Evidor

Evidor is an open-source, provider-agnostic runtime for building AI agents. It provides a clean, modern interface for conversational agents across multiple model providers, with first-class support for stateless LLM calls and native provider-managed conversations.

Install

Install the provider adapter(s) you need:

pip install "evidor[openai]"
# or: evidor[anthropic], evidor[gemini], evidor[mcp], evidor[otel], evidor[langfuse], evidor[phoenix], evidor[prometheus], evidor[all]

Quickstart

from evidor import Agent, OpenAIProvider

agent = Agent(OpenAIProvider(model="gpt-4.1-mini"))
response = agent.send("Give a one-sentence explanation of dependency inversion.")
print(response.text)

Core Architecture: Two Execution Paradigms

Evidor cleanly separates AI execution into two paradigms:

1. Stateless LLM Calls: Direct single-turn or user-managed multi-turn requests with tools, structured outputs, and raw message histories (provider.generate(request)). Evidor does not automatically manage, mutate, or store conversation state.
2. Provider-Managed Conversations: A unified Evidor interface (Agent) over native provider state and context features, exposing provider-specific differences clearly without doing custom client-side token slicing or summarization.

1. Stateless LLM Calls

In the stateless paradigm, you interact directly with the underlying ModelProvider using GenerationRequest. Evidor executes the call directly against modern provider endpoints without maintaining or mutating any hidden conversation state.

Simple Requests

from evidor import GenerationRequest, Message, OpenAIProvider

provider = OpenAIProvider(model="gpt-6-astra")

# Single-prompt shortcut
response = provider.generate(GenerationRequest(prompt="Explain Raft consensus in one paragraph."))
print(response.text)

# Multi-message request
request = GenerationRequest(
    messages=[
        Message(role="system", content="You are a distributed systems expert."),
        Message(role="user", content="What is quorum?"),
    ]
)
response = provider.generate(request)
print(response.text)

User-Managed Message History

When your application handles its own persistence (e.g. database, chat storage, or custom state machine), you maintain the message list and pass it directly to provider.generate(). Evidor never modifies, prunes, or summarises your messages behind your back:

from evidor import GenerationRequest, Message, OpenAIProvider

provider = OpenAIProvider(model="gpt-6-astra")
chat_history: list[Message] = [
    Message(role="system", content="You are a concise coding assistant."),
]

# Turn 1
chat_history.append(Message(role="user", content="Recommend a fast key-value store in Go."))
response = provider.generate(GenerationRequest(messages=chat_history))
chat_history.append(Message(role="assistant", content=response.text))
print(f"Assistant: {response.text}")

# Turn 2
chat_history.append(Message(role="user", content="Does it support persistence?"))
response = provider.generate(GenerationRequest(messages=chat_history))
chat_history.append(Message(role="assistant", content=response.text))
print(f"Assistant: {response.text}")

Stateless Tool Calling

You can supply tool definitions directly in a GenerationRequest. The model emits tool calls if needed, which you can execute statelessly in your own application loop:

from evidor import GenerationRequest, Message, OpenAIProvider, Tool

tool_def = Tool(
    name="get_stock_price",
    description="Get the current price for a stock ticker.",
    parameters={
        "type": "object",
        "properties": {"ticker": {"type": "string"}},
        "required": ["ticker"],
    },
    func=lambda ticker: f"${ticker.upper()} is $150.00",
)

request = GenerationRequest(
    messages=[Message(role="user", content="What is the price of AAPL?")],
    tools=(tool_def,),
)
response = provider.generate(request)

if response.tool_calls:
    for call in response.tool_calls:
        print(f"Model requested tool: {call.name} with args {call.arguments}")

2. Provider-Managed Conversations

Modern LLM providers (OpenAI, Google Gemini, Anthropic) provide native server-side context and session management. Rather than running error-prone client-side token counters, sliding windows, or summarization loops, Evidor delegates context management directly to each provider's native mechanism via Agent.

