Everfern: Advanced

This is the seventh of eight pages for Everfern. It covers six power features that go past the first weeks of use, each with what it is for, how to turn it on, and what it trades away. Then it points to the shared advanced pages that apply to Everfern.

Everfern's README names most of these features but does not explain how to use them. So the labels below come from its source code on GitHub, read on 14 November 2026. The labels were checked in the code at the v1.3.0 tag, the version you download. This guide did not run the app to check them. Try each one only after the jobs in the playbook feel routine.

The Linux VM for commands

What it is for. The README says every shell command runs in an isolated Linux VM, so that "nothing touches your real system by accident" when a command goes wrong. Everfern uses it for the work behind documents and code. The Windows settings text lists "PDF generation, spreadsheets, and code execution" among its jobs.

How to turn it on. During setup, the code offers Install WSL 2 on Windows or Get Docker Desktop on a Mac. You can come back later through the settings section Linux VM, which shows Install Linux VM on Windows and Set up Linux VM on a Mac. The same section has a Skill Toolchain & Dependencies card for the Python setup in ~/.everfern/venv.

As page 3 explains, the Mac container shares your whole /Users folder, and WSL on Windows normally sees your C: drive. So this VM separates commands from your system without hiding your files from them.

• You gain: Commands run in a separate Linux, so a bad install or a broken script lands there rather than in Windows or macOS. Without the VM, the code blocks most commands, so many document jobs need it.
• You risk: Your own files are still within reach of those commands through /host/Users or /mnt/c, so a mistaken delete in the VM can still remove a real document. On Windows, WSL also adds a whole Linux system that you have to remove separately, as page 6 shows.
• Makes sense when: you want Everfern's document jobs to work at all, and you keep giving it only the everfern-tryout folder. Treat the VM as a tidier place for commands, not as the wall that Agent Zero's container is.

Auto-approval rules

What it is for. Each time Everfern wants to run a command on your real computer, it shows a card with Deny, Always allow, and Allow once. Cards for other risky actions say Reject and add Allow prefix, which saves a rule for every command that starts the same way. A saved rule skips that card for matching commands. On a long job that keeps running the same harmless command, that saves many clicks.

How to turn it on. Clicking Always allow or Allow prefix on a card saves a rule. In the settings section Tool Permissions, you can also click Add New Auto-Approval Rule and fill in Tool Name, Match Type, and Pattern / Command Base. Match Type is either Exact Tool/Command Name or Prefix Match (e.g. npm, git). The rules are saved in ~/.everfern/tool-approvals.json, and Revoke All clears them.

The code does try to protect prefix rules. It refuses to auto-approve a command that chains a second command on, such as ls; rm -rf ~ under an ls rule, but this guide did not test that check.

• You gain: Fewer interruptions on a job you have already watched, such as one that runs the same safe command many times.
• You risk: Every rule runs with your own permissions on your real computer, and nobody is asked again. A broad prefix such as npm or git covers many commands you have never seen, including ones that change or delete files.
• Makes sense when: you have watched the exact same command many times, it only reads or only writes inside everfern-tryout, and you choose Exact Tool/Command Name rather than a prefix. It does not make sense for a beginner, so keep this list empty in your first weeks.

Custom skills

What it is for. The README says Everfern turns what it learns into "saved, reusable instructions" that speed up later jobs. You can also write a skill yourself, such as a fixed way to format your weekly notes, so you do not repeat the same instructions in every chat.

How to turn it on. In the settings section Custom Skills, click Create New Skill and fill in Skill Name, Description, and SKILL.md Content (Optional). That section says custom skills are stored in ~/.everfern/custom_skills/. The Privacy & Data section lists ~/.everfern/skills/ instead, so this guide could not confirm which folder your copy uses. The website also shows skills to install, such as email-triage and slack-notify, but this guide could not confirm the steps for installing one.

• You gain: Repeat jobs become shorter to ask for and more consistent each time.
• You risk: Everfern follows a skill without you retyping it, so a careless line in it, or in one you install from someone else, repeats in every job that uses it. Skills that touch email, Slack, or money let the agent act as you.
• Makes sense when: you write the skill yourself, it covers a job you have already done by hand, and it touches only files in everfern-tryout. Read any installed skill's text before you use it.

Scheduled tasks

What it is for. A scheduled task runs a prompt by itself at set times, such as a summary of new notes every morning. The pricing page says the Free plan includes "Background schedules and cron triggers."

How to turn it on. The code has a Scheduled Tasks panel, and when it is empty, it tells you to click the plus icon to create a task. The plus icon opens a form with Task Name, Description, AI Instructions, Repeat Pattern, Starts At / Run Time, and Expires At, and Schedule Task saves it. You can also ask in the chat, because the code gives the agent its own tool for creating tasks. That tool takes patterns such as "every X minutes," "every X hours," and "daily at HH:mm" for its timing. This guide could not confirm where the panel opens from, or whether Everfern has to stay open for a task to run.

Everfern's built-in instructions tell the model to confirm before "Interactive desktop control outside a scheduled task." In other words, a scheduled task is excused from that question.

• You gain: A job you trust repeats without you, and Expires At lets it end by itself.
• You risk: Nobody is watching, and the model is told it may control the desktop inside a scheduled task without asking first. The agent can also create a task from a chat, so a misread request could leave a job repeating every few minutes.
• Makes sense when: the job has worked by hand at least three times, it writes only its own notes inside everfern-tryout, and its prompt forbids computer use and the web. Set Expires At on every task, and check the Scheduled Tasks list each week.

