_observe.init(infra_exclude=["celery", "grpc"])

Coexistence with Existing OTel

If your application already uses OpenTelemetry instrumentation, waxell-observe works alongside it:

• No double-patching: OTel instrumentors have a singleton guard. If you already called RedisInstrumentor().instrument(), our call is a safe no-op.
• Additive: We add our span processor to the global TracerProvider alongside your existing exporters (Datadog, Jaeger, Honeycomb, etc.). Your spans still flow to your backend AND also appear in Waxell.
• Context-gated: Outside of agent runs, our processor does nothing — your app's normal telemetry is untouched.

Manual Instrumentation

@waxell_agent Decorator

Wrap any function to create an observed agent run:

from waxell_observe import waxell_agent

@waxell_agent(agent_name="my-agent")
async def run_agent(query: str, waxell_ctx=None) -> str:
    # waxell_ctx is automatically injected when declared as a parameter
    response = openai_client.chat.completions.create(
        model="gpt-4o",
        messages=[{"role": "user", "content": query}],
    )

    # Record steps for hierarchical trace visualization
    waxell_ctx.record_step("llm_response", {"model": "gpt-4o"})

    return response.choices[0].message.content

WaxellContext Context Manager

For more control, use the context manager directly. Works as both async with and plain with:

from waxell_observe import WaxellContext

# Async
async with WaxellContext(agent_name="my-agent") as ctx:
    result = await my_agent.run(query)
    ctx.record_llm_call(model="gpt-4o", tokens_in=150, tokens_out=80)
    ctx.record_step("retrieval", {"documents": 5})
    ctx.set_result({"output": result})

# Sync — ideal for batch scripts, CLI tools, ETL pipelines
with WaxellContext(agent_name="my-agent") as ctx:
    result = my_agent.run(query)
    ctx.record_llm_call(model="gpt-4o", tokens_in=150, tokens_out=80)
    ctx.record_step("retrieval", {"documents": 5})
    ctx.set_result({"output": result})

LangChain Integration

from waxell_observe.integrations.langchain import create_langchain_handler

handler = create_langchain_handler(agent_name="my-langchain-agent")
chain.invoke(input, config={"callbacks": [handler]})
handler.flush_sync()

Claude Code & Cowork

Add observability and security guardrails to Claude Code sessions:

# Set up hooks (basic observability)
wax observe claude-code setup

# With governance (policy enforcement on Bash/Edit/Write)
wax observe claude-code setup --governance

# With MCP tools (Claude can check policies proactively)
wax observe claude-code setup --governance --mcp

Every session is traced as an agent run — tool calls, LLM usage, and cost are captured automatically. The local guard provides instant security checks:

• Destructive command blocking — rm -rf /, fork bombs, pipe-to-shell
• Sensitive file protection — .env, private keys, credentials
• Git safety — block force push, hard reset, config edits on protected branches
• Path boundary enforcement — block writes outside project directory
• Network access control — block WebFetch to internal/private URLs
• CI/CD protection — require confirmation for Dockerfiles, CI configs, Terraform
• Cowork conflict detection — warn when teammate modified the same file

Customize via .waxell/guard.json:

{
  "git_protected_branches": ["main", "master", "staging"],
  "max_file_changes": 30,
  "path_boundary_enabled": true
}

See the full Claude Code docs for server-side policies, MCP tools, and all configuration options.

Session Tracking

Group related agent runs into sessions:

from waxell_observe import generate_session_id, waxell_agent

session = generate_session_id()  # "sess_" + 16 hex chars

@waxell_agent(agent_name="chat-agent", session_id=session)
async def handle_message(msg: str) -> str:
    ...

Tags and Metadata

Attach searchable metadata to spans:

@waxell_agent(agent_name="my-agent")
async def run_agent(query: str, waxell_ctx=None) -> str:
    waxell_ctx.add_tags(environment="production", version="1.2.0")
    waxell_ctx.add_metadata(user_id="u_123", prompt_template="v3")
    ...

Governance

Policy Enforcement

Policies configured in the Waxell platform are automatically enforced:

from waxell_observe import waxell_agent, PolicyViolationError

@waxell_agent(agent_name="my-agent", enforce_policy=True)
async def run_agent(query: str) -> str:
    ...  # Raises PolicyViolationError if blocked by policy

# Disable policy enforcement for development
@waxell_agent(agent_name="my-agent", enforce_policy=False)
async def run_agent_dev(query: str) -> str:
    ...

Mid-Execution Governance

Enable cooperative policy checks during agent execution:

@waxell_agent(agent_name="my-agent", mid_execution_governance=True)
async def run_agent(query: str, waxell_ctx=None) -> str:
    # Each record_step() triggers a server-side policy check.
    # If a policy blocks, PolicyViolationError is raised.
    waxell_ctx.record_step("step_1", {"tokens_so_far": 5000})
    waxell_ctx.record_step("step_2", {"tokens_so_far": 12000})  # May halt here
    ...