Multi-turn Conversations

The Agent maintains session state across conversation turns using the underlying provider's native strategy. Use send() (or send_async()) to advance the conversation:

from evidor import Agent, AnthropicProvider

agent = Agent(AnthropicProvider(model="claude-3-7-sonnet-20250219"))

agent.send("My favorite color is navy blue.")
response = agent.send("What color did I say I like?")
print(response.text)

System Prompts

Configure an initial system prompt when creating an Agent. Evidor delivers the instructions natively to the provider backend across turns:

from evidor import Agent, GeminiProvider

agent = Agent(
    GeminiProvider(model="gemini-3.8-flash"),
    system_prompt="You are a concise technical writer. Avoid buzzwords and respond in bullet points.",
)

response = agent.send("How does a TCP handshake work?")
print(response.text)

How Provider Context Management Works

Evidor exposes each provider's context management strategy through agent.strategy:

Inspecting and Resetting Provider State

You can inspect provider identifiers, strategy, or reset the conversation at any time:

agent = Agent(OpenAIProvider(model="gpt-6-astra"))
agent.send("Hello!")

# Inspect provider context strategy
print(agent.strategy)  # "response_chain"

# Inspect the latest provider response/interaction ID
print(agent.last_response_id)  # e.g. "resp_12345"

# Inspect the message history captured during this session
for msg in agent.messages:
    print(f"{msg.role}: {msg.content}")

# Reset the conversation state and provider response chain
agent.reset()  # or agent.clear_history()
print(agent.last_response_id)  # None

Provider Retries & Capabilities

Evidor delegates retry handling directly to the underlying model provider SDKs (openai, google-genai, anthropic). Each provider SDK automatically handles transient network disruptions, connection resets, and rate limits using its native retry policies out of the box, avoiding redundant application-level retry wraps.

Configurable Thinking & Reasoning Mode

Evidor provides first-class support for reasoning models across OpenAI (gpt-6-astra, gpt-6.1-sol, gpt-6-luna), Anthropic (claude-3-7-sonnet-20250219), and Gemini (gemini-3.8-flash, gemini-3.8-live-extended-thinking, gemini-3.1-pro-preview). Thinking can be configured either using provider-specific typed configuration classes or natural shorthands.

from evidor import (
    Agent,
    AnthropicProvider,
    AnthropicThinkingConfig,
    GeminiProvider,
    GeminiThinkingConfig,
    OpenAIProvider,
    OpenAIThinkingConfig,
)

# 1. OpenAI: Configure reasoning effort ('low', 'medium', 'high', 'xhigh', 'max')
# Via shorthand:
provider = OpenAIProvider(model="gpt-6-astra", thinking="high")
# Or via typed config:
provider = OpenAIProvider(model="gpt-6-astra", thinking=OpenAIThinkingConfig(effort="high"))

# 2. Anthropic: Configure token budget or adaptive effort
# Shorthand int (sets budget_tokens >= 1024; auto-bumps max_tokens if needed):
provider = AnthropicProvider(model="claude-3-7-sonnet-20250219", thinking=4096)
# Shorthand string (adaptive thinking):
provider = AnthropicProvider(model="claude-3-7-sonnet-20250219", thinking="high")
# Or via typed config:
provider = AnthropicProvider(
    model="claude-3-7-sonnet-20250219",
    thinking=AnthropicThinkingConfig(budget_tokens=4096),
)

# 3. Gemini: Configure thinking level ('minimal', 'low', 'medium', 'high') or budget
# Shorthand string (Gemini 3.x thinking_level):
provider = GeminiProvider(model="gemini-3.8-flash", thinking="high")
# Shorthand bool (False disables thinking; True enables dynamic thinking):
provider = GeminiProvider(model="gemini-3.8-flash", thinking=False)
# Or via typed config:
provider = GeminiProvider(
    model="gemini-3.8-flash",
    thinking=GeminiThinkingConfig(thinking_level="high", include_thoughts=True),
)

Inspecting Thinking Steps & Reasoning Tokens

When thinking is enabled and the provider outputs thought tokens, the response provides both the visible text and the internal reasoning scratchpad:

agent = Agent(AnthropicProvider(model="claude-3-7-sonnet-20250219", thinking=2048))
response = agent.send("Solve this logic riddle step-by-step...")

# Final visible response
print(response.text)

# Model's internal reasoning scratchpad (if available)
if response.thinking:
    print(f"Reasoning Steps:\n{response.thinking}")

# Reasoning tokens recorded in usage metadata
if response.usage and response.usage.reasoning_tokens:
    print(f"Reasoning tokens consumed: {response.usage.reasoning_tokens}")

Multi-Turn Tool Calling & Cryptographic Signatures

Reasoning models from Anthropic and Gemini output thought blocks alongside cryptographic signatures (such as Anthropic's signature field or Gemini's thought parts). In multi-turn tool loops, these signatures must be preserved and replayed in subsequent conversation turns to avoid API verification errors. Evidor automatically captures and replays them in agent.messages and thought_blocks without requiring any manual handling.

