Evo Harness: a chat that added a retry to web_fetch, next to the edited fetch.rs with its git changes in the Files panel **A local-first desktop workspace for AI agents.** Evo Harness brings agent chats, files, terminal, browser, git, connectors and skills into one desktop app. A headless Rust engine runs every chat on your machine; the desktop app is a thin shell around it. [![License: Apache 2.0](https://img.shields.io/badge/License-Apache%202.0-2563EB.svg)](LICENSE) [![Desktop app](https://img.shields.io/badge/Product-Desktop-1764FF)](apps/desktop/) [![Rust 1.89+](https://img.shields.io/badge/Rust-1.89+-000000?logo=rust&logoColor=white)](Cargo.toml) [![Tauri v2](https://img.shields.io/badge/Tauri-v2-FFC131?logo=tauri&logoColor=white)](apps/desktop/) **[Website](https://evoelsewhere.asia/)** · **[Features](#core-features)** · **[Download](#download)** · **[Run from source](#run-from-source)** · **[Documentation](#documentation)**
## Highlights - **One engine process per chat.** A crashing chat, tool or plugin does not take the window down. Sessions are saved as they run and resume after a crash. - **Bring your own model** from hosted, cloud or local providers. Models, limits, costs and thinking options come from models.dev, not from code. - **Workspace tools beside the chat:** Files with a code editor and Git changes, Terminal, a Browser the agent can drive, Memory and a side chat. - **Plugins with their own screens, code and AI.** Ask Evo to build one with you (Creator Mode), or install one: it can add pages and panels, run a backend, sign in to services and use your own models, only as you allow. - **You stay in control:** approval prompts, a sandbox, project trust, and the usage and cost of every response, all kept on your machine. ## Core features Animated walkthroughs of Customize, Browser, Computer App Control and Office are on the [website](https://evoelsewhere.asia/#core). ### Customize Shape how the agent works, per person or per project, from **Customize** in the sidebar. Changes apply without restarting the app. - **Skills and slash commands** in the open Agent Skills (`SKILL.md`) format, from Evo's folders, `.agents/skills` or plugins. Each skill has a switch, a "who can use it" setting and a usage count; type `/` in the composer to run one. - **Connectors (MCP servers)** over stdio, streamable HTTP or SSE, with OAuth sign-in for remote servers. One shared connection pool serves every chat. Each tool is set to Allow, Ask or Block, and connectors can be switched off per chat from the composer's `+` menu. - **Plugins and marketplaces** bundle skills, commands, subagents, hooks and connectors, and can add views, a backend, accounts and AI of their own ([below](#plugins-and-creator-mode)). Install them from a GitHub `owner/repo`, any git URL or a folder. - **Hooks** run a command or a web request on session, prompt and tool events, and can block, allow or rewrite a tool call or add context. - **Instructions** for every chat (Personal) and for one project. A project's skills, hooks, instructions and `.mcp.json` servers are only used after you trust the folder and approve them. [Spec](docs/features/customize-skills-plugins-connectors/spec.md) ### Plugins and Creator Mode A plugin is a folder Evo installs. Besides skills and connectors, it can bring its own screens, code and AI, and it gets only what you approve in its consent dialog. Each plugin has a page in **Customize → Plugins** with its settings, accounts, permissions, runtime, data and models. - **Creator Mode.** Ask in a chat for the plugin you want. The bundled `evo-creator` skill leads the agent through the folder, the manifest, the code and a design that matches Evo, then you add it from its folder and turn on Developer mode to see each edit at once. - **Views in the app:** a sidebar page, a side panel, chat cards and composer actions. They are native webviews, one origin per plugin, that reach Evo only through `window.evo`, with every call checked against the permissions you approved. Incognito chats are invisible to them. - **Its own code and accounts.** A backend runs on demand under Deno, limited to the sites and folders you allowed, or with uv for Python. Settings can be secret, sealed in a vault whose key is in your OS keychain. Accounts sign in with OAuth, a device code, an API key or a web login, and Evo renews tokens. The plugin's pages see an account's state, never its token. - **Your AI models, natively.** With `ai`, a plugin streams answers from the models you connected, chosen by need (fast, balanced or best) among those you allow for it, with one fallback when a provider fails. Its runs can call functions of its own page, tools of its backend, and Evo's read, edit, command, web, browser and terminal tools under your approval mode. Runs keep threads per subject and can attach the open chat or files you allowed. The plugin never sees a key; its **Models** tab shows what it spent and its runs, and sets caps per day and month. [Plugin UI](docs/features/plugin-ui/spec.md) · [Settings, secrets and runtime](docs/features/plugin-settings/spec.md) · [Plugin AI](docs/features/plugin-ai/spec.md) · Samples: [every view slot](examples/plugin-ui-sample/README.md), [settings, accounts and AI](examples/plugin-settings-sample/README.md) ### Browser The agent gets its own browser inside the app and works in it while you keep chatting. - **Floating cards.