\n<\/p>\n
Multi-platform agentic AI harness. Runs on Telegram, Discord, Microsoft Teams, and the Terminal, with a pluggable backend (Claude Agent SDK, Kilo, OpenCode, Codex, or OpenAI Agents) and full tool access through MCP.<\/p>\n
Multi-frontend
\nTelegram (Grammy + GramJS userbot), Discord (discord.js), Microsoft Teams (Bot Framework), Terminal with live tool visibility<\/p>\n
Pluggable backend
\nClaude Agent SDK, Kilo, OpenCode, Codex, OpenAI Agents \u2014 selectable per-process via backend config. Streaming, model fallback, context-overflow recovery.<\/p>\n
MCP tools
\nMessaging, media, history, search, web fetch, cron jobs, triggers, goals, stickers, file system, admin controls<\/p>\n
Plugins
\nHot-reloadable plugin system. Built-in: GitHub, MemPalace, Playwright, Brave Search<\/p>\n
Background agents
\nHeartbeat (hourly by default \u2014 advances goals, proactively messages when something matters) and Dream (memory consolidation + diary)<\/p>\n
Goals
\nPersistent multi-day objectives the agent commits to in chat; every heartbeat run re-reads them, makes progress, and records what it did<\/p>\n
Skills
\nAgent-authored reusable scripts (bash\/python\/node) \u2014 procedures worked out once get saved and replayed locally at zero token cost<\/p>\n
Triggers
\nSelf-authored watcher scripts (bash\/python\/node) that wake the bot when conditions are met<\/p>\n
Per-chat settings
\nModel, effort level, and pulse toggle per conversation via inline keyboard<\/p>\n
Model registry
\nModels discovered from the active backend at startup \u2014 new models appear in all pickers automatically<\/p>\n
git clone https:\/\/github.com\/dylanneve1\/talon.git && cd talon
\nnpm install<\/p>\n
# Interactive setup (select frontend, configure tokens, pick model)
\nnpx talon setup<\/p>\n
# Start
\nnpx talon start # configured frontend (daemon mode)
\nnpx talon chat # terminal chat mode
\nPrerequisites:<\/p>\n
Node.js 24+
\nBackend-specific:<\/p>\n
claude backend: Claude Code installed and authenticated (claude CLI on PATH).
\nkilo backend: nothing extra \u2014 @kilocode\/sdk spawns a local server. Free models are accessible without auth; routed models use Kilo’s own credentials.
\nopencode backend: nothing extra \u2014 @opencode-ai\/sdk spawns a local server.
\ncodex backend: install the codex CLI (npm i -g @openai\/codex) and authenticate with codex login, CODEX_API_KEY, TALON_CODEX_KEY, or codexApiKey. OPENAI_API_KEY is used only as a fallback when no Codex login exists.<\/p>\n
Talon runs from a normal source or package install; standalone compiled binaries are not supported.<\/p>\n
index.ts Composition root
\n |
\n +– core\/ Platform-agnostic engine
\n | +– agent-runtime\/ Backend capability interfaces, events, stores
\n | +– models\/ Model layer: catalog, per-chat active model,
\n | | reasoning-effort vocabulary
\n | +– prompt\/ System-prompt assembly + prompts\/system templates
\n | +– background\/ Agents that run without a user message:
\n | | heartbeat, dream, pulse, cron, triggers
\n | +– tools\/ MCP tool definitions + spawn\/env contract
\n | +– engine\/ Message flow: dispatcher (per-chat serial,
\n | | cross-chat parallel), HTTP gateway for MCP
\n | | tool calls, backend lifecycle controller
\n | +– plugin.ts Plugin loader, registry, hot-reload
\n |
\n +– backend\/
\n | +– registry.ts Bootstrap-decoupled backend lookup
\n | +– shared\/ Cross-backend helpers (stream state, flow violation,
\n | | delivery contract, metrics, prompt format,
\n | | model retry, system prompt, usage)
\n | +– remote-server\/ Shared infrastructure for agent-server backends
\n | | (MCP registration, sessions, providers, lifecycle)
\n | +– claude-sdk\/ Claude Agent SDK (in-process MCP, hooks)
\n | +– kilo\/ Kilo HTTP server backend (streaming via SSE)
\n | +– opencode\/ OpenCode HTTP server backend
\n | +– codex\/ Codex CLI backend (`@openai\/codex-sdk`)
\n | +– openai-agents\/ OpenAI Agents SDK backend (Responses API)
\n |
\n +– frontend\/
\n | +– shared\/ Cross-frontend presentation helpers
\n | +– telegram\/ Grammy bot + GramJS userbot
\n | +– discord\/ discord.js v14
\n | +– teams\/ Bot Framework + Graph API
\n | +– terminal\/ Readline CLI with tool call visibility
\n |
\n +– storage\/ Sessions, history, chat settings,
\n | cron jobs, media index, daily logs
\n +– util\/ Config, logging, workspace, paths, time<\/p>\n
Dependency rule: core\/ imports nothing from frontend\/ or backend\/. Frontends and backends depend on core types, never on each other. All five backends (Claude SDK, Kilo, OpenCode, Codex, OpenAI Agents) implement the same Backend capability interface from core\/agent-runtime\/capabilities.ts. Kilo and OpenCode additionally share the remote-server\/ infrastructure because they wrap forks of the same upstream HTTP agent server.