Tools & Function Calling

Evidor provides first-class primitives for defining tools and equipping agents with automatic function-calling execution loops.

Defining Tools with @tool

Transform any Python function into a provider-neutral tool using the @tool decorator. Evidor automatically derives the JSON Schema for parameters from type annotations and extracts descriptions from docstrings (Google, Sphinx, or plain styles):

from evidor import tool

@tool
def get_weather(location: str, unit: str = "celsius") -> str:
    """Get the current weather forecast for a given location.

    Args:
        location: City and country or state, e.g. 'San Francisco, CA'.
        unit: Temperature scale ('celsius' or 'fahrenheit').
    """
    return f"Weather in {location}: 22° {unit}, clear skies."

@tool(name="calc_add", description="Add two numbers together.")
def add(a: float, b: float) -> float:
    return a + b

Equipping Agents with Tools

Pass your tools directly into Agent(..., tools=[...]). When a model decides to call one or more tools, the agent automatically executes them, feeds the results back to the model as role="tool" messages, and loops until the model generates a final response:

from evidor import Agent, GeminiProvider, tool

@tool
def search_database(query: str) -> list[str]:
    """Search internal records by keyword."""
    return [f"Record 1 matching '{query}'", f"Record 2 matching '{query}'"]

agent = Agent(
    GeminiProvider(model="gemini-3.8-flash"),
    tools=[search_database],
    max_tool_iterations=10,  # Maximum tool execution turns per send() (default: 10)
)

response = agent.send("Can you check our records for 'alpha project'?")
print(response.text)

Async Tools & Execution Timeouts

Evidor seamlessly supports asynchronous functions and execution timeouts:

import asyncio
from evidor import Agent, GeminiProvider, tool

# Async functions are automatically executed without blocking or crashing
@tool(timeout=5.0)  # Timeout in seconds
async def fetch_webpage(url: str) -> str:
    """Fetch content from a URL."""
    await asyncio.sleep(0.1)
    return f"Contents of {url}"

# Configure an agent with a default timeout for all its tools
agent = Agent(
    GeminiProvider(model="gemini-3.8-flash"),
    tools=[fetch_webpage],
    tool_timeout=10.0,      # Default tool timeout in seconds
    max_tool_iterations=10, # Max tool turns before forcing final synthesis
)

If a tool times out, raises an error, or if the model emits malformed JSON arguments, Evidor safely captures the error and feeds it back to the model as a role="tool" message so the LLM can self-correct.

Native Asynchronous Conversations (send_async)

In async environments (like FastAPI, Tornado, or asynchronous data pipelines), use await agent.send_async(...). When called, all async tools execute natively on the caller's event loop without thread transitions, cleanly preserving shared connection pools, locks, and contextvars:

import asyncio
from evidor import Agent, OpenAIProvider, tool

@tool
async def lookup_account(account_id: str) -> dict:
    return {"id": account_id, "status": "active"}

async def main() -> None:
    agent = Agent(OpenAIProvider(model="gpt-4.1-mini"), tools=[lookup_account])
    response = await agent.send_async("Check status for account 1234")
    print(response.text)

asyncio.run(main())

Manual Tool Definition

For dynamic or programmatic tools without a standard Python function signature, instantiate Tool directly:

from evidor import Tool

custom_tool = Tool(
    name="query_sql",
    description="Run a read-only SQL query.",
    parameters={
        "type": "object",
        "properties": {
            "query": {"type": "string", "description": "SQL statement"}
        },
        "required": ["query"],
    },
    func=lambda query: f"Results for: {query}",
    timeout=5.0,
)

Evidor's tool calling works consistently across OpenAI, Anthropic, and Gemini, adapting each provider's underlying tool schema and message format transparently.

Built-in Tools

Evidor includes a couple of dependency-free tools you can opt into. They are not attached to agents automatically:

from evidor import Agent, OpenAIProvider, calculator, get_current_time

agent = Agent(
    OpenAIProvider(model="gpt-4.1-mini"),
    tools=[calculator, get_current_time],
)

get_current_time returns an ISO 8601 timestamp and accepts an IANA timezone name (UTC by default). calculator evaluates basic arithmetic without executing Python code.

