Roadmap

Planned work, loosely prioritized. Nothing here is committed — priorities shift. Open an issue if something matters to you.

Near-term

More channels — WhatsApp (Twilio / Meta Cloud API) and Slack
Voice support — Real-time API streaming for voice conversations through Telegram and Discord
Profile-based tool loadouts — Named tool profiles (coding, research, minimal) to reduce schema payload size and inference latency
Per-tool permission UI — Per-session allowlist/denylist without config reload
Chat-mediated Claude Code approvals — Pause for user approval when Claude Code needs permission, via file-based IPC
File read size cap — Configurable limit to prevent context blowout on large files
Per-session rate limiting — Throttle LLM calls per user/session to protect against runaway loops

Webhook hardening — HMAC signature verification for Telegram webhook mode
MCP stress testing — Dedicated test harness for adversarial transport conditions; replace custom client when an official Julia MCP SDK exists
Agent harness expansion — before_llm/after_llm hooks, on_error hook, headless run_once() entry point, session replay for regression testing
RAG-backed memory — Vector-indexed memory with semantic retrieval instead of full MEMORY.md dump each turn
Streaming responses — Token-by-token streaming to chat platforms via edit-in-place
Structured output — JSON Schema validation on LLM responses before tool dispatch
Multi-modal input — Forward images, voice, and documents from channels to the LLM
Image generation output — Display OpenAI image_generation results in Telegram/Discord

Additional LLM providers — Anthropic Claude (Messages API), local models via Ollama/llama.cpp
Versioned context directory — Migration layer for evolving skill and bootstrap doc schemas
Web dashboard — Local read-only UI for browsing session history, memory, and cron jobs
Platform comparison — Test deployment on Fly.io, Railway, bare VPS and document trade-offs