** Each page the agent uses floats over its chat as a live card (up to three per chat), so you do not have to open a panel. The card shows what the agent is doing, and you can minimise it, drag it, resize it or open it in the side panel. - **A virtual cursor** shows where the agent points and clicks. You can click and type in the card yourself, for example to sign in. - **Desktop-sized pages.** Pages the agent drives always lay out at 1280 × 800 and are scaled to fit. - **Real browser tools:** navigate, read the page as an outline, text or Markdown, click, type, drag, wait, handle dialogs, upload files and run up to 20 steps in one batch. The agent can also set the viewport, colour scheme, locale and time zone, edit cookies and storage, mock requests, read the console and network, save a PDF and record performance traces. - **Approvals by site.** Reading a page never asks. Clicks and typing on a site that is not on your computer ask once per site per chat in Ask every time. Uploads, storage writes, mocks and scripts ask in Ask every time and Approve edits. Trusted input, screenshots and page cards are Windows-only for now; on macOS and Linux, clicks and typing are scripted. [Spec](docs/features/agent-browser/spec.md) · [Floating cards](docs/features/floating-cards/spec.md) **WebBridge** (in development) will let the agent work in a tab of your own Chrome or Edge, where you are already signed in. Nothing is shared until you pick a tab, a shared tab belongs to one chat, and you can stop sharing at any time. [Spec](docs/features/browser-bridge/spec.md) ### Computer App Control (preview, Windows) Ask for work in a desktop app you already use, such as a spreadsheet, a document or a chat app. The app's window moves off your screen while the agent works in it (it is not minimised or hidden), and a live card shows it with a virtual cursor. - **Never your mouse or keyboard.** Evo does not move your pointer, change your keyboard focus or front window, or use the clipboard. - **Every action asks first** by default, even in Approve for me. The prompt lists each click and keystroke as it will run. - **It checks its own work.** The agent acts on identified UI elements first, reads the app back after each step and says what changed. When something cannot be done in the background, it says so. - **You stay in charge.** **Stop** ends the current action until you choose **Allow again**. **Show the app** brings the window back and ends control for the turn. **Close** hands the app back. The window returns exactly where it was when control ends. Turn it on in **Settings → Computer apps** (off by default), where you can also allow or block apps. Evo itself, system processes, apps running as administrator and password fields are never controlled. [Spec](docs/features/computer-app-control/spec.md) ### Office documents Word, PowerPoint and Excel files open in **Files**, rendered natively with no download, network access or scripts. Bundled `docx`, `pptx` and `xlsx` skills let the agent create and edit them. - **A real viewer:** zoom, search and thumbnails. Decks have Normal, Sorter and Reading views with speaker notes. Workbooks have sheet tabs, a name box, the formula bar and a print layout. - **Live build.** A deck or workbook fills in slide by slide or sheet by sheet while the agent builds it, and new files open by themselves. - **Point, then ask.** Use **Select an area to edit** on a slide, or **Select cells to edit** in a sheet, and write what should change. The agent changes only what you marked. - **Every version kept.** Undo, Redo and History step through the last 50 saves. For exact print pages, install the optional LibreOffice runtime from the viewer's banner. [Spec](docs/features/office/spec.md) · [Viewer and editing](docs/features/office-v3-parity/spec.md) ## More features | Feature | What it does | | --- | --- | | [Subagents and nested sessions](docs/features/subagents-and-nested-sessions/spec.md) | Delegate work to a child agent with a fresh context: general, read-only explore or plan, or your own types in `.evo/agents`. Background agents report back when they finish. Nested sessions are full chats under their parent. Either can work in its own Git worktree, which you review and then apply or discard. | | [Code intelligence](docs/features/code-intelligence/spec.md) | Language servers for 21 languages give the agent diagnostics, definitions, references and symbols. They install on demand from **Customize → Code intelligence**, and servers you already have are reused. | | [Git in Files](docs/features/files-git/spec.md) | Changed files are marked in the tree and changed lines in the editor gutter. Click a gutter marker to see the committed version. | | [Terminal](docs/features/agent-terminal/spec.md) | Every shell on your machine, including PowerShell, Git Bash, WSL and Visual Studio developer shells. The agent can run commands, wait for output and send keys in its own tabs, which are marked with a bot icon. | | [Approval modes](docs/features/approval-modes/spec.md) | Choose Ask every time, Approve edits or Approve for me from the composer. Reads never ask. A change applies to every running chat at once. | | [Memory and Scheduler](docs/features/memory-and-scheduler/spec.md) | Memories