Create filesystem tools scoped to an existing directory for listing, searching, reading, creating, updating, and deleting files. Paths supplied by the model are resolved under that root, and file reads and writes and result counts have configurable limits:

from evidor import Agent, OpenAIProvider, calculator, filesystem_tools, get_current_time

agent = Agent(
    OpenAIProvider(model="gpt-4.1-mini"),
    tools=[
        calculator,
        get_current_time,
        *filesystem_tools("./project", max_file_bytes=100_000, max_results=100),
    ],
)

filesystem_tools returns six tools:

• list_files lists files in the root or a subdirectory.
• read_file reads a UTF-8 text file.
• search_files searches UTF-8 text files for a literal string.
• create_file creates a new UTF-8 text file; it fails if the path already exists. Its parent directory must already exist.
• write_file replaces the contents of an existing UTF-8 text file.
• delete_file deletes an existing file (directories cannot be deleted).

All paths stay within the configured root. max_file_bytes limits the size of a file read or written (including initial content passed to create_file); max_results bounds listing and search results, and max_files_scanned bounds recursive search work.

Web Search

Use web_search() to create a provider-neutral agent tool. The tool depends only on the WebSearchProvider interface, so the selected search service can be changed without changing the agent setup. The bundled Tavily, Exa, and Brave adapters use only Python's standard library; set the provider's API key environment variable (or pass api_key= explicitly).

export TAVILY_API_KEY="..."
# or EXA_API_KEY="..." / BRAVE_SEARCH_API_KEY="..."

from evidor import Agent, OpenAIProvider, TavilySearchProvider, web_search

search = web_search(TavilySearchProvider(), max_results=5)
agent = Agent(OpenAIProvider(model="gpt-4.1-mini"), tools=[search])

response = agent.send("Find the current Python release notes and cite the sources.")
print(response.text)

Switch providers by replacing TavilySearchProvider() with ExaSearchProvider() or BraveSearchProvider(). To use another service, implement the small WebSearchProvider contract: search(query: str, *, max_results: int) -> list[SearchResult]; both types are exported from evidor.

Webpage Reader

Use read_webpage (or read_webpage_tool()) to let agents fetch and read clean, token-bounded text from web pages, stripping HTML boilerplate automatically:

from evidor import Agent, OpenAIProvider, read_webpage, read_webpage_tool

# Ready-to-use default tool (cleans HTML boilerplate, bounded to 15,000 chars)
agent = Agent(OpenAIProvider(model="gpt-4.1-mini"), tools=[read_webpage])

# Or customized with explicit character and timeout limits
custom_reader = read_webpage_tool(max_chars=8_000, timeout=10.0)

Structured Evidence Store (EvidenceStore)

Prevent context pollution and hallucinations across multi-step research or analysis tasks by equipping agents with an in-memory structured evidence store:

from evidor import Agent, EvidenceStore, GeminiProvider

store = EvidenceStore()

# store.tools provides record_evidence, list_evidence, and compare_evidence
agent = Agent(GeminiProvider(model="gemini-3.8-flash"), tools=[*store.tools])
agent.send("Analyze trade-offs between PostgreSQL and SQLite and record your findings.")

# Inspect structured findings in Python or format as Markdown
print(store.to_markdown())
for item in store.items:
    print(f"[{item.id}] {item.topic}: {item.finding}")

Model Context Protocol (MCP)

Evidor includes first-class support for the Model Context Protocol (MCP). Agents can connect to any MCP server (over local stdio, modern Streamable HTTP, or legacy SSE), automatically discover tools, execute them, and feed responses back into the model.

Install MCP support:

pip install "evidor[mcp]"

Quickstart: Connecting an Agent to an MCP Server

You can pass mcp_servers directly when instantiating an Agent. For example, integrating an Amazon MCP server (e.g. AWS Bedrock, Amazon search, or a custom store MCP server):

from evidor import Agent, OpenAIProvider

agent = Agent(
    OpenAIProvider(model="gpt-4.1-mini"),
    mcp_servers={
        "amazon": {
            "command": "uvx",
            "args": ["amazon-mcp-server"],
            "env": {"AWS_PROFILE": "default"},
        }
    },
)

# Tools exposed by the Amazon MCP server are automatically discovered and made available
response = agent.send("Search for Kindle Paperwhite on Amazon and give me a summary.")
print(response.text)

