Choose who does the reasoning and where the work runs.

Use a native Agent for a Workspace role

A native Agent is the assistant you configure in My Agents with its instructions, reference knowledge, Skills, and selected Tools. It suits a recurring role such as proposal review when you want that setup available for future Workspace Sessions. Keep each client’s current inputs in its Project.

Start with a supervised example and inspect the saved file. Reuse the role once its method works. If you later need a hosted Loop or Schedule, configure and test that workflow separately. Creating the Agent alone does not start recurring execution, and a normal Session does not automatically have access to a computer.

Use MCP when you want to keep an external assistant

A compatible Claude, ChatGPT, OpenClaw, or Hermes client can connect to hosted Workspace MCP through browser authorization. The external assistant supplies its own reasoning and uses the Workspace tools you permit. This is useful when you already prefer its conversation and want Project sources and deliverables in Clanker.

The private external conversation is not copied automatically. Save relevant decisions, results, and a handoff deliberately. Changing tools use the managed Session workflow, which records public progress and the checked result. A selected Project names the task context; it does not narrow the OAuth grant to that Project. Review the connected user’s accessible Workspace and granted permissions.

Use BYOA for a supported harness with its own runtime

Where available, BYOA lets you deploy supported cloud agents or connect a compatible enrolled computer or your own agent host. The harness retains its own runtime. The documented preview supports hosted OpenClaw, Hermes, and Codex; hosted Claude Code installation is unavailable. Check the current picker and setup guide before choosing.

BYOA hosting is bounded task compute, with successful private saves used to restore supported state. It is not a promise of a process running forever. An existing host needs the supported adapter, and a connected computer must be approved and online. Direct Cloud Browser/Desktop control and some native Workspace tools are outside the documented BYOA coverage.

Compare the operation you need, not the label

Write the required steps: read sources, create a workbook, interact with a website, run a harness, or repeat a brief at set times. Confirm the current tools and environment for each. A remote MCP connection, BYOA deployment, and Sandbox API runtime are different interfaces; do not infer feature parity from the word Agent.

If a task requires a specific connected computer, verify that device’s availability. If it requires recurring native work, test the Loop or Schedule path. Regular Channels, Loops, and Schedules do not run native BYOA harnesses in the documented preview. Choose the environment that supports the required operation, then test the actual result.

Separate inference from hosting and tools

Model calls made by an external assistant remain with its provider. Clanker-hosted model work and Workspace compute use Clanker credits. Reading MCP tools or recording a message does not itself start a Clanker model conversation. This makes a source-reading task different from asking Clanker to run paid compute or generation.

BYOA hosting and inference are independent. Supported Hermes and OpenClaw configurations can use Clanker inference where offered; Codex and Claude Code use their provider inference in the documented preview. Review setup time, active task compute, selected paid tools, and model usage before estimating the cost of the whole workflow.

Keep the trial small enough to compare

Use the same fictional packet and acceptance criteria in the surfaces you are considering. Record setup steps, unavailable operations, saved output quality, correction effort, and observed usage. If your existing assistant already completes the task well, the reason to add Workspace should be a concrete benefit in saved work, review, reuse, or operation.

Whichever surface you select, save a handoff identifying current inputs, output locations, unresolved questions, and the next authorized step. A later assistant can read that deliberate record. Recheck the choice when the workflow changes; success at file drafting does not establish website control, persistent hosting, or unattended execution.

Try it with a small example.

Before you use the result.

• The choice names who reasons and where execution happens.
• Every required operation has a verified supported surface.
• External conversations and saved Workspace context are distinguished.
• Hosting, inference, tool charges, and recurrence are reviewed separately.
• The same small task can be inspected before extending the workflow.