PRAXO-CORE: Python Agent Framework

PRAXO-CORE is a dependency-free Python framework for building tool-using AI agents. It provides the agent runtime, not a model or hosted service, so you keep control of your models, tools, storage, and sandboxing.

It is for teams that want a small, explicit agent loop instead of a large orchestration platform:

• Async-first models, tools, events, and sessions.
• Policy hooks, approvals, path checks, PII redaction, and call limits.
• Context budgets, checkpoints, middleware, and fallback models.
• Streaming, cancellation, and OpenAI-compatible, Anthropic, and Gemini adapters.
• No runtime dependencies; Python 3.11 or newer.

Registry and Coding-Agent Entry

• Registry introduction: REGISTRY.md
• Machine-readable entry: llms.txt and llms.md
• Examples: runnable language-specific examples.
• Coding-agent contract: docs/getting-started/coding-agents.md
• Registry: PyPI

Install

pip install praxo-core==0.6.0

For a source checkout:

pip install ./language/python

Quickstart

For a complete application with explicit tool scope, policy denial tests, and virtual-environment package checks, use the independent Python consumer.

import asyncio

from praxo.agent import Agent, AgentOptions, Session
from praxo.event import Event
from praxo.model import Message, Response, Usage
from praxo.tool import FuncTool, Registry


class EchoModel:
    async def generate(self, request):
        return Response(message=Message(role="assistant", content="hello"))


class LogSink:
    async def emit(self, event: Event) -> None:
        print(event.to_wire())


async def main() -> None:
    tools = Registry()
    tools.register(FuncTool("echo", "Echo input", lambda args: args))
    agent = Agent(
        AgentOptions(model=EchoModel(), tools=tools, events=LogSink())
    )
    result = await agent.run(Session(id="session-1"), "hi")
    print(result.output, result.steps, result.usage)


asyncio.run(main())

What you get

Harmovela event sink

HarmovelaSink posts lifecycle events to {base_url}/events as JSON envelopes with spec_version, id, type, source, created_at, and a payload that merges the event data with session_id and run_id when present.

from praxo.adapters.harmovela import HarmovelaSink

sink = HarmovelaSink("http://127.0.0.1:8790/aep/api")
agent = Agent(AgentOptions(model=EchoModel(), events=sink))

Pass transport to run without network access; it receives (method, url, headers, body) and returns (status, body). Limits: the sink adds no authentication, retries, buffering, acknowledgements, ordering, or exactly-once delivery. Filter sensitive event data at the caller, and use HTTPS for remote endpoints.

MCP client (stdio)

MCPClient connects to a local MCP server over newline-delimited JSON-RPC 2.0, runs the initialize / notifications/initialized handshake, lists tools with pagination checks, and calls tools through the Core registry seam.

from praxo.adapters.mcp import MCPClient, MCPServerConfig
from praxo.tool import Registry

config = MCPServerConfig(command="my-mcp-server", args=["--stdio"])
async with MCPClient(config) as client:
    tools = await client.list_tools()
    result = await client.call_tool("search", {"query": "praxo"})

    registry = Registry()
    await client.register_tools(registry)

list_tools() follows nextCursor, preserves inputSchema on each MCPTool, and rejects repeated cursors and duplicate names. call_tool() returns structuredContent when present, otherwise a text string or an ordered list of text blocks; isError results raise MCPToolError. Registering tools does not bypass policy: the agent still authorizes each tool.invoke before calling the registered adapter. Limits: stdio only; no hosted service, authentication, sandboxing, schema promotion, resources, prompts, sampling, or server-initiated requests. Tool output is untrusted. Timeouts send notifications/cancelled, but local cancellation does not prove remote work stopped. MCP failures use MCP_* codes such as MCP_PROTOCOL, MCP_TRANSPORT, MCP_HANDSHAKE, MCP_TIMEOUT, MCP_PAGINATION, MCP_DUPLICATE_TOOL, MCP_TOOL_ERROR, MCP_RPC, and MCP_CLOSED; Registry.specs() still carries only names and descriptions, so read MCPTool.input_schema when you need the model-facing schema.

Contract and status

The normative contract is spec/; Go remains the reference implementation. This package supports SPEC_VERSION 0.2.0 and passes all eleven conformance vector groups. Source version 0.7.0 is prepared in the repository; the latest source tag and PyPI package are python/v0.6.0 and praxo-core==0.6.0. VFS and plugin hosts are not yet implemented.

Requirements

• Python 3.11 or newer
• No runtime dependencies

Contract Surface

Wire codecs preserve the Go names, including image_url, runId, contentBlocks, and plugin plugin_id.

Conformance

From this directory:

python conformance/run.py
python -m unittest discover -s tests -v

Both commands read ../../spec/vectors/*.json. A released Core implementation must pass every vector and keep spec/parity.md current.

API Reference and Build

See API.md. Refresh the checked-in symbol snapshot after a public API change:

python api-snapshot.py
python api-snapshot.py --check
python api-signatures.py --check
python -m build --wheel --sdist .

The name snapshot and signature snapshot are checked independently. The signature baseline covers class constructors, dataclass fields/defaults, methods, properties, functions, and exported constants. Both are source-API gates, not runtime or ABI guarantees.

CI installs the built wheel and sdist into separate clean virtual environments and runs tests/installed_smoke.py from outside the source tree. This verifies the packaged artifacts, not the checkout. It does not prove PyPI publication or installation from a package registry.

License

MIT. See LICENSE.