# Cleanly close subprocesses when done
agent.close()

Or using a context manager:

with Agent(provider, mcp_servers={"amazon": {"command": "uvx", "args": ["amazon-mcp-server"]}}) as agent:
    response = agent.send("Find top noise-cancelling headphones on Amazon.")
    print(response.text)

Async MCP with send_async

In asynchronous applications, use send_async() and async with:

import asyncio
from evidor import Agent, AnthropicProvider

async def main():
    async with Agent(
        AnthropicProvider(model="claude-3-7-sonnet-20250219"),
        mcp_servers={
            "amazon": {
                "command": "npx",
                "args": ["-y", "@modelcontextprotocol/server-amazon"],
            }
        },
    ) as agent:
        response = await agent.send_async("Check the status of order ORD-12345.")
        print(response.text)

asyncio.run(main())

Remote & Hosted MCP Servers (Streamable HTTP & SSE)

Evidor natively supports connecting to remote, hosted MCP servers. In modern MCP specifications, Streamable HTTP (streamable-http or http) is the primary standard for HTTP-based communication, while legacy SSE is supported for backwards compatibility:

from evidor import Agent, OpenAIProvider, MCPServerConfig

# Modern Streamable HTTP (recommended for hosted MCP servers)
agent = Agent(
    OpenAIProvider("gpt-4.1-mini"),
    mcp_servers={
        "hosted-aws": {
            "url": "https://mcp.my-hosted-aws.com/mcp",
            "transport": "streamable-http",  # default for non-/sse URLs
            "headers": {"Authorization": "Bearer YOUR_API_KEY"},
        }
    },
)

# Or explicitly using factory methods:
config = MCPServerConfig.http(
    name="hosted-service",
    url="https://mcp.service.internal/mcp",
    headers={"x-api-key": "..."},
)

Connecting Multiple MCP Servers

Connect multiple MCP servers simultaneously. Evidor automatically prevents tool naming collisions across servers:

from evidor import Agent, GeminiProvider, MCPServerConfig

agent = Agent(
    GeminiProvider(model="gemini-3.8-flash"),
    mcp_servers=[
        # Stdio server
        MCPServerConfig.stdio(
            name="amazon",
            command="uvx",
            args=["amazon-mcp-server"],
        ),
        # Remote Streamable HTTP server
        MCPServerConfig.http(
            name="remote-inventory",
            url="https://inventory.internal/mcp",
            headers={"Authorization": "Bearer secret-token"},
        ),
        # Legacy SSE server
        MCPServerConfig.sse(
            name="weather",
            url="http://localhost:8080/sse",
        ),
    ],
)

response = agent.send("What is the weather in Seattle and recommend waterproof hiking boots on Amazon?")
print(response.text)

mcp_servers vs mcp_client: When to Use Which

An Agent supports two parameters for MCP integration depending on whether the agent manages the connection lifecycle:

Using mcp_client to Share Connections Across Agents

from evidor import Agent, MCPClient, OpenAIProvider

# Boot a single shared client for your application
shared_mcp = MCPClient.from_file("claude_desktop_config.json")
shared_mcp.connect()

# Both agents reuse the exact same client without duplicating subprocesses:
researcher = Agent(OpenAIProvider("gpt-4.1-mini"), mcp_client=shared_mcp)
assistant = Agent(OpenAIProvider("gpt-4.1-mini"), mcp_client=shared_mcp)

response1 = researcher.send("Search records for client ABC.")
response2 = assistant.send("Verify credentials for client ABC.")

# Clean up once when the application terminates
shared_mcp.close()

Advanced: Standalone MCPClient and Tool Composition

You can also use MCPClient standalone to inspect tools, retrieve resources, or mix MCP tools with standard Python tools:

import asyncio
from evidor import Agent, MCPClient, OpenAIProvider, tool

@tool
def internal_calculator(x: int, y: int) -> int:
    return x * y

async def main():
    async with MCPClient({"amazon": {"command": "uvx", "args": ["amazon-mcp-server"]}}) as mcp:
        # Retrieve all Evidor Tool objects from the MCP server
        mcp_tools = await mcp.get_tools_async()

        # Combine MCP tools with local tools
        agent = Agent(OpenAIProvider("gpt-4.1-mini"), tools=[*mcp_tools, internal_calculator])
        response = await agent.send_async("Calculate total price for 3 units of item B08N5WRWNW.")
        print(response.text)

asyncio.run(main())

Telemetry & Observability

Evidor includes a first-class, decoupled telemetry subsystem built on an asynchronous actor-like producer-consumer runtime. It captures distributed trace trees across agent turns, model generations, tool executions, and external MCP server invocations with zero overhead on the agent's critical path.