persist across chats; project memories are kept outside the repository. Scheduled prompts run as new chats, manually, once, hourly, daily, weekly or on a cron schedule. | | [Usage and cost](docs/features/usage-and-cost/spec.md) | The tokens and models.dev cost of every response, broken down by model, chat and source, with a price check against other models. Nothing leaves your computer. | | [Web search and web fetch](docs/features/web-search-and-fetch/spec.md) | Search that works with any model and needs no key, and public pages read as Markdown. | | [Notifications](docs/features/notifications/spec.md) | A notification bell, toasts and OS notifications for finished runs, approvals and scheduled tasks, plus Do not disturb. | | [Model providers](docs/features/app-database/spec.md) | Connect providers in **Settings → Providers** with an API key or by signing in. Model lists come from models.dev, refresh without a restart and have a bundled snapshot for offline use. Credentials stay in an owner-only file, not in the app database. | The full list, with the status of each feature, is in the [feature catalogue](docs/features/README.md). ## Download Download the latest desktop release from [GitHub Releases](https://github.com/evoelsewhere/evo-harness/releases). Packages are available for macOS (Intel and Apple Silicon), Windows and Linux. Each package includes the `evo-engine` sidecar. Linux packages should be installed through the system package manager, for example: ```sh sudo apt install ./Evo*_amd64.deb ``` The LibreOffice runtime for Office document previews is optional. The app offers to install it the first time you open an Office file. ## Run from source ### Requirements - Rust (the toolchain is pinned in `rust-toolchain.toml`; MSRV 1.89) - [Bun](https://bun.sh/) - [Tauri prerequisites](https://v2.tauri.app/start/prerequisites/) for your operating system - On Windows, [NASM](https://www.nasm.us/) for a clean build ### Setup ```sh git clone https://github.com/evoelsewhere/evo-harness.git cd evo-harness/apps/desktop bun install ``` Start the development app with: ```sh # Build evo-engine, stage it as the sidecar, then run the Tauri app bun run dev:app ``` Use bun only in `apps/desktop`; do not use npm, pnpm or yarn there. `bun run build:app` builds a release with installers for the current platform. The engine and library crates build and test from the repository root: ```sh cargo build --workspace cargo test --workspace ``` On first launch, connect a model provider in **Settings → Providers**, then open a folder and start a chat. ## Documentation Each feature above links to its spec. Detailed contracts and guides live under [`docs/`](docs/): - [Documentation index](docs/README.md) - [Architecture (ADD) and decisions (ADRs)](docs/architecture/README.md) - [Feature catalogue: specs, plans and tasks (Spec-Driven Development)](docs/features/README.md) - [Desktop UI map and source layout](apps/desktop/README.md) - [Visual design language](DESIGN.md) - [Contributor and agent guide](AGENTS.md) ## Architecture at a glance Evo Harness architecture: the React UI talks to the Tauri shell over IPC; the shell starts one evo-engine process per chat and relays JSONL to it, and hosts the shared connector pool; the engine works on project folders, saves sessions to local state, calls model providers over HTTPS and reads model data from models.dev. The engine owns agent behaviour: the agent loop, tools, sessions, compaction, skills, hooks, memory, model selection and credentials. The shell owns native features (windows, terminals, files, browser webviews, the MCP connection pool, plugins) and relays the protocol. Plugins' model calls run in one more engine process (`evo-engine --mode plugin-ai`) that the shell gates. Model providers are replaceable, while the harness handles context, tools, permissions, memory and verification. ## Project layout ```text crates/ Rust workspace: engine, protocol, agent loop, tools, providers apps/desktop/ Tauri 2 shell (src-tauri/) and React 19 UI (src/) examples/ sample plugins (views; settings, accounts and AI) and the native plugin used by extension tests scripts/ Office runtime build, README screenshots docs/ architecture (ADD, ADRs) and feature spec / plan / tasks ``` The crates are released together and depend only on the layers below them: Evo Harness crate layers, top to bottom: App (evo-engine, apps/desktop/src-tauri); Protocol (evo-protocol, evo-mesh); Runtime (evo-harness, evo-durable, evo-runtime, evo-extensions, evo-plugin-sdk); Agent (evo-agent, evo-tools, evo-worktrees); LLM (evo-ai); Leaf services with no internal dependencies (evo-store, evo-mcp, evo-plugins, evo-hooks, evo-lsp, evo-web, evo-office, evo-telemetry). Diagram and banner sources are in [`docs/assets/readme/`](docs/assets/readme/); [`scripts/readme-shots/`](scripts/readme-shots/) regenerates the screenshots. ## Contributing Issues and pull requests are welcome. Keep changes focused and include the smallest relevant test run. [AGENTS.md](AGENTS.md) lists the checks for each layer and the design-document workflow. Report vulnerabilities privately through GitHub Security Advisories. ## License Evo Harness is released under the [Apache License 2.0](LICENSE). Parts of the `evo-*` crates are derived from MIT-licensed upstream work; see [NOTICE](NOTICE).