A desktop AI agent where the chat box is the entire control panel

What it does

Deva is a cross-platform Electron desktop agent client (TypeScript, MIT) where the conversation is the only interface. Creating personas, scheduling cron jobs, installing skills, attaching MCP servers, and mounting local project directories all happen through chat, gated by user confirmation. Once a workspace is mounted, the agent can read and modify code, run commands, and dispatch parallel subagents for large tasks. All data stays on the local machine.

The interesting bit

The prompt architecture is layered with non-overlapping responsibilities: a built-in system spec (tools, safety, engineering discipline), project conventions from AGENTS.md, memory from memory.json, and persona files — personas shape tone but never touch tool or safety rules, so switching characters doesn’t degrade engineering capability. Subagents deliberately get no persona injection, keeping research output uncolored by roleplay.

Key highlights

• Five subagent types (General, Explore, Research, Plan, fork) with independent contexts, parallel and background execution, and follow-up via send_to_subagent
• /rewind checkpoints: roll back to any turn, with code and conversation restorable separately — and the rollback itself is undoable
• Unattended scheduled tasks run in an isolated environment: no ask_user, read-only memory, results land in a dedicated conversation with a system notification
• Auto-compaction at 75% of the context window, plus incremental delivery of skills/memory so prompt caching survives updates
• Continue the same session from your phone via Feishu/Lark or Telegram bots; presets for Anthropic, OpenAI, Gemini, DeepSeek, Kimi, GLM, Qwen, and Ollama
• Built-in persona 小码酱 for prompt-injection research and jailbreak work on Chinese models (prompts sourced from YuJunZhiXue/dsh-purge)

Caveats

• Installers aren’t code-signed (macOS is ad-hoc only), so expect OS warnings and manual overrides on first run
• The 小码酱 persona is tuned specifically for Chinese models (DeepSeek, Kimi, GLM, Qwen); effectiveness elsewhere is explicitly not guaranteed
• Scheduled tasks only run while the app is open; missed runs are caught up only once

Verdict

Worth a look if you want a local-first, chat-driven agent client with unusually thoughtful prompt layering and checkpointing. If you need a signed, enterprise-grade desktop app or a server-side deployment, this isn’t that.