Zero Extra Dependencies for Custom Telemetry:
Custom telemetry implementations and built-in sinks (ConsoleSink, InMemorySink) require no external packages (pip install evidor).
Optional adapters are available for industry-standard backends:

• OpenTelemetry: pip install "evidor[otel]"
• Langfuse: pip install "evidor[langfuse]"
• Arize Phoenix: pip install "evidor[phoenix]"
• Prometheus: pip install "evidor[prometheus]"

Design Principles

• Zero-Dependency Core: Evidor core defines framework-neutral domain events and lifecycle operations. Telemetry adapters own protocol conversions and client exports.
• Non-Blocking Emission: Event emission on the agent thread takes under 1 microsecond via an in-memory bounded queue. Formatting, batching, and network transport happen on a dedicated background worker thread (Evidor-Telemetry-Worker).
• Trace Context Propagation: Automatically tracks parent-child span hierarchy across synchronous and asynchronous operations using contextvars.
• Fault-Isolated: Downstream observability backend timeouts, outages, or serialization errors never crash the agent.

Built-in Sinks (Zero Dependencies)

Evidor includes built-in sinks that require zero external dependencies:

1. ConsoleSink: Real-Time Developer Visibility

Stream structured events directly to sys.stderr or sys.stdout:

from evidor import Agent, OpenAIProvider
from evidor.telemetry import ConsoleSink

sink = ConsoleSink()
agent = Agent(OpenAIProvider("gpt-4.1-mini"), telemetry=sink)
agent.send("What is the speed of light?")

2. InMemorySink: Testing & Offline Assertions

Capture all events in memory to assert on token counts, latency, and tool invocations:

from evidor import Agent, OpenAIProvider
from evidor.telemetry import InMemorySink, LLMCallEndEvent

sink = InMemorySink()
agent = Agent(OpenAIProvider("gpt-4.1-mini"), telemetry=sink)
agent.send("Hello agent!")

# Inspect recorded events
llm_events = sink.filter(LLMCallEndEvent)
print(f"Total tokens used: {llm_events[0].token_usage.total_tokens}")

Observability Adapters

Evidor provides official, ultra-compact adapters for industry-leading observability platforms. Because Evidor's core runtime manages all context propagation, queuing, and background dispatch, each adapter requires under ~100 lines of logic:

1. OpenTelemetry & OpenInference (evidor[otel])

Convert domain events into distributed trace spans compliant with OpenTelemetry GenAI and OpenInference semantic conventions:

pip install "evidor[otel]"

from opentelemetry import trace
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import BatchSpanProcessor, ConsoleSpanExporter
from evidor import Agent, OpenAIProvider, OpenTelemetrySink

provider = TracerProvider()
provider.add_span_processor(BatchSpanProcessor(ConsoleSpanExporter()))
trace.set_tracer_provider(provider)

otel_sink = OpenTelemetrySink(tracer_provider=provider)
agent = Agent(OpenAIProvider("gpt-4.1-mini"), telemetry=otel_sink)
response = agent.send("Search the documentation for vector embeddings.")
agent.close()

2. Langfuse (evidor[langfuse])

Trace agent executions, evaluate responses, monitor token usage, and track latency on the open-source Langfuse platform:

pip install "evidor[langfuse]"

from evidor import Agent, OpenAIProvider, LangfuseSink

# Reads LANGFUSE_PUBLIC_KEY, LANGFUSE_SECRET_KEY, and LANGFUSE_HOST from env
langfuse_sink = LangfuseSink()

agent = Agent(OpenAIProvider("gpt-4.1-mini"), telemetry=langfuse_sink)
response = agent.send("Summarize the quarterly earnings report.")
agent.close()  # Automatically flushes pending events and shuts down Langfuse client

