Everyone’s raising their AI lobster. Not everyone has a server to run it on.

# ai# go

邱敬幃 Pardn Chiu

Agenvoy is a pure Go agentic backend — no framework overhead, no bloated runtime. Verified...

Agenvoy is a pure Go agentic backend — no framework overhead, no bloated runtime. Verified running on a Raspberry Pi Zero 2W with just 512MB RAM.

Multi-Provider LLM with Intelligent Routing

Agenvoy integrates seven AI backends — GitHub Copilot, Claude, OpenAI, Gemini, Nvidia NIM, and any OpenAI-compatible endpoint (Compat/Ollama) — behind a unified Agent interface. A dedicated planner LLM automatically selects the most appropriate provider for each request, eliminating the need to manually switch models. Named compat[{name}] instances allow multiple local model endpoints to coexist, each with independent URL and credential configuration.

Skill-Based Agentic Execution

Skills are declarative Markdown files (SKILL.md) that define a task's system prompt and tool allowlist. At runtime, a Selector LLM picks the best matching skill across 9 standard scan paths, then drives a tool-call loop of up to 128 iterations until the task completes. When the iteration limit is reached, the engine automatically triggers summarization rather than returning an error. Skill extensions are auto-synced from GitHub on startup via SyncSkills.

25+ Built-in Tools Across Six Categories

The executor ships a comprehensive toolchain: filesystem operations (read_file, write_file, patch_edit, glob_files, search_content), web access (search_web, fetch_page, download_page, fetch_google_rss), scheduling (add_task, add_cron, write_script), error memory (remember_error, search_errors, get_tool_error), a math calculator, and arbitrary HTTP requests. Every rm is redirected to .Trash and all writes use atomic tmp-then-rename to prevent partial file corruption.

JSON-Driven API Extension Architecture

External REST APIs are defined as JSON files under extensions/apis/ and loaded at runtime through the API adapter — no Go code required. Thirteen public-domain APIs are bundled out of the box: Yahoo Finance, CoinGecko, Wikipedia, World Bank, USGS Earthquake, Nominatim, Open-Meteo, HackerNews, REST Countries, TheMealDB, IP-API, and exchange rates. Custom endpoints follow the same schema, making the tool surface arbitrarily extensible without recompilation.

Discord Bot Mode with Task Scheduler

cmd/server launches a persistent Discord bot handling both direct messages and slash commands with per-channel session state. The integrated scheduler supports one-time tasks (via +5m or absolute timestamps) and recurring cron jobs (standard 5-field expressions validated by go-scheduler). Each task is linked to a Discord channel ID: when a script completes, the planner agent processes stdout and posts results back to the originating channel. The schedule-task skill routes natural-language timing intent to the scheduler automatically.

Cross-Session Persistent Memory

At the end of each turn, the agent emits a structured JSON summary that is deep-merged with the previous session summary using field-level deduplication, then stored in ~/.config/agenvoy/. Subsequent sessions inject this summary alongside the last N conversation turns, allowing the agent to recall decisions, constraints, and conclusions without replaying full history. Tool-execution errors are persisted with SHA-256 keys so the agent can look up past root causes before retrying.

OS Keychain Credential Management

API keys are stored in the OS-native keychain (macOS Keychain, Linux Secret Service) rather than plain environment variables. GitHub Copilot uses OAuth Device Code Flow with automatic token refresh. The interactive agenvoy add / agenvoy remove commands manage credentials across all six providers with an embedded model registry for guided selection. A multi-layer path deny-list blocks access to SSH keys, shell configs, cloud credentials (.aws, .gcloud, .docker), .env files, and private key formats across all file and command tools.