Custom Assistants, Projects & MCP tools
Specialise Jan with a reusable Assistant and external tools
Beyond plain chat, Jan lets you organise and specialise your work. Custom Assistants let you save a model plus a set of instructions as a reusable persona; Projects group related threads; and Agents can act more autonomously. Jan also supports the Model Context Protocol (MCP), a standard way to give a model access to external tools so it can do agentic work rather than only answer questions. For research, this turns Jan from a chat box into a configurable, private workbench.
- 1 Create a custom Assistant in Jan: give it a name, choose a model, and write a system instruction such as:
You are a careful bioinformatics methods reviewer. Always ask for the organism, sample size, and sequencing platform before advising. - 2 Start a thread with that Assistant and give it a task from your work. Notice it behaves according to your saved instruction every time, without re-typing it.
- 3 Group related threads into a Project so a multi-week analysis stays organised in one place.
- 4 Explore MCP in Jan's settings: enable an MCP integration to give a model access to an external tool, so it can take actions (agentic use) rather than only generating text.
- 5 Reflect: an Assistant is a saved expert, a Project is a folder of work, and MCP is how the model reaches the outside world — all running through your own machine.
You created a reusable custom Assistant with its own instruction and ran at least one thread through it; optionally you enabled an MCP tool.
Build a specialised Assistant for a recurring research task
A well-written Assistant instruction is the difference between generic answers and answers shaped to your field. Build one you will actually reuse.
Create one custom Assistant tuned to a task you repeat (e.g. a protocol reviewer, a figure-caption writer, or a plain-language summariser) and run two real inputs through it.
- 1 Decide on the Assistant's job and write a precise system instruction for it, including what it should ask for before answering.
- 2 Save it as a named Assistant with a model of your choice.
- 3 Run two genuine inputs from your work through the Assistant.
- 4 Refine the instruction once based on what it got wrong, and note the improvement.
A saved Assistant, its final instruction text, and two example outputs showing it behaving as specialised.