3. Arize Phoenix (evidor[phoenix])

Send traces directly to Arize Phoenix (local or cloud) formatted with native OpenInference metadata (input.value, output.value, tool.parameters, token breakdowns):

pip install "evidor[phoenix]"

from evidor import Agent, OpenAIProvider, PhoenixSink

# Connects to Phoenix at http://localhost:6006/v1/traces by default
phoenix_sink = PhoenixSink(
    endpoint="http://localhost:6006/v1/traces",
    project_name="my-evidor-agent",
)

agent = Agent(OpenAIProvider("gpt-4.1-mini"), telemetry=phoenix_sink)
response = agent.send("Analyze customer sentiment from survey data.")
agent.close()

4. Prometheus Metrics (evidor[prometheus])

Export operational agent metrics (evidor_agent_runs_total, evidor_llm_calls_total, evidor_llm_retries_total, evidor_tokens_total, evidor_tool_calls_total, evidor_errors_total, and latency histograms) to Prometheus:

pip install "evidor[prometheus]"

from evidor import Agent, OpenAIProvider, PrometheusSink

prom_sink = PrometheusSink()
agent = Agent(OpenAIProvider("gpt-4.1-mini"), telemetry=prom_sink)
agent.send("Calculate total revenue.")

# Export metrics exposition text (e.g., in a FastAPI /metrics endpoint):
metrics_text = prom_sink.export_text()
print(metrics_text)

Span Tree & Semantic Attributes

Evidor automatically builds the full execution tree across agent runs, LLM generations, and tool calls:

agent.run                     [openinference.span.kind = "AGENT"]
  ├── llm.gpt-4.1-mini        [openinference.span.kind = "LLM", status = "ok"]
  └── tool.search_docs        [openinference.span.kind = "TOOL"]
        └── mcp.docs_server.search  [openinference.span.kind = "TOOL"]

Standard attributes automatically populated:

• openinference.span.kind: AGENT, LLM, or TOOL
• input.value, output.value, tool.parameters
• gen_ai.system, gen_ai.request.model, gen_ai.response.model
• gen_ai.usage.input_tokens, gen_ai.usage.output_tokens, gen_ai.usage.total_tokens
• llm.input_prompt, llm.input_messages, llm.output_text, llm.output_tool_calls
• evidor.agent.run_id, evidor.tool.name, evidor.mcp.server_name

Thinking Visibility in Telemetry (capture_thinking)

Internal thinking scratchpads can be lengthy and often contain confidential reasoning traces. All Evidor telemetry sinks default to capture_thinking=False, which redacts the raw thinking text while preserving reasoning token counts (gen_ai.usage.reasoning_tokens and Prometheus evidor_tokens_total{token_type="reasoning"}).

To make intermediate thinking steps visible in your traces, logs, or UI dashboards, enable capture_thinking=True:

from evidor import Agent, AnthropicProvider
from evidor.telemetry import ConsoleSink, OpenTelemetrySink, PhoenixSink, LangfuseSink

# Enable thinking visibility on any sink
console_sink = ConsoleSink(capture_content=True, capture_thinking=True)
otel_sink = OpenTelemetrySink(capture_content=True, capture_thinking=True)
phoenix_sink = PhoenixSink(capture_content=True, capture_thinking=True)
langfuse_sink = LangfuseSink(capture_content=True, capture_thinking=True)

agent = Agent(
    AnthropicProvider(model="claude-3-7-sonnet-20250219", thinking=2048),
    telemetry=[console_sink, phoenix_sink],
)
response = agent.send("Analyze system security risks...")
agent.close()

Supported Telemetry Thinking Attributes & Semantics:

• OpenTelemetry (evidor[otel]): Exports gen_ai.reasoning.content and evidor.thinking.content span attributes, plus gen_ai.usage.reasoning_tokens.
• Arize Phoenix (evidor[phoenix]): Automatically outputs OpenInference reasoning blocks (message_content.type = "reasoning") rendered directly in the Phoenix UI, and logs llm.token_count.completion_details.reasoning.
• Langfuse (evidor[langfuse]): Attaches reasoning text to generation traces and populates usage["reasoning"].
• Prometheus (evidor[prometheus]): Tracks token consumption under evidor_tokens_total{token_type="reasoning"}.

Privacy & PII Protection (Metadata-Only Mode)

In privacy-sensitive or regulated environments (such as healthcare, finance, or customer data processing), you may not want to persist raw user prompts, model outputs, or tool arguments to external observability backends, while still requiring distributed trace hierarchies, token usage, latency timings, and error analytics.

