Talon: A multi-platform AI agent harness with pluggable backends and MCP tool access

Talon is a new open-source AI agent harness that aims to be a universal runtime for interactive AI agents. It supports multiple frontends — Telegram, Discord, Microsoft Teams, terminal, and a cross-platform desktop/mobile companion app built with Flutter — all powered by a pluggable backend system. You can choose from five backends: the Claude Agent SDK, Kilo, OpenCode, Codex, or OpenAI Agents, each implementing the same core interface. This means your agent can run identically across chat platforms, while the underlying model and tooling can be swapped via a config file.

A key architectural decision is strict separation of core, frontend, and backend code. The core module manages prompts, tool registration, and memory, while each backend implements a shared capability interface. This makes it relatively easy to add new backends or frontends without touching the rest of the codebase. Talon also bundles support for the Model Context Protocol (MCP), allowing agents to use a growing ecosystem of MCP servers for browser automation, web search, GitHub API access, and more.

For developers and power users, Talon ships self-contained binaries for Linux, macOS, and Windows — no Node.js runtime required for the basic agent CLI. The desktop companion app communicates with the daemon via a versioned HTTP/SSE protocol (the Talon Client Bridge Protocol), opening the door for third-party GUI clients. The project also includes structured JSON logging, health endpoints, and systemd/docker deployment configurations.

Talon's memory subsystem supports two long-term backends: a local "mempalace" using vector search and knowledge graphs, or the hosted mem0 platform. The agent can consolidate memories automatically in a "Dream" mode, which could be useful for maintaining persistent context across sessions.