Hope Agent
A local-first desktop assistant that carries your tasks and memory across devices.
- Source repo
- shiwenwen/hope-agent
- Stars
- ★ 1.9k
- Last updated
- today
- License
- MIT
- Primary language
- Rust
- FA score
- 68/100 · Some gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · OpenAI API · Claude API
- Cost
- Free software; you pay for model usage
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- An individual tracking a multi-step research or work outcome on a Mac can use Goals, Workflows, Loops, and background Tasks to continue and resume execution.
- Not a fit if
- Users who need a fully supported Windows or Linux desktop experience; both are experimental
- Users unwilling to supply a Provider API key or install Ollama for local models
- Source review
- 68/100 · Some gaps
What does this agent do, and when should you use it?
Hope Agent is a personal AI assistant centered on a local desktop experience, with an option to run persistently as a service with a Web GUI. Its architecture combines a Rust core, a Tauri 2 desktop shell, and a React 19 frontend across Desktop, Server + Web GUI, and ACP modes. It keeps projects, chats, memories, goals, workflows, and background tasks in one system, with state stored by default in ~/.hope-agent/. Beyond chat, it can operate on files, run bash, call MCP services, control a browser and macOS with approval, and coordinate sub-agents. It also includes a Markdown-based knowledge space and a design space for producing exportable deliverables.
During onboarding, the user selects a Provider template and supplies an API key or signs in through Codex OAuth. Chats can be placed in Projects; global, project, and Agent memories are stored in layers and retrieved through full-text and vector search. A user can define a Goal and success criteria, while Workflow records stages, conditions, parallel work, tools, diffs, reviews, and verification for an execution; Loop resumes work on an interval, condition, or event, and Task shows progress. Subject to centralized approval, it can run bash, read and write files, call MCP through OAuth 2.1 transports, connect workspaces, control macOS, and operate pages through a controllable browser. The design space can produce web pages, mobile prototypes, presentations, and other artifacts for export as HTML, PNG, PDF, PPTX, MP4, or ZIP; the knowledge space reads and writes Markdown and supports search, backlinks, graphs, and Obsidian vault binding.
- An individual tracking a multi-step research or work outcome on a Mac can use Goals, Workflows, Loops, and background Tasks to continue and resume execution.
- A user running a personal assistant on a NAS, home server, or VPS can use hope-agent server, access the Web GUI in a browser, and keep IM channels and scheduled work online.
- A knowledge worker maintaining Markdown notes or an existing Obsidian vault can archive material, retrieve it, and review AI-generated organization suggestions in the knowledge space.
- A product or design practitioner who needs web pages, mobile prototypes, presentations, or posters from text or reference images can preview, refine, version, and export results in the design space.
- An advanced user combining APIs, shell commands, files, and MCP services can delegate work between the main chat and dynamic sub-agents under approval controls.
How do you install or deploy this agent?
On macOS, run:
brew tap shiwenwen/hope-agent
brew install --cask hope-agentFor self-hosting, run:
docker run -d --name hope-agent -p 127.0.0.1:8420:8420 -v hope-data:/data ghcr.io/shiwenwen/hope-agent:latestAfter the container starts, open http://127.0.0.1:8420. First use requires selecting a Provider template and providing that provider’s API key, or signing in with Codex OAuth; the local-model path installs and configures Ollama from Settings.
How do you use this agent?
Launch the desktop app, or run hope-agent server start and open http://127.0.0.1:8420. After configuring a Provider, start a chat; for ongoing work, define a Goal and success criteria, execute it with a Workflow, and set a Loop condition for future progress. For an IDE backend, run hope-agent acp. To access it from another device, set the server listener to 0.0.0.0:8420 and configure an API key; public deployment should place an HTTPS reverse proxy in front of it.
What are this agent's strengths and limitations?
- The same ha-core powers desktop, headless Server + Web GUI, and ACP modes, allowing chats, memory, and task state to continue between local and self-hosted environments.