All Evidor telemetry sinks accept capture_content: bool = True and capture_thinking: bool = False parameters. Setting capture_content=False strips all raw text payloads before export:

from evidor import Agent, OpenAIProvider
from evidor.telemetry import OpenTelemetrySink, LangfuseSink, PhoenixSink

# Sinks record span trees, token usage, and latencies, but redact all prompts, thinking, and outputs
otel_sink = OpenTelemetrySink(capture_content=False)
langfuse_sink = LangfuseSink(capture_content=False)
phoenix_sink = PhoenixSink(capture_content=False)

agent = Agent(
    OpenAIProvider("gpt-4.1-mini"),
    telemetry=[otel_sink, langfuse_sink],
)
response = agent.send("Sensitive patient record or financial data...")
agent.close()

For custom sinks, the pure function redact_event(event, capture_content=..., capture_thinking=...) is exported directly from evidor.telemetry.

Advanced: Custom Sinks & Runtime Tuning

Implementing a Custom Sink

Implement the TelemetrySink protocol to send events to any custom database, webhook, or logging system:

from collections.abc import Sequence
from evidor.telemetry import TelemetryEvent, TelemetrySink

class WebhookSink(TelemetrySink):
    def write(self, events: Sequence[TelemetryEvent]) -> None:
        for event in events:
            # Send batch payload over HTTP or write to database
            attrs = event.to_attributes()
            ...

    def flush(self) -> None:
        pass

    def close(self) -> None:
        pass

Tuning the TelemetryRuntime

Configure queue capacity, batching thresholds, and overflow strategies:

from evidor import Agent, OpenAIProvider
from evidor.telemetry import TelemetryRuntime, ConsoleSink

# Actor runtime with custom batching and bounded buffer
runtime = TelemetryRuntime(
    sinks=[ConsoleSink()],
    queue_size=2000,                # Max buffered events
    batch_size=100,                 # Batch size dispatched to sinks
    flush_interval_seconds=0.5,     # Max wait before flushing available events
    overflow_strategy="drop_oldest" # "drop_newest" (default) or "drop_oldest"
)

agent = Agent(OpenAIProvider("gpt-4.1-mini"), telemetry=runtime)

Providers

Supported providers, example models, and reasoning/thinking support:

API keys can be supplied explicitly via the api_key constructor argument or read automatically from environment variables.

Modern Provider APIs & Zero Legacy Fallbacks

Evidor is in active beta and embraces breaking changes to maintain clean, modern integrations:

• Google Gemini (GeminiProvider): Exclusively uses Google GenAI's modern Interactions API (client.interactions.create) from the official google-genai SDK (from google import genai). All legacy generate_content / generate_content_stream fallbacks and outdated candidate-parsing logic have been removed. Requires google-genai>=1.0.0.
• OpenAI (OpenAIProvider): Exclusively uses OpenAI's modern Responses API (client.responses.create) with full support for previous_response_id chaining, reasoning models (gpt-6-astra, gpt-6.1-sol, gpt-6-luna), and modern tool schemas. Legacy Chat Completions methods have been removed. Requires openai>=1.0.0.
• Anthropic (AnthropicProvider): Modern Messages API with native ephemeral prompt caching (cache_control={"type": "ephemeral"}) and adaptive/budget-based extended thinking.

Custom Providers

Agent depends only on the ModelProvider protocol. To support another LLM or backend, implement the generate(request) method:

from evidor import GenerationRequest, GenerationResponse, ModelProvider

class CustomProvider:
    def __init__(self, model: str = "custom-model") -> None:
        self.model = model

    def with_model(self, model: str) -> "CustomProvider":
        return CustomProvider(model=model)

    def generate(self, request: GenerationRequest) -> GenerationResponse:
        prompt_text = request.prompt
        # or iterate over request.messages
        return GenerationResponse(text="Custom response", model=self.model)

Optional Extras Summary

Build and release

Create distributable artifacts locally:

python -m pip install -e ".[dev]"
python -m pytest
python -m build
python -m twine check dist/*

Releases are performed by GitHub Actions and use Conventional Commits to select the next version:

• fix: ... creates a patch release.
• feat: ... creates a minor release.
• feat!: ... or a BREAKING CHANGE: footer creates a major release.

Pushes to dev publish -dev.N prereleases to TestPyPI. Pushes to main publish stable releases to PyPI. Configure TEST_PYPI_API_TOKEN as a repository or dev environment secret, and configure PyPI trusted publishing for the prod environment before enabling the workflow.