evoelsewhere/evoflux
Evoflux is an open-source, local-first workspace where AI agents build software, conduct deep research, automate browser tasks, and collaborate in parallel. Connect any model, keep control of your workspace and data, and take complex work from idea to completion—all in one place.
安装Install
dsh plugin --profile web add github:evoelsewhere/evoflux把 evoelsewhere/evoflux 加入你的 DSH 配置（web profile）即可启用。
READMEREADME
EvoFlux
The local-first desktop workspace for AI agent teams.
Give EvoFlux an outcome. A lead agent plans the work, brings in specialists, uses the right tools, and verifies the result — while you stay in control.
Cowork and software engineering. Any model. Your machine.
Download EvoFlux · Product tour · Quick start · How it works · Architecture · Capabilities
Plan with the team, inspect the implementation, and navigate the repository without leaving the workspace.
[!NOTE] Since 30 June 2026, fixes, optimizations, and new EvoFlux features have been developed and delivered using EvoFlux Coding mode. The agents build, review, and ship themselves.
Why EvoFlux
| Delegate outcomes | Keep the whole job together | Choose every model | Own the execution |
|---|---|---|---|
| A lead coordinates focused specialists and verifies their handoffs. | Chat, files, terminal, browser, memory, git, and previews live in one workspace. | Mix providers, models, reasoning levels, skills, and tools per agent. | Local runtime, scoped access, outbound redaction, and inspectable history. |
Product tour
One app, two specialized modes
One desktop app. One harness. Two different kinds of work.
| Work | Coding | |
|---|---|---|
| Product role | Cowork | Software engineering workspace |
| Workspace | Temporary sandbox | Persistent repo or multi-repo project |
| Best for | Research, documents, data, browser work, quick scripts | Build, test, refactor, review, git operations |
| Default specialists | Executor, Explorer, Consultant, Debate | Coder, Explorer, Architect, Debate |
| Verification | Artifact and tool-result review | Tests, diffs, code context, git |
Work is a fast execution sandbox for research, documents, data, browser tasks, files, and quick scripts. Start with a request instead of a repository.
Coding opens one or more real repositories and keeps them available across sessions. Agents can understand the codebase, edit and test code, review diffs, and use the complete git surface.
 |
 |
| Start with the outcome Drop into Work and describe what you want accomplished. |
Bring every tool into view Open terminal, processes, browser, files, side chat, memory, and scheduler beside the conversation. |
Assemble the right team and models
Create role-focused agent teams, then tune the model and capabilities of every member independently. EvoFlux ships with 19 provider integrations, including direct APIs, subscription OAuth, cloud platforms, local runtimes, and model routers.
 |
 |
| Lead-and-specialists Give each agent one role, one model, and a focused capability set. |
Bring your own model Connect hosted, subscription, routed, cloud, or local providers from one catalog. |
Understand the codebase, not just matching text
EvoFlux indexes symbols and relationships across every repository in a Coding project. Explore the graph visually, trace callers and dependencies, and pair structural context with repository-aware language servers.
Local control is part of the product
Language servers provide semantic feedback in the active project. Sandbox controls scope filesystem and process access, mask or block sensitive outbound data, and keep execution boundaries visible rather than hidden behind the agent.
 |
 |
| Semantic feedback Detect project languages and reuse managed or system language servers. |
Explicit boundaries Control outbound data, host environment access, shell behavior, and denied paths. |
Download
Current stable release: EvoFlux v0.0.8
| Platform | Package | SHA-256 |
|---|---|---|
| macOS · Apple Silicon | Download DMG | Checksum |
| macOS · Intel | Download DMG | Checksum |
| Windows · x64 | Download installer | Checksum |
| Linux · x64 | Download DEB | Checksum |
Install the Linux package with sudo apt install ./EvoFlux_*_amd64.deb.
Linux updates use the same package-managed flow instead of replacing
dpkg-owned files through the in-app updater.
The desktop packages include the native Python sidecar. The optional WebBridge browser companion is not bundled in EvoFlux release assets; its source and installation instructions live in the separate evo-webbridge repository.
[!NOTE] The v0.0.8 macOS packages use an ad-hoc signature and the Windows installer is not Authenticode-signed, so the operating system may show a trust prompt. Tauri updater archives remain minisign-signed.
Quick Start
Install the desktop app
Choose the package for your platform in Download, install it, and launch EvoFlux. The packaged app includes its Python sidecar.
On macOS and Windows, updater-aware builds check the latest GitHub Release after
startup. You can also run a manual signed update check from Settings > About,
the application menu, or the tray menu. Install one current package manually;
later releases can update in place. Linux updates use a newly downloaded DEB
through apt or dpkg.
On first launch:
- Connect an LLM provider.
- Start a Work session or open a repository for Coding.