Configuration

Configuration is resolved in priority order:

1. Explicit init() arguments (highest)
2. Environment variables
3. Config file (~/.waxell/config)

init() Parameters

waxell_observe.init(
    # Core
    api_key="wax_sk_...",        # API key (or WAXELL_API_KEY env var)
    api_url="https://...",        # API URL (or WAXELL_API_URL env var)
    debug=False,                  # Enable debug logging + console span export

    # Tracing
    capture_content=False,        # Include prompt/response content in traces
    resource_attributes=None,     # Custom OTel resource attributes on all spans
                                  # e.g. {"deployment.environment": "production"}

    # AI/ML Instrumentation
    instrument=None,              # List of AI/ML libraries (None = auto-detect all)
                                  # e.g. ["openai", "anthropic"]

    # Infrastructure Instrumentation
    instrument_infra=True,        # Enable infra auto-instrumentation
    infra_libraries=None,         # Allowlist (None = auto-detect all installed)
                                  # e.g. ["redis", "httpx", "psycopg2"]
    infra_exclude=None,           # Blocklist — exclude specific libraries
                                  # e.g. ["celery", "grpc"]

    # Prompt Guard
    prompt_guard=False,           # Enable client-side PII/injection detection
    prompt_guard_server=False,    # Enable server-side ML-powered detection
    prompt_guard_action="block",  # "block", "warn", or "redact"
)

Environment Variables

Variable	Description	Default
WAXELL_API_KEY	API key (wax_sk_...)	—
WAXELL_API_URL	Platform API URL	—
WAXELL_OBSERVE	Kill switch — set to false to disable all telemetry	true
WAXELL_INSTRUMENT_INFRA	Enable/disable infrastructure instrumentation	true
WAXELL_INFRA_EXCLUDE	Comma-delimited list of infra libraries to exclude	—
WAXELL_DEBUG	Enable debug logging	false
WAXELL_CAPTURE_CONTENT	Capture prompt/response content	false
WAXELL_PROMPT_GUARD	Enable prompt guard	false

Config File (~/.waxell/config)

INI-format config file with profile support:

[default]
api_url = https://api.waxell.dev
api_key = wax_sk_...
instrument_infra = true
infra_exclude = celery,grpc
debug = false
capture_content = false

[local]
api_url = http://localhost:8001
api_key = wax_sk_...
instrument_infra = true

The config file is created automatically by wax login or can be edited manually.

Kill Switch

Disable all observability with a single environment variable:

export WAXELL_OBSERVE=false  # Disables init(), auto-instrumentation, and span emission

The agent code runs identically — only telemetry emission is suppressed.

Architecture

Your Agent Code
    │
    ├─► waxell-observe SDK (this package)
    │       ├─► Auto-instrumented AI/ML spans (OpenAI, Anthropic, 145+ libraries)
    │       ├─► Auto-instrumented infra spans (HTTP, Redis, PostgreSQL, 30 libraries)
    │       ├─► HTTP REST API (runs, LLM calls, steps, policy checks)
    │       └─► OTel OTLP/HTTP (spans with gen_ai.* semantic conventions)
    │
    └─► Waxell Platform
            ├─► OTel Collector (tenant routing via X-Scope-OrgID)
            ├─► Tempo (distributed traces)
            ├─► Loki (structured logs)
            ├─► PostgreSQL (runs, LLM records, policy audit)
            └─► Grafana (dashboards, trace explorer)

Key design principles:

• Dual data path: OTel spans flow to Tempo for trace visualization; HTTP REST calls persist structured data to PostgreSQL for cost tracking, governance, and analytics.
• Zero latency impact: OTel spans export in a background thread via BatchSpanProcessor. Agent execution sees <0.01ms p99 overhead.
• Fail-safe: If the Waxell backend is unreachable, telemetry is silently dropped. Agent execution continues unimpaired.
• Multi-tenant: Tenant isolation via API key resolution at SDK init. Spans include waxell.tenant_id as an OTel resource attribute for collector routing.

Requirements

• Python 3.10+
• httpx (included in base install)
• OpenTelemetry SDK 1.20+ (optional, via [otel] extra)
• Infrastructure instrumentors (optional, via [infra] extra)

Development

# Install with dev dependencies
pip install -e "observe/waxell-observe[dev,otel,infra]"

# Run SDK tests
pytest observe/waxell-observe/tests/ -v

# Run integration tests (requires Django)
cd controlplane/waxell-controlplane
DJANGO_SETTINGS_MODULE=config.settings pytest ../../tests/integration/ -v

# Run benchmarks
pytest tests/benchmarks/ -v --tb=short

License

Apache 2.0 — see LICENSE for details.