Give your agents a place to do lasting work.

What makes a workspace useful to an agent?

An agent needs more than a prompt when a job spans several conversations. It needs to know which evidence is current, what the user authorized, where to save its work, and what remains unfinished. A useful workspace makes those answers recoverable from files and task records that a person can inspect.

In Clanker, a Project groups related files and instructions. A Session records a particular conversation and its work. An Agent supplies a role, method, and selected capabilities. These parts let you preserve an engagement or research effort while changing the assistant or the next task. They do not remove the need to check an answer.

Choose Clanker for a complete work cycle

Consider Clanker when the job should produce a durable decision brief, source table, document, spreadsheet, or App, and someone needs to return to that work. For example, an agent can compare an approved requirements file with supplier proposals, save the evidence and decision, then leave an unresolved-questions note for a later review.

The concrete benefit is the assembled workflow: persistent Project material, a reusable Agent, selected Tools, a Session record, and visible saved outputs. Evaluate that combination on your own task. A chat answer, a generated file, and an accepted business decision are separate milestones; identify which one your workflow actually needs.

Keep a compatible assistant and connect its tools

Claude, ChatGPT, OpenClaw, Hermes, and other compatible clients can reach hosted Workspace through the supported MCP connection. Client version, account access, organization policy, and approved permissions determine whether the route works. First ask the client to find the intended Project and read one known file.

The external assistant keeps its own reasoning and provider usage. Workspace supplies the tools exposed to that connection. MCP and the separately authenticated HTTP API have different coverage; neither implies every native Workspace tool is available. Selecting a Project identifies task context, while the OAuth grant can cover the connected user’s accessible Workspace.

Sources: Clanker: external-agent access, permissions, and managed Sessions

Separate lasting work from active compute

Saved Project context is different from a running machine. Ordinary writing and file tasks can start in a browser Session. A task needing scripts, a website interaction, or a specific local application may need a cloud environment or an explicitly connected computer. Choose the environment after identifying that dependency.

A saved BYOA registration is not a permanently running process. Supported cloud agent tasks use bounded compute, and successfully saved state can be restored for later work. Unsaved process state can be lost. A standalone Sandbox API runtime has its own identity, expiry, and storage rules; it is not the same resource as your Workspace Project.

Use the workspace to build a reusable specialist

Once a task produces a result you can check, turn its stable method into a custom Agent. Define the role, accepted inputs, source rules, selected capabilities, exception behavior, and output checks. Keep changing client or project evidence with the Project rather than embedding every temporary fact into the reusable role.

This is the practical agent-factory path: specify a job, configure a specialist, test it with normal and incomplete inputs, and inspect its saved work. Build several narrow roles only when each has a distinct responsibility. A researcher, reviewer, and formatter should exchange named artifacts rather than rely on an assumed shared conversation.

Prove the handoff before expanding the workflow

Start with a small approved packet and one deliverable. Open the saved file in Preview or Files, verify consequential statements, then ask for one correction. Finish with a handoff that identifies source versions, current outputs, unresolved questions, and the next permitted step. Read that handoff in a fresh task.

Add recurring execution only after the task and input refresh process work. Native Loops and Schedules have their own settings; connecting an assistant does not create one. Review actual usage, stop unneeded environments, and confirm the exact data arrangement for sensitive work. A persistent workspace is useful when the next task can resume from checked evidence.

Try it with a small example.

Before you use the result.

• The Project contains the actual approved sources and checked output.
• The active client can use the tools required for this task.
• A later task can resume from the saved handoff.
• Compute lifetime, access, and usage are understood separately.