\nPrompts: everything the model reads at session start is assembled by core\/prompt\/ from the files in prompts\/ \u2014 see prompts\/README.md for the assembly order, file ownership (user-editable vs package-owned templates), and the per-backend delivery contracts.<\/p>\n
Select via the backend field in ~\/.talon\/config.json. All backends implement the same Backend capability interface \u2014 heartbeat, dream, and chat handlers are backend-agnostic.<\/p>\n
Backend
\nbackend value
\nTransport
\nNotes<\/p>\n
Claude SDK
\n“claude”
\nIn-process via @anthropic-ai\/claude-agent-sdk
\nRequires the claude CLI on PATH. Hook-based turn termination.<\/p>\n
Kilo
\n“kilo”
\nLocal HTTP server via @kilocode\/sdk
\nSSE-streamed turns. Routes to many model providers via Kilo’s auth.<\/p>\n
OpenCode
\n“opencode”
\nLocal HTTP server via @opencode-ai\/sdk
\nSSE-streamed turns; same MCP and session shape as Kilo (upstream fork).<\/p>\n
Codex
\n“codex”
\nPer-turn subprocess via @openai\/codex-sdk
\nRequires the codex CLI from @openai\/codex and Codex auth (codex login, CODEX_API_KEY, TALON_CODEX_KEY, or codexApiKey). MCP servers configured via TOML overrides at thread start.<\/p>\n
OpenAI Agents
\n“openai-agents”
\nIn-process via @openai\/agents
\nResponses API (or any OpenAI-compatible endpoint via TALON_AGENTS_URL \/ openaiBaseUrl). Persistent per-chat MCP bundles.<\/p>\n
The Kilo and OpenCode backends share infrastructure (backend\/remote-server\/) since the upstream HTTP API is the same; each backend supplies its own SDK client, port, and delivery suffix. Codex is its own integration on top of the Codex CLI’s JSONL event stream.<\/p>\n
GitHub API access via the official GitHub MCP server. Gives the agent access to repositories, issues, PRs, code search, and more.
\nRequirements: Docker installed and running.
\n{
\n “github”: {
\n “enabled”: true,
\n “token”: “ghp_…”
\n }
\n}
\nThe token is optional — defaults to the output of gh auth token if the GitHub CLI is authenticated.<\/p>\n
Structured long-term memory with vector search. The agent can store, search, and retrieve memories semantically. Integrates with Dream mode for automatic memory consolidation and personal diary entries.
\nRequirements: Python 3.10+ with the mempalace package.
\n# Set up a Python environment
\npython -m venv ~\/.talon\/mempalace-venv
\n~\/.talon\/mempalace-venv\/bin\/pip install mempalace # Unix
\n# or: ~\/.talon\/mempalace-venv\/Scripts\/pip install mempalace # Windows
\n{
\n “mempalace”: {
\n “enabled”: true,
\n “palacePath”: “~\/.talon\/workspace\/palace”,
\n “pythonPath”: “~\/.talon\/mempalace-venv\/bin\/python”
\n }
\n}
\nBoth paths are optional — defaults to ~\/.talon\/workspace\/palace\/ and the venv Python respectively.<\/p>\n
Headless browser automation via the Playwright MCP server. The agent can browse websites, take screenshots, generate PDFs, fill forms, and scrape content.
\nRequirements: None — @playwright\/mcp is bundled with Talon.
\n{
\n “playwright”: {
\n “enabled”: true,
\n “browser”: “chromium”,
\n “headless”: true
\n }
\n}
\nSupported browsers: chromium (default), chrome, firefox, webkit, msedge.<\/p>\n
Web search via the Brave Search MCP server. Replaces the built-in WebSearch\/WebFetch tools with higher-quality search results.
\n{
\n “braveApiKey”: “BSA…”
\n}
\nGet an API key at brave.com\/search\/api.<\/p>\n
Plugins add MCP tools and gateway actions without modifying core code. SOLID interface — only name is required.
