Connection
For a different port or connection name, use matching options:
Tool scope
An external connection can manage workspaces, projects, documents, tasks, execution and results through the public tools. Pass the workspace and record identifiers requested by each tool’s schema. The server’s tool descriptions provide the current arguments and response expectations. Built-in Coordinators manage their assigned workspace and its member projects, plus explicitly service-wide harness, notification and update settings. They cannot switch workspaces or start and stop their own conversation. External clients can create workspaces and control the built-in Coordinator. Workers and Reviewers keep tools scoped to their assigned execution. Saved application settings and workflows are available through MCP. Appearance, unsent drafts, file pickers and browser/OS notification permission belong to the browser. A Coordinator setting that requires a fresh native session must be changed after the active turn stops, through the GUI or an external agent. Permission and cleanup tools relay your explicit decisions for the exact request. They do not let an agent infer approval, approve its own permission request or assume a process has stopped. For updates, use the current revision returned by the relevant read. If another edit changes the record, read it again and reconcile your intended change. Code approval binds to the exact result and candidate being reviewed.Configure and recover work
Discover tools and their schemas through MCP; the server supplies current arguments and constraints.
Changing settings does not authorize enabling a paused queue or delivering code. Native permissions and saved-check grants require your explicit authorization; code approval binds to the exact result and candidate.
Read documents
Use the revision returned by a document read or search when requesting its text.
get_document_text accepts:
Outline results identify heading ranges. Read a section using its
offset and end_offset, then follow the returned continuation until the section is complete. Offsets count Unicode characters.
search_documents returns next_after when more document pages remain, including when the current page has no matches. For text reads, copy next_request or continue with next_offset on the same revision. Workers and Reviewers navigate their assigned sources through read_context; required sources must be read fully.
Read execution measurements
get_context_metrics reads numeric prompt, tool and output sizes for one execution. Supply workspace_id, source and source_id from the metrics evidence, then follow next_after for additional records.
These values measure UTF-8 payload bytes. Use them to investigate large tool responses, repeated reads and failed requests. Use Metrics for token consumption and estimated cost.
Session instructions, messages and tool outputs are separate measurements. Native output may contain Flowfield tool results, so adding the categories can double-count content. Read the returned coverage information when comparing executions: it identifies unavailable measurements, shortened previews and event limits.