Overview
Jan Agent is an AI agent that lives in your terminal. You give it a task in plain language, and it
reads files, edits them, runs commands, and searches the web until the work is done. Plain jan
auto-approves tool calls; use jan --safe to ask for approval before writes, shell commands, and
MCP tool calls. Approval and shell sandboxing are separate controls.
It can inherit enabled HTTP providers configured in Jan Desktop, provided their endpoints and required credentials are usable. To use a Desktop local model, start its Local API Server and configure Jan Agent to use that endpoint.
janThat single command opens the interactive console in whatever directory you're in.
Jan Agent and Jan Desktop
They're two front ends onto the same stack, built for different jobs. Jan Desktop is where you run models on your own hardware and work in a GUI. Jan Agent is where you point a model at a folder and let it work.
| Jan Desktop | Jan Agent | |
|---|---|---|
| Interface | Desktop app | Terminal, interactive or headless |
| Where models run | Locally via llama.cpp or MLX, or a cloud provider | Cloud providers, or any OpenAI-compatible endpoint |
| Works on | Conversations, projects, and files you attach | A project folder on disk |
| Files and shell | Through MCP servers you add | Built in - read, write, edit, and run commands, auto-approved by default; use --safe for approval prompts |
| MCP servers | Yes | Yes, the same ones |
| Skills and memory | Desktop's own agent workspace | Jan Agent's own user and project stores (separate default stores) |
Neither replaces the other. A common setup is Jan Desktop serving a local model, with Jan Agent pointed at it - see running against local models.
Jan Agent can inherit eligible Desktop HTTP provider settings and MCP servers, but
skills and memory are not automatically shared. Desktop keeps its
permanent agent store in <data-folder>/agent-workspace; Jan Agent uses user and project stores
under ~/.jan. Cowork can also read skills from an attached project’s Jan Agent store; this does
not merge the permanent stores or share memory.
Install
Jan Agent is a preview. The installers below pull from the agent-nightly channel - expect
nightly-quality builds.
curl -fsSL https://delta.jan.ai/jan-cli/install-jan-agent.sh | bashirm https://delta.jan.ai/jan-cli/install-jan-agent.ps1 | iexInstalls to %USERPROFILE%\.local\bin. To also add it to your user PATH:
& ([scriptblock]::Create((irm https://delta.jan.ai/jan-cli/install-jan-agent.ps1))) -AddToPathThen check it landed:
jan --versionThe Quickstart covers connecting a provider and finishing a first task. Looking for the desktop app instead? Download Jan Desktop.
What it can do
- Read and change a project - open files, make targeted edits, create new ones.
- Run commands - build, test, grep, and read the output to decide what to do next.
- Search the web - look things up mid-task when the answer isn't in your project.
- Use your MCP servers - every connector configured in Jan Desktop is available here too.
- Follow your project's skills - reusable procedures it loads when a task calls for them.
- Remember across sessions - durable facts about the project, kept as files you can review.
- Delegate - split large jobs across subagents that work in parallel.
- Keep working toward a goal -
/goalkeeps it going until a condition you set is true.
Jan Agent ships no inference engine of its own, so a model always runs somewhere else. That can be a
cloud provider, or your own machine - point it at any OpenAI-compatible endpoint, such as Jan
Desktop's local API server, with jan config set --base-url.
Get started
Install Jan Agent, connect a provider, and finish your first task.
Slash commands, keyboard shortcuts, and how a session works.
What the agent can do, and how approval keeps you in control.
Providers, agent.toml, and per-project instructions.
Delegate parts of a task to agents that run in parallel.
Every flag and subcommand, including headless runs for scripts and CI.
Staying in control
The agent edits files and runs commands without stopping to ask. File paths are checked against the
project root, and anything in [tools] deny is refused outright - but
shell commands are not sandboxed by default, so one runs with your own access.
Three ways to change that:
- Plan mode (
jan --plan) makes the session read-only. The agent investigates and proposes, but cannot change anything. - Safe mode (
jan --safe) prompts for writes, shell commands, and MCP tool calls instead of auto-approving them. Policy allow rules and thread-level grants can allow matching calls without a prompt; dedicated memory and skill tools do not prompt. It does not enable the shell sandbox. - Sandbox (
jan --sandbox) confines shell commands to the project with an OS sandbox. Off by default on the CLI, always on for the desktop chat agent. Combine it with--safeto request approval as well as confinement.
See Tools & permissions for the full picture.