Recoleta
GitHub
Source note

Show HN: Toolnexus for Python – MCP, agent skills,a2a for any LLM

Authors
muthuishere
Source
Hacker News
Published
Collection
Software Intelligence
LLM AgentsMCPAgent ToolsA2aMulti Agent Systems
Open originalMarkdown source

Summary

Toolnexus is a Python toolkit that connects LLM agents to MCP servers, local skills, custom functions, HTTP endpoints, built-in tools, and A2A agents through one tool interface. It targets developers who want a small agent runtime with streaming, retries, memory, metrics, and tool execution across model providers.

Problem

  • LLM agents often need several tool sources at once: MCP servers, local skills, custom code, REST APIs, shell and file tools, and peer agents.
  • Teams need the same agent code to work across OpenAI-style, Anthropic-style, OpenRouter, Gemini schema adapters, and local execution loops.
  • Production use needs memory, streaming, retries, metrics, and failure isolation around remote agents.

Approach

  • create_toolkit(...) loads tools from mcp.json, a skills/ folder, Python functions, HTTP endpoints, 10 built-in tools, and remote A2A agent cards.
  • create_client(...) points at an OpenAI- or Anthropic-style endpoint and runs a loop that sends tool schemas, executes tool calls, and returns a RunResult.
  • ask(prompt, tk, id=...) stores and reloads transcripts through ConversationStore; stream(...) emits text and tool events and saves the thread on completion.
  • A2A support exposes SKILL.md skills through a JSON-RPC 2.0 subset with Agent Card discovery, SendMessage, and GetTask; outbound A2A skills appear as tools.
  • Metrics are exposed through on_metric callbacks and Prometheus text without third-party dependencies.

Results

  • The excerpt reports no benchmark, accuracy, latency, reliability, or user-study results.
  • It claims a working agent can be built in 5 lines when using the default built-in tools.
  • It ships 10 built-in tools: bash, read, write, edit, grep, glob, webfetch, question, apply_patch, and todowrite.
  • The Python package is version 0.4.0, supports Python 3.11 or newer, and publishes a 117.4 kB source archive plus a 53.2 kB wheel.
  • The same library behavior is claimed across 5 language ports: Python, JavaScript, Go, Java, and C#.
  • A2A support is limited in v1: JSON-RPC 2.0, Agent Card discovery, SendMessage, and polling GetTask are included; streaming, push, and auth are absent.

Link