- Choose the model, reasoning level, skills, tools, and permissions for each agent.
Run the desktop app from source
Requirements: Python 3.12+, uv, Bun, Rust, Tauri CLI, and the Tauri prerequisites for your operating system.
git clone https://github.com/evoelsewhere/evoflux.git
cd evoflux
uv sync
cd web && bun install && cd ..
# Choose browser development — local API + React development server
make dev-web
# Or choose desktop development — local API + React + Tauri desktop shell
make dev-desktop
make dev remains an alias for make dev-web. localhost:5173 is the internal frontend development server used by Tauri during development. EvoFlux is shipped and positioned as a desktop product, not a standalone web app.
Agent working model
EvoFlux operates under a lead-and-specialists model. Each request is analyzed by the Lead Agent to determine scope and complexity.
- A simple task stays with the Lead.
- A complex task is broken into well-defined subtasks with explicit goals, outputs, and constraints.
- Specialists activate on demand, work in parallel, and exchange results through a shared mailbox.
- The Lead evaluates handoffs and evidence, requests rework when needed, and synthesizes the final response.
Configurable per agent
| Configuration | Why it matters |
|---|---|
| LLM model | Use a fast model for routine execution and a stronger reasoning model for architecture or review |
| Thinking level | Tune latency and reasoning depth by role and model capability |
| Skills and tools | Add agent-specific capabilities or disable code-owned defaults with explicit opt-outs |
| Permissions and access scope | Limit what an agent can read, write, execute, or approve |
The result is higher parallel capacity, less context noise, the right model for each job, verified delivery, and an execution history that can be inspected instead of trusted blindly.
Agent Markdown is the user-owned override surface. Runtime and Settings compile
the same effective config from the mode profile plus frontmatter additions and
tools_opt_out; reads and validation never materialise
configuration files. See
documents/architecture/application-harness.md.
Architecture
EvoFlux is desktop-only:
Tauri Desktop → React UI → local FastAPI sidecar → local state / model providers
The production app launches a local sidecar through an ephemeral port and token handshake. The React interface, agent runtime, repository-local code indexes, memory engine, scheduler, permissions, and MCP client all run on the user's machine.
What makes it a harness
A language model generates reasoning. The harness turns that reasoning into controlled action:
| Layer | Responsibility |
|---|---|
| 1. Tool orchestration | Shell, filesystem, git, browser automation, MCP, and agent-to-agent actions |
| 2. Guardrails | Permissions, policies, approvals, filesystem sandboxing, command checks |
| 3. Context and memory | Workspace state, sessions, code indexes, compaction, knowledge wiki |
| 4. Verification loops | Test, compare, review, debate, reject, rework, and evidence |
| 5. Observability | Streaming events, telemetry, logs, metrics, diagnostics, and audit history |
The model is replaceable. The harness — context, action, policy, verification, and state — is the product.
Core capabilities
Multi-agent teams
Agents are Markdown files with YAML frontmatter (name, role, model, thinking_level), making teams readable, diffable, and versionable. A team has one Lead and any number of on-demand members. Multiple instances of the same blueprint can work in parallel without becoming always-on background processes.
Durable Goal mode
Start an autonomous objective in any mode with /goal <objective>. Goal state,
elapsed time, token usage, and an optional token budget survive reconnects and
app restarts. The team continues through hidden internal turns until the Lead
records completion, the budget pauses execution, the user pauses it, or the
same concrete blocker is reported three turns in a row. Goal mode never expands
the session's permissions or sandbox scope.
Use /goal to inspect status, /goal:budget <tokens|none> to change the budget,
and /goal:pause, /goal:resume, or /goal:stop to control the objective.
Repository-local code context
EvoFlux ships a repository-local code index based on stable source keys and desired-state reconciliation. The implementation is part of the application and adds no indexing framework dependency. Twenty-five tree-sitter parsers cover Python, TypeScript/TSX, JavaScript, Go, Rust, Java, C#, C, C++, Swift, Kotlin, PHP, Ruby, Scala, Dart, Objective-C, Lua, Luau, R, Pascal, Svelte, Vue, Astro, and Liquid.
Each repository owns a managed SQLite target in the EvoFlux cache. A refresh fingerprints source bytes, parser and pipeline implementations, and project settings; parses only additions and changes; removes deleted components; and atomically replaces their committed source snapshot, AST-aware overlapping chunks, local code vectors, symbols, relations, and FTS rows. The vectorizer is implemented with Python's standard library, so the runtime adds no model or vector-database package. Parse failures preserve the last good component and are surfaced in status/query limitations. The application database stores projects and sessions but no code-index or graph data.
Cross-repository links are resolved at query time across only the repositories authorized for the active project. Resolution prefers same-file and lexical definitions, import bindings and module paths, then a unique cross-repository definition. There is no persisted cross-repository guess, resolver tier, background resolver job, or model-facing scope switch.