MCP servers

What it is for. An MCP server is an add-on program that gives an agent new tools, such as reading a calendar or a database. The README's architecture diagram lists MCP among Everfern's tools, and the first pre-release, in April 2026, listed "Built-in MCP compatibility" as a feature.

How to turn it on. This guide found no documented way, and no settings screen, for adding an MCP server by hand. Instead, the code gives the agent a tool named connect_mcp_server. Its description reads "Dynamically connect to an MCP server using a command or docker image," so the agent starts the server itself. A second tool, search_mcp_registry, suggests servers for Gmail, Slack, Google Drive, GitHub, PostgreSQL, and Spotify. The settings section Registered Tools lists the agent's tools. This guide could not confirm that a connected server appears there. The v1.3.0 code for connect_mcp_server has no approval step, so the agent can start a server program on your computer without a card. Add "Do not connect any MCP server" to prompts unless you want one.

• You gain: Everfern can work with apps through their own connections instead of clicking around their screens, which is faster and less likely to misclick.
• You risk: The code starts the server as a program on your real computer, not in the VM, so the server can do anything your account can do. Servers for Gmail or Slack let the agent read and send as you, and a sent message usually cannot be undone.
• Makes sense when: you know exactly which server you are adding, it comes from the app's own maker, and it can only read. Skip anything for email, chat, or money until you have used Everfern for months.

To compare MCP servers with skills, and to see how other agents add them, read Skills and MCP.

Build from source, and the unfinished terminal version

What it is for. Building from source lets you run the newest code, which on 14 November 2026 was ahead of the v1.3.0 download. It also lets you check what Everfern really does before you trust it.

How to turn it on. The README says you need Node.js v18+ on Windows 10/11, macOS, or Linux, and then lists these steps:

git clone https://github.com/Everfern-AI/Everfern.git
cd Everfern
npm install
npm run dev

The repository also has a terminal version in its cli folder. Its commands include node src/cli.js status and node src/cli.js doctor. Its own README marks "Interactive input handling" and "Agent integration" as not done yet, so it is not a working way to give Everfern jobs.

• You gain: The newest fixes before they reach a release, and a way to read the code that the README's safety claims depend on.
• You risk: Code on main has had no release testing, and npm install downloads and runs many packages from other people. This guide could not confirm whether a build from source keeps its data apart from the desktop app's ~/.everfern folder. If it does not, a bug in it could reach your saved chats and settings.
• Makes sense when: you already use Node.js, you have backed up or reset ~/.everfern, and you want to test a fix rather than do real work. For everyday use, stay on the release download.

Not covered: chat apps and Dispatch

The first pre-release, in April 2026, listed "Seamless integration with Discord and Telegram" as a feature, and the source code still has parts for both. The current setup code marks integrations "Coming Soon," though, and the README describes a desktop app only. The settings code also lists a section named EverFern Dispatch marked Beta. This guide could not confirm that any of these work in v1.3.0, so treat Everfern as having no phone or chat-app control.

Shared advanced topics

These shared pages apply to Everfern:

• Local models: Everfern runs Ollama or LM Studio, and computer use needs a local model that can read pictures.
• Mixing models: Everfern's settings code has a separate Vision Grounding section, so the model that reads your screen can differ from the one you chat with.
• Host access and unattended work: computer use always acts on your real computer, and scheduled tasks run with nobody watching.
• An always-on machine: scheduled tasks can only run when the computer is on.
• Relays: OpenRouter and EverFern Cloud both sit between Everfern and the company that runs the model, and the pricing page calls the second one "EverFern Cloud routing."

Sources

Opened 2026-11-14. Source-code links point to the v1.3.0 tag.

• Everfern README, for the VM claim, reusable skills, MCP in the architecture diagram, the build-from-source steps, and Node.js v18+.
• Everfern release list, for the April 2026 pre-release with MCP, Discord, and Telegram.
• Everfern pricing, for the Free plan's schedules and "EverFern Cloud routing."
• Everfern website, for skills such as email-triage and slack-notify.
• Settings screen source, for Linux VM, Install Linux VM, Set up Linux VM, and Skill Toolchain & Dependencies. The same file has Tool Permissions, Add New Auto-Approval Rule, Custom Skills, and Create New Skill. It also has Registered Tools, EverFern Dispatch, and Vision Grounding.
• Linux VM setup screen source, for Install WSL 2 and Get Docker Desktop.
• Approval rules store, for ~/.everfern/tool-approvals.json and the chained-command check.
• Host-command approval card, for Deny, Always allow, Allow once, Reject, and Allow prefix.
• Scheduled Tasks panel and scheduled task form, for the panel and form labels.
• Scheduled-task tool for the agent, for the repeat patterns.
• Built-in permission instructions to the model, for the scheduled-task exception.
• MCP registry tool and MCP code, for connect_mcp_server, search_mcp_registry, servers started as local programs, and no approval step before connecting.
• Everfern CLI README, for the commands, the unfinished items, and the shared ~/.everfern/ folder.
• Setup screen source, for integrations marked "Coming Soon."

Next

Continue to Everfern: Persona and rules.