\n{
\n “plugins”: ({ “path”: “\/path\/to\/my-plugin”, “config”: { “apiKey”: “…” } })
\n}
\nexport default {
\n name: “my-plugin”,
\n version: “1.0.0”,
\n mcpServerPath: resolve(import.meta.dirname, “tools.ts”),
\n validateConfig(config) {
\n \/* return errors or undefined *\/
\n },
\n getEnvVars(config) {
\n return { MY_KEY: config.apiKey };
\n },
\n handleAction(body, chatId) {
\n \/* gateway action handler *\/
\n },
\n getSystemPromptAddition(config) {
\n return “## My Plugin\\n…”;
\n },
\n init(config) {
\n \/* one-time setup *\/
\n },
\n destroy() {
\n \/* cleanup *\/
\n },
\n};
\nPlugins support hot-reload via the reload_plugins MCP tool — no restart required.<\/p>\n
talon setup Interactive setup wizard
\ntalon start Start as a background daemon
\ntalon stop Stop the daemon
\ntalon chat Terminal chat mode (always available)
\ntalon status Health, sessions, plugins, disk usage
\ntalon config View or edit configuration
\ntalon logs Tail structured log file
\ntalon doctor Validate environment and dependencies<\/p>\n
Config file: ~\/.talon\/config.json<\/p>\n
Field
\nDefault
\nDescription<\/p>\n
frontend
\n“telegram”
\n“telegram”, “discord”, “teams”, “terminal”, or an array<\/p>\n
backend
\n“claude”
\n“claude”, “kilo”, “opencode”, “codex”, or “openai-agents”<\/p>\n
botToken
\n—
\nTelegram bot token<\/p>\n
model
\n“default”
\nDefault model. Interpretation depends on the active backend.<\/p>\n
codexApiKey
\n—
\nCodex-only OpenAI API key. Prefer this over openaiApiKey for Codex API-key auth. codex login takes precedence over shared openaiApiKey.<\/p>\n
concurrency
\n1
\nMax concurrent AI queries (1–20)<\/p>\n
pulse
\ntrue
\nPeriodic group engagement<\/p>\n
heartbeat
\nfalse
\nBackground maintenance agent<\/p>\n
heartbeatIntervalMinutes
\n60
\nHeartbeat interval<\/p>\n
braveApiKey
\n—
\nBrave Search API key<\/p>\n
timezone
\n—
\nIANA timezone (e.g. “Europe\/London”)<\/p>\n
plugins
\n()
\nExternal plugin packages<\/p>\n
disabledToolTags
\n—
\nHide whole tool groups from the model (e.g. (“stickers”, “web”)) \u2014 each registered tool costs context tokens per session<\/p>\n
disabledTools
\n—
\nHide individual tools by name (end_turn cannot be disabled)<\/p>\n
adminUserId
\n—
\nTelegram user ID for \/admin commands<\/p>\n
allowedUsers
\n—
\nWhitelist of Telegram user IDs<\/p>\n
apiId \/ apiHash
\n—
\nTelegram API credentials for full message history<\/p>\n
github
\n—
\nGitHub plugin config (see above)<\/p>\n
mempalace
\n—
\nMemPalace plugin config (see above)<\/p>\n
playwright
\n—
\nPlaywright plugin config (see above)<\/p>\n
Tool calls shown in real-time with parameters. Streaming phase indicators (thinking \/ responding \/ using tools). Per-turn stats: duration, tokens, cache hit rate, tool count.
\nCommands: \/model, \/effort, \/reset, \/status, \/help<\/p>\n
Docker:<\/p>\n
Systemd: unit file at packaging\/systemd\/talon.service \u2014 copy to \/etc\/systemd\/system\/, set User= and WorkingDirectory=, then systemctl enable –now talon.
\nHealth endpoint: GET http:\/\/localhost:19876\/health returns JSON with uptime, memory, queue depth, active sessions, and last activity timestamp.
\nLogging: Structured JSON via pino to ~\/.talon\/talon.log. Rotated on startup when the file exceeds 10MB.
\nResilience: Dynamic model fallback on overload, session auto-retry on expiry, rate limit handling with backoff, atomic file writes, graceful shutdown with 15-second drain timeout.<\/p>\n
npm run dev # watch mode
\nnpm test # 2300+ tests across unit \/ SDK-stub \/ MCP-functional \/ integration tiers
\nnpm run test:coverage # with coverage report
\nnpm run typecheck # tsc –noEmit
\nnpm run lint # oxlint
\nnpm run format # prettier –write<\/p>\n