EverFern is open source — self-host in minutes. Star on GitHub ↗

The free, open-source alternative to Claude Co-work & Manus Desktop

Computer use, browser automation, document processing, and multi-agent workflows — all running locally on your machine. No subscriptions. No data leaving your device. Free forever.

Download for Windows

Why EverFern

Claude Co-work costs $20/mo.
Manus Desktop costs $39/mo.
EverFern costs nothing.

Same capabilities. Runs locally. MIT licensed. Your data never touches a third-party server unless you want it to.

Computer use

Sees your screen.
Operates your apps.

No integrations. No APIs. EverFern moves the mouse, clicks buttons, reads UI, fills forms — across any desktop app exactly like a human.

everfern — computer-use

# task: "Create Q3 report in Notion"

→ screenshot() // reading screen state

→ click(342, 198) // "New page" button

→ type("Q3 Report")

→ select("Weekly template")

✓ Page created

→ navigate("database view")

See it in action

Skills marketplace

Install a skill.
Extend anything.

📄invoice-parser

🌐navis-browser

📊sheet-writer

📧email-triage

🔔slack-notify

🗂file-organiser

Self-hosted & private

Your data never
leaves your machine.

Cloud requiredNever

Data ownership100% yours

Monthly fee$0 — free forever

LicenseMIT open source

Linux VM sandbox

Shell access.
Zero risk to your system.

everfern — linux-vm

$ npm run build

⚠ Shell execution requested

→ Awaiting user approval

✓ Approved — running in sandbox

[VM] Build complete in 4.2s

Cost visibility

Real-time token &
cost tracking.

Token usage stored 100% locally

Per-session cost breakdown

Works with OpenRouter & all providers

Multi-provider

10+ AI providers.
One interface.

OpenAI (GPT-4o)Connected

Anthropic ClaudeConnected

Ollama (local)Private

OpenRouterConnected

100% free — forever

Everything Claude Co-work does.
Without the $20/mo price tag.

Download free →

Why EverFern

Not another
SaaS agent tool.

While Claude Co-work and Manas Desktop are paid commercial tools that lock your data in the cloud, EverFern is 100% free, open source, and runs entirely on your machine.

The others

Commercial & paid

Data sent to their servers

Monthly subscription fees

Closed-source, no audits

Limited customisation

Vendor lock-in

EverFern

Open, local-first

Runs 100% on your machine

Free and open source (MIT)

Fully auditable codebase

Build and share your own skills

You own your data, always

Get started

In action

From tasks to results,
handled by AI agents

Describe what you want in plain English. EverFern finds the right skill, takes control, and reports back — on your machine, privately.

Get started

Chat

Terminal

Browser

+ New Chat

Fix layout spacing

Generate CSV parser

Automate email sync

Audit system logs

Can you search for python tutorials, read the first link, and write a quick script to test tool calling?

I'll start by opening a browser to search for python tutorials, extract the content, and then execute a python script locally to verify the results.

🌐 BrowserSearching: python tutorials

🌐 BrowserNavigated to python.org

⚡ SystemExecuting: python3 test_tool.py

✓ All tests completed!

Ask EverFern...

Peer Agents

Agents that argue.
Until they find the best answer.

EverFern's Peer Agent mode pits specialized AI agents against each other in a structured debate. They critique, challenge, and refine ideas until one solution wins.

Debate

Summary

Verdict

3 agents active

⚡

Optimizer

Performance

🏗️

Architect

Structure

🎯

Pragmatist

Delivery

TopicHow should EverFern handle multi-step task failures — retry, reroute, or escalate?

⚡

Optimizerjust now

Automatic retry with exponential backoff handles 80% of transient failures without user intervention. We should implement a 3-strike rule with jittered delays: 2s, 8s, 30s. Anything that fails 3 times gets flagged for escalation. This keeps the system fast without overloading the user with manual decisions.

🏗️

Architect

⚠ Challenging

Blind retry misses the root cause. If a tool failure is due to a missing API key or a changed DOM element, retrying 3 times wastes 40 seconds and delivers the same error. We need semantic failure classification first — is it transient (network), structural (wrong selector), or logical (wrong understanding)? Only transient failures should retry; structural ones should reroute to an alternative approach.

🎯

Pragmatist

✓ Proposed synthesis

Both valid. Ship a hybrid: classify failure type via a lightweight error taxonomy, then branch. Transient → retry with Optimizer's backoff. Structural → reroute with a fallback tool or DOM strategy. Logical → pause and ask the user for clarification rather than spinning in loops. This covers all cases without over-engineering retry or ignoring root causes.

✦

Consensus Verdict

Classify failure type first. Retry transient errors (max 3×), reroute structural errors, escalate logical errors to the user. Ship FailureRouter v1 before next sprint.

3–5

Avg. rounds to consensus

Unlimited

Agent types supported

≥ 80%

Verdict confidence threshold

Everything included

Every tool you need.
Zero compromises.

No technical knowledge required. Describe what you want in plain English — EverFern figures out the rest.

Computer Use

Sees your screen, moves the mouse, clicks, types, and navigates any app exactly like a human would. Works with any desktop application — no integrations needed.

Navis — Built-in Browser Agent

Navigate websites, fill forms, scrape data, and interact with web apps in plain English. Powered by our open-source Navis extension.

Document Processing

Read, analyze, and create PDFs, Word docs, Excel sheets, PowerPoints, and CSVs. Works with your existing files.

Code Assistant

Write, review, debug, and refactor code with full project context. Works inside your actual editor.

Skills System

Reusable task modules you can install from the community or build yourself. Shareable, auditable, and yours.

Workflow Builder

Chain actions together, save them, trigger on a schedule. Automate anything you do repeatedly.

Linux VM Execution

Shell commands run in an isolated sandbox so nothing can accidentally break your system.

Peer Agent Debate

Multiple specialized agents debate the best solution before anything gets executed. Plans are stress-tested before they touch your machine.

20+ Built-in Tools

Desktop, Browser, Files, Code, Data, and System tools all built in. Plus custom tools via MCP — connect anything.

Multi-Provider Support

Use local models (Ollama, LMStudio) or connect to 10+ cloud providers — OpenAI, Anthropic, DeepSeek, Gemini, OpenRouter, Groq, and more.

Integrations

Works with the tools
your team already uses

Unlike paid, cloud-locked alternatives, your data stays safe on your machine with EverFern. Use free online AI providers or connect high-tier models — your choice, your control.

Browse all integrations