The model receives one native code_context tool:
| Question | Action |
|---|---|
| Find code from a concept or source phrase | search |
| Match a syntax shape with metavariables | grep |
| Locate a known symbol | definition |
| Follow incoming or outgoing calls | callers or callees |
| Inspect direct or transitive relationships | references, impact, or neighborhood |
The first query normally uses refresh=true; immediate follow-ups over the same indexed version can use refresh=false. Structural results are static evidence, so runtime-only behavior still requires tests, logs, LSP, or debugger evidence. The full storage, query, ambiguity, and tool contract is documented in documents/architecture/coding-agent-code-context.md.
Coding's repository-local LSP, automatic post-edit feedback, Guarded ChangeSets,
Problems hub, explicit AI editor/Git actions, and Search Everywhere contracts are
documented in documents/architecture/coding-semantic-intelligence.md.
Memory and Dream
The scheduled or manually triggered Dream agent consolidates sessions and notes into an inspectable Markdown wiki: topics/, entities/, notes/, and imports/, with INDEX.md, an append-only LOG.md, source citations, confidence, and related-page metadata.
Bring your own model
Nineteen provider integrations ship behind one streaming abstraction, including Anthropic, OpenAI, Google Gemini, AWS Bedrock, Ollama, DeepSeek, xAI, Vertex AI, and GitHub Copilot. Models can be selected independently for each agent.
Skills and MCP
Twenty-nine built-in skills cover mode-scoped Work and Coding workflows, specialized artifacts/design, EvoFlux configuration/installers, portable plugin development, and provider-neutral PR lifecycle operations. Work and Coding each expose one implicit router; broad specialists are explicit-only so they do not compete on every request. Custom skills can be created, edited, diagnosed, and filtered as Work, Coding, or Both in Settings. A bounded 2%/8K metadata catalog is always available for model-driven selection, while SKILL.md bodies and bundle resources load only after exact activation. EvoFlux is also an MCP client for stdio, HTTP, and SSE servers; connected tools inherit the same permission rules as native tools.
The built-in Plugin Center implements the portable Agent Plugins 1.0 core. It can scaffold, validate, import, developer-link, pack, update, enable, disable, and uninstall local plugins containing immediate-child Agent Skills and isolated stdio or Streamable HTTP MCP servers. .evoplugin is a deterministic ZIP distribution wrapper; the unpacked package remains standards-compatible through root plugin.json and optional mcp.json.
Plugin packages remain portable content bundles: they contribute Skills and MCP server declarations, while EvoFlux owns validation, lifecycle, permissions, credentials, installation data, and runtime status. New imports remain disabled until the user reviews executable commands, remote hosts, environment-field names, and capabilities. Plugin MCP servers run in an isolated manager instead of being merged into the user's global MCP configuration. See the Agent Plugin setup guide to use a package and the portable Agent Plugin architecture for the package contract, runtime boundaries, storage model, and failure isolation rules.
Permissions and sandboxing
Wildcard (tool, pattern) → allow | deny | ask rules use last-match-wins evaluation. The denylist filesystem sandbox protects EvoFlux state and cache directories, rejects symlinks into blocked roots, and tokenizes shell commands for denied-path checks.
Git and session UX
Coding mode exposes diff review, commits, branches, merge, rebase, cherry-pick, stash, and worktrees to agents and the source-control UI. Long sessions support prompt navigation, revert/undo boundaries, context compaction, four-pane Split view, and a unified Monitor view.
WebBridge
WebBridge is an independently distributed browser companion for the EvoFlux desktop app — not a web version of EvoFlux. Its extension source and release lifecycle live in the dedicated repository rather than this desktop repository.
It connects an agent to the user's real Chrome or Edge session through a persistent, policy-checked relay. Control flows from the desktop agent to the browser over CDP; selections, page context, and human handoff flow back to the desktop session.
| Capability | What it does |
|---|---|
| Secure connection | Pairs the browser extension with scoped credentials, uses one-time session tickets, enforces domain policies, and maintains a complete audit trail. |
| Safe context sharing | Lets users intentionally share a selection, link, or page while sanitizing metadata, preserving provenance, and treating browser content as untrusted input. |
| Live collaboration | Streams the agent session into the browser side panel, supports questions and element selection, and allows seamless control handoff between the user and agent. |
| Teach and monitor | Records meaningful browser actions without capturing raw keystrokes, redacts sensitive fields, creates reviewable workflows, and requires confirmation before monitored results are shared. |
Pairings, tickets, tab bindings, and Teach drafts are persisted through Alembic migrations. Revoking a pairing closes the live relay and invalidates outstanding tickets. EvoFlux shows live connection status and pairing controls in its WebBridge panel; extension installation and distribution are handled by the evo-webbridge repository.
