Talon: self-hosted AI agent harness for Telegram, Discord, Teams, Terminal

Talon is a multi-platform harness for agents that run continuously in the background and respond to users across Telegram, Discord, Microsoft Teams, a terminal CLI, or a Flutter-based desktop and mobile app. The tool treats the interface (frontend), the AI backend, and the agent's capabilities as separate, pluggable layers.

On the frontend side, users can deploy the same agent across any combination of Telegram (using Grammy and GramJS), Discord (discord.js v14), Microsoft Teams (Bot Framework and Graph API), a Terminal with live tool visibility, or a local/remote Desktop and Mobile client built with Flutter. The Desktop/Mobile app communicates via a versioned HTTP and Server-Sent-Events bridge (the Talon Client Bridge Protocol), which allows remote agents over LAN or the internet with optional HTTPS and token authentication.

For the AI backend, developers select one of five implementations in the config file: Claude Agent SDK (in-process via the claude CLI), Kilo (local HTTP server without extra auth for free models), OpenCode (local HTTP server, same MCP shape as Kilo), Codex (per-turn subprocess requiring the Codex CLI and auth), or OpenAI Agents (in-process via Responses API or any OpenAI-compatible endpoint). All backends share the same capability interface, so chat handlers, heartbeat, and dream logic are backend-agnostic; the system also supports streaming, model fallback, and context-overflow recovery.

The agent's execution model centers on tasks—every chat turn, heartbeat run, dream cycle, and cron/trigger job is registered in a live task table accessible via talon ps or talon kill. Goals (persistent multi-day objectives committed in chat) are re-read and advanced on each heartbeat run, with progress recorded. Background agents include a Heartbeat (hourly by default, advancing goals and proactively messaging when something matters) and Dream (memory consolidation and diary writing). Triggers are self-authored watcher scripts (bash, Python, or Node.js) that wake the bot when conditions are met. An event bus provides typed pub-sub across the task and turn lifecycle, allowing subsystems to subscribe instead of importing each other.

Tools are exposed through MCP (the Model Context Protocol) and include messaging, media, history search, web fetch, cron jobs, triggers, goals, file system, and admin controls. The VFS (virtual filesystem) unifies the namespace at ~/.talon/ns with a FUSE-backed live view of the task table, event bus, and plugin registry, allowing tools and users to interact with the agent via plain ls and cat commands.

Plugins (hot-reloadable modules for additional functionality) and Skills (reusable SKILL.md workflow bundles) are managed via CLI. Built-in plugins include GitHub (via Docker-backed GitHub MCP server), MemPalace (local vector search and knowledge graph for semantic memory), and Playwright (browser automation). Long-term memory can use MemPalace (local, Python 3.10+ required) or mem0 (hosted or self-hosted).

Installation is straightforward: npm install followed by npx talon setup (interactive configuration) and npx talon start. The tool also ships self-contained binaries for Linux and macOS (x64 and arm64) and Windows (x64), available via Homebrew on macOS/Linux or direct download, requiring only SHA256 verification. The code is open-source on GitHub (github.com/dylanneve1/talon), with a clear architecture separating core (platform-agnostic engine), backend (Claude, Kilo, OpenCode, Codex, OpenAI), and frontend (Telegram, Discord, Teams, Terminal, Desktop) layers.

Talon addresses a growing need for AI agent infrastructure that does not lock developers into a single platform or model provider. By decoupling the interface layer (where users interact) from the backend layer (which AI powers the agent) and the tooling layer (MCP for external integrations), it allows teams to swap any component without rewriting the agent logic. The multi-day goals system, background tasks (heartbeat and dream), and persistent task table show that the tool is designed for agents that do more than respond to a single message—they maintain state, pursue objectives, and consolidate learning over time.

The plugin and skills ecosystems—with hot-reload for plugins and per-session skills installation—suggest that authors expect different use cases to require different capabilities. The inclusion of long-term memory backends (MemPalace and mem0) and memory consolidation in the Dream background agent indicates that persistence and semantic recall are treated as first-class features, not afterthoughts.

The availability of both npm and binary distributions broadens access: developers can use npm for a lightweight setup, while organizations that want a zero-dependency deployment path can use the Homebrew or direct-download binaries.

No comments yet. Be the first to share your thoughts!