A local-first AI assistant that remembers and grows — available via Desktop, Cloud Server, and 12 IM channels.
Hope Agent is a local-first AI assistant designed for everyday users, running a shared core across Desktop GUI, HTTP/WS daemon, and ACP stdio modes.
Core differentiators include: (1) Cross-session persistent memory — SQLite + FTS5 full-text search + vector semantic retrieval, organized in Global/Project/Agent three-layer scope, with automatic "dream" consolidation of high-value memories during idle time and user profile extraction; (2) Auto-evolving skills — automatic skill draft generation after complex tasks, conditional activation, sub-Agent forked execution with tool whitelist isolation, compatible with agentskills.io open standard; (3) Five-layer progressive context compression — ensuring tool-call pairs are never split in long conversations, with automatic restoration of recently edited file content after summarization.
Workflow and tooling: Plan Mode with a 5-state state machine (Off → Planning → Review → Executing → Completed), Project containers for session grouping with inherited memory/instructions/shared files, Agent Team for one-command expert group assembly, natural-language Cron scheduling, Dashboard with four-dimensional visualization and 11-chapter AI recap reports. Toolbox covers controllable browser (CDP), AI image generation (7 providers), Web search (8 providers with failover), Bash execution (optional Docker sandbox), file read/write/grep/find, URL preview, crash logs, and self-diagnostics. MCP client supports all four transports (stdio / Streamable HTTP / SSE / WebSocket) with full OAuth 2.1 + PKCE flow.
Model support: 36 built-in Provider templates, 166 preset models, covering international providers (Anthropic, OpenAI, Google Gemini, etc.), Chinese providers (DeepSeek, Moonshot, Qwen, etc.), and local Ollama with one-click model installation. 12 IM channels (Telegram, Discord, Slack, Feishu, QQ Bot, WeChat, etc.) with automatic multimodal context conversion for images/voice/files, each group/account bindable to an independent Agent.
Security: local-first with zero third-party relay, all data stored in ~/.hope-agent/, tool approval gates, Docker sandbox, three-tier SSRF policies, auto config snapshots with one-click rollback, three-layer crash self-healing (Guardian process + OS-level launchd/systemd + per-component watchdogs). Frontend built with React 19 + TypeScript + Tailwind CSS v4 + shadcn/ui, supporting 12 interface languages.
Current version 0.1.0 with 1,400+ commits, actively developed. No binary release yet — build from source required. Windows platform not yet fully tested.