- It combines Goals, Workflows, Loops, and Tasks with persistent execution records that can be paused, resumed, and conservatively recovered for work that extends beyond a single chat.
- It includes an MCP client, bash, file operations, browser control, and macOS control, with centralized approval for side effects and an optional Docker sandbox.
- Its Markdown knowledge space and design space cover both knowledge maintenance and production of deliverables exportable as HTML, PNG, PDF, PPTX, MP4, and ZIP.
- It documents 40+ Provider templates, 300+ preset models, Ollama local models, and API-key rotation, avoiding a single-provider model path.
- The complete desktop experience is explicitly centered on macOS; Linux and Windows are labeled experimental, and Windows is x64 only.
- Cloud models require the user’s own Provider API key or Codex OAuth, while local models require Ollama installation and model downloads.
- File writes, commands, browser control, and computer control have real side effects; approvals and Docker sandboxing still require the user to configure permissions and review actions.
- Remote access requires changing the listener and configuring an API key; public exposure additionally requires the user to set up an HTTPS reverse proxy.
- State lives by default in ~/.hope-agent/; the materials document Obsidian binding but do not document a general-purpose migration or import/export process.
How does this agent compare with similar options?
Compared with using Ollama directly, Hope Agent presents Ollama as a GUI-managed local runtime and OpenAI-compatible endpoint; models such as Qwen and Gemma are still distributed through the Ollama model library.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Hope Agent This agent | 68 · Some gaps | Desktop appFree + model costs | ★ 1.9k | today | Rust | Codex · OpenAI API · Claude API |
| Atom Platform | 67 · Some gaps | CLIFree + model costs | ★ 911 | today | Python | OpenAI API · Claude API |
| Commonly | 70 · Some gaps | Self-hosted serviceFree + model costs | ★ 1.4k | today | JavaScript | Codex · Claude Code |
| DeepChat | 70 · Some gaps | Desktop appFree + model costs | ★ 6.4k | today | TypeScript | OpenAI API · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
AGENTS.md describes concrete product boundaries around the permission engine, per-call approvals, SSRF checks, workspace paths, and token handling. The browser skill requires stopping for login, CAPTCHA, and permission prompts, while the Feishu skill calls for confirmation before approval actions. These support ordinary high-risk boundaries. Deductions: a launch flag can skip every tool approval for that process, and the packet does not include the full permission implementation or data flows. Publisher authorship is visible, but maintenance identity and update responsibility are not independently established.
The browser host and runtime entry point describe handling for disconnects, protocol mismatches, malformed frames, diagnostic errors, and update rollback. AGENTS.md also lists boundary checks and targeted verification entry points. Deduction: the packet does not include lockfile contents, complete failure paths, or executed verification results, so dependency availability is only partially assessable.
The materials cover desktop users, cross-device tasks, HTTP/WS, ACP, MCP, and headless deployment, with scenarios for browser use, IM, skills, goals, and workflows. Installation and tool-use constraints address multiple runtime environments. Deduction: the main README body is not included, so product boundaries and prerequisites for each deployment mode cannot be fully checked.
Root AGENTS.md is mainly contributor development and architecture guidance, not product-facing instructions; it provides a system documentation index and maintenance rules. The packet also contains detailed Feishu workflows, error-code guidance, browser procedures, and limitations, plus a clear MIT license. Deductions: README.en.md only points to the omitted README.md, and version history and maintenance responsibility are not fully established by the sample.
The claims about goals, memory, dynamic workflows, and cross-device agency align with repository components for desktop, server, ACP, skills, browser, and evaluation. Skill examples provide usable workflows and output formats. Deductions: the system has a broad scope and depends on external models, runtimes, browser extensions, and service configuration; the packet does not explain user costs or deployment complexity sufficiently.
The repository guidance, Cargo/package metadata, entry-point code, browser host, skills, and security policy corroborate parts of the product description and security claims. The evaluation CLI also shows implementation entry points for planning, running, aggregation, and evidence verification. Deductions: some security and evaluation assurances appear only in documentation or comments, and the packet lacks the related implementation, results, or full README for independent confirmation.