Beyond the real-browser bridge
EvoFlux also includes direct control of its persistent in-app browser, PDF/HTML intake through markitdown, read-only DOCX/XLSX/PPTX workspace previews, cron-driven agent prompts, OpenTelemetry, Prometheus, and DuckDB-backed observability summaries.
How EvoFlux compares
Compare deployment, models, memory, code intelligence, and browser integration
| EvoFlux | Claude Code | Cursor | Devin | OpenAI Codex | OpenHands | |
|---|---|---|---|---|---|---|
| Interface | Desktop app | CLI, IDE, desktop, web | VS Code fork | Cloud + desktop + CLI | CLI, cloud, IDE | Web, CLI, API |
| Deployment | Local, self-hosted | Local + optional cloud | Local IDE + cloud agents | Cloud/VPC + local desktop | Local + cloud sandbox | Self-hosted or cloud |
| Open source | Apache-2.0 | No | No | No | CLI only | MIT |
| Bring your own model | 19 providers | Partial proxy setups | Partial BYOK | Provider choice | OpenAI only | Any model |
| Non-project cowork | Work | Ad hoc | No | Limited | No | Yes |
| Multi-agent | Lead + on-demand specialists + mailbox | Subagents and teams | Agent fleets + worktrees | Sub-Devins | Up to six subagents | Parallel delegation |
| Code understanding | Structural graph, 25 parsers, cross-repo | Search + optional LSP | Embedding search | Codebase Q&A | Repo-aware loop | Agent-computer interface |
| Persistent memory | Inspectable wiki + Dream | Markdown + auto-memory | Project Memories | Org knowledge base | AGENTS.md + session memory |
Condenser + skills |
| Real-browser bridge | WebBridge, two-way | No | No | No | No | No |
| Pricing | Free; pay model costs | Subscription or API | Subscription | Subscription + usage | ChatGPT or API | Free self-hosted / paid cloud |
EvoFlux leans into local ownership, model choice, inspectable memory, general cowork, and structural code intelligence. Commercial products lead in vendor-specific coding models, cloud infrastructure for long unattended runs, and editor-native maturity.
Competitor information reflects publicly reported product capabilities and pricing around mid-2026 and may change.
Tech stack
| Layer | Technology |
|---|---|
| Desktop | Tauri v2, Rust, bundled Python sidecar |
| Frontend | React 19, TypeScript 5.9, Vite 7, Tailwind CSS v4, Zustand, TanStack Query and Router |
| Backend | Python 3.12+, FastAPI, SQLModel, Alembic |
| Streaming | Server-Sent Events through sse-starlette; one feed per session |
| Data | SQLite WAL or PostgreSQL/MySQL; Markdown knowledge wiki |
| Code intelligence | tree-sitter, tree-sitter-language-pack, SQLite FTS5 |
| Observability | OpenTelemetry, Prometheus, DuckDB-backed aggregation |
Project layout
app/ Local FastAPI sidecar — agents, code context, memory, scheduler, MCP
web/ React interface embedded by the Tauri desktop app
desktop/ Tauri v2 shell and Python sidecar packaging
seed/ Work and Coding blueprints, skills, and config
tests/ Backend and frontend tests
documents/ Design notes, analyses, and README media
Contributing
Issues and pull requests are welcome. Keep changes focused and include the smallest relevant test run. Please report vulnerabilities privately through GitHub Security Advisories.
License
EvoFlux is released under the Apache License 2.0.
相关推荐Related
- deepseek-ai/deepseek-harness DeepSeek Harness: Everything is a Plugin.
- deepseek-ai/deepseek-harness DeepSeek Harness (dsh) — 由 DeepSeek AI 开发的开源 agent harness（智能体框架）。一切皆插件：核心 + 插件 + 技能 = 可热插拔的 AI 编程工作台。DeepSeek Harness: …
-
nexu-io/open-design
🎨 Best DeepSeek Harness Design Plugin. The open-source Claude Design alternative. 🖥️ Local-first desktop app. 🖼️ Your…
-
ruvnet/ruflo
🌊 The original agent meta-harness. Deploy intelligent multi-player swarms, coordinate autonomous workflows, and build c…
-
amruthpillai/reactive-resume
A one-of-a-kind resume builder that keeps your privacy in mind. Completely secure, customizable, portable, open-source a…
-
esengine/DeepSeek-Reasonix
DeepSeek-native AI coding agent for your terminal. Engineered around prefix-cache stability — leave it running.
-
volcengine/OpenViking
Self-evolving Context Database for AI Agents. Unify Agent Memory, Knowledge RAG and Skills.
-
Molunerfinn/PicGo
:rocket: The Ultimate Image Uploader for Efficient Creators. Supports Obsidian, Typora, VS Code etc. and 60+ image hosti…