Talon

A multi-platform agentic AI harness.

Runs on Telegram, WhatsApp, Discord, Microsoft Teams, the Terminal, and a cross-platform Desktop and Mobile companion app (Flutter), with a pluggable backend (Claude Agent SDK, Kilo, OpenCode, Codex, or OpenAI Agents) and full tool access through MCP.

Features

Multi-frontend

- Telegram (grammY and GramJS userbot), WhatsApp (Baileys multi-device), Discord (discord.js), Microsoft Teams (Bot Framework), Terminal with live tool visibility, and a Desktop and Mobile app (Flutter) over a local or remote bridge. One or several at once.

Pluggable backend

- Claude Agent SDK, Kilo, OpenCode, Codex, and OpenAI Agents. Selectable per process. Streaming, model fallback, and context-overflow recovery.

MCP tools

- Messaging, media, history, search, web fetch, cron jobs, triggers, goals, stickers, file system, and admin controls.

Plugins

- Hot-reloadable plugin system with install, enable, and disable commands from npm, git, or local sources. Built-in: GitHub, MemPalace, Playwright, and Brave Search.

Background agents

- Heartbeat (hourly by default) advances goals and proactively messages when something matters. - Dream consolidates memory and writes diary entries.

Goals and skills

- Persistent multi-day objectives the agent commits to in chat. Every heartbeat re-reads them, makes progress, and records what it did. - SKILL.md workflow bundles the agent authors and reuses, installable from local folders, git, or owner/repo. The Anthropic skills ecosystem installs directly.

Triggers and task table

- Self-authored watcher scripts in bash, python, or node that wake the bot when conditions are met. - Every unit of agent work (chat turns, heartbeat, dream, isolated cron and trigger jobs) is registered live in a task table.

Event bus and VFS

- Typed internal pub-sub spine with task and turn lifecycle events. - Unified namespace at ~/.talon/ns over workspace, skills, scripts, and logs, plus /proc-style live views of the task table, event bus, and plugin registry, backed by FUSE.

Per-chat settings

- Model, effort level, and pulse toggle per conversation.

Model registry

- Models discovered from the active backend at startup. New models appear in all pickers automatically.

Frontends

- Telegram: inline keyboards, reactions, media groups, stickers, polls, admin commands, and optional userbot history access. - WhatsApp: drives a real WhatsApp account over Baileys multi-device, paired by phone code or QR. Media, reactions, edits, deletes, forwards, polls, locations, contacts, and group admin. - Discord: slash commands, guild and channel allowlists, and presence text. - Microsoft Teams: inbound over a Power Automate webhook, outbound over Graph. - Terminal: local readline chat, always available even when another frontend is configured. - Native: HTTP and Server-Sent Events bridge used by the Flutter companion app.

Backends

- Claude SDK: in-process, requires the Claude Code CLI on PATH. - Kilo: local HTTP server, streaming via SSE, routes to many model providers. - OpenCode: local HTTP server, same MCP and session shape as Kilo. - Codex: per-turn subprocess, requires the Codex CLI and Codex authentication. - OpenAI Agents: in-process, uses the Responses API or any OpenAI-compatible endpoint.

Desktop and mobile app

The native frontend turns the daemon into a client bridge, a versioned HTTP and Server-Sent Events JSON API that any GUI client can speak. The reference client is Talon Companion, a single Flutter codebase that runs on Windows, macOS, Linux, and Android. It provides multi-chat history, live streaming with reasoning and tool-call visibility, per-chat model and effort settings, and settings sync with the daemon. The app keeps itself up to date by watching releases and installing the next one.

Requirements

- Bun 1.3+ or Node.js 24+ for running from source - Backend-specific tools depending on the chosen backend - Optional: Docker for the GitHub plugin, Python 3.10+ for the MemPalace memory backend

Note

Plugins that ship their MCP server as TypeScript source need a runtime that can execute TypeScript. A standalone binary does not provide that, so those plugins want the npm install. Plugins declaring an explicit mcpServer command work everywhere.

License

MIT