- The --dangerously-skip-all-approvals launch flag skips all tool approvals for the current process; understand its effect before using it.
- The packet omits the README.md body, lockfile contents, and full permission implementation, leaving some installation requirements, dependency state, and security assurances unverified by this static review.
Maintain this project? Add the FARS badge to your README
[](https://followagents.com/en/agents/hope-agent)<a href="https://followagents.com/en/agents/hope-agent"><img src="https://followagents.com/badge/hope-agent.svg" alt="FARS review"></a>Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
v0.72.0 Hope Agent v0.72.0
Release notes excerpt
[简体中文](https://github.com/shiwenwen/hope-agent/blob/v0.72.0/docs/release-notes/v0.72.0.md) · [English](https://github.com/shiwenwen/hope-agent/blob/v0.72.0/docs/release-notes/v0.72.0.en.md) Hope Agent v0.72.0 更多模型选择 在设置中添加服务商时，可选择 GPT-6.1 Terra、Kimi K3.1 Pro 及其高速型号、Unbiased Pareto 26.10 和火山方舟 DeepSeek V4.1 Flash。API 与 Codex 订阅的可用范围、上下文和费用各自独立，已有配置保留；请按账号权限与费用选择，新增预设不代表已完成真实账号验收。详见[模型与服务商](https://github.com/shiwenwen/hope-agent/blob/v0.72.0/docs/user-guide/02-模型与Provider.md)。 长任务与设计检查更准确 …View on GitHub ↗ -
v0.71.0 Hope Agent v0.71.0
Release notes excerpt
Hope Agent 0.71.0 — 图像与转写新预设，减少配置误操作 > 简体中文 · [English](https://github.com/shiwenwen/hope-agent/blob/v0.71.0/docs/release-notes/v0.71.0.en.md) > 发布日期：2026 年 10 月 30 日 > 完整改动列表见 [CHANGELOG](https://github.com/shiwenwen/hope-agent/blob/v0.71.0/CHANGELOG.md#0710---2026-10-30)。 新建图像服务商时有更多画幅选择 在设置中新建 Google 图像服务商，首选预设更新为 Nano Banana 2.1，可选更宽或更长的画幅。图像工具与 Agent 的图像产物生成可使用最多 14 张参考图；设计空间首页仍最多 5 张。现有模型和默认链不会自动更换，当前接入不支持蒙版重绘或视频、PDF 输入。新型号的真实画质与费用尚未验收。 按录音方式选择转写预设 …View on GitHub ↗ -
v0.70.0 Hope Agent v0.70.0
Release notes excerpt
Hope Agent 0.70.0 — 更多模型选择与消息接收修复 > 简体中文 · [English](https://github.com/shiwenwen/hope-agent/blob/v0.70.0/docs/release-notes/v0.70.0.en.md) > 发布日期：2026 年 10 月 27 日 > 完整改动列表见 [CHANGELOG](https://github.com/shiwenwen/hope-agent/blob/v0.70.0/CHANGELOG.md#0700---2026-10-27)。 更容易找到入口与模型 侧栏图标下方现在常驻显示入口名称，长名称仍可悬停查看完整提示；小窗口可通过“更多”找到折叠入口。 添加服务商时，可选择 Unbiased 的 Pareto 稳定版或预览版；Google 模板新增 Gemini 4 Argon，Mistral 模板新增 Z.ai GLM 5.3 Flash。填入对应服务商的 API Key 后即可配置使用。ChatGPT 登录的新建 GPT-6.1 Sol 预设采用 500K 上下文；已保存的上下文和当前模型不会自动改写。账号资格、模型可用性和实际费用仍由服务商决定，模板参数不代表真实账号验收。 减少消息误触发与接收损坏 …View on GitHub ↗
FAQ
Where is data stored by default?
Can it run without a cloud model?
How can other devices access it safely?
What happens if a background task exits unexpectedly?
Thanks — we use this to improve the page.