Skip to main content
Repository context gives your agent manual site inputs, saved search observations and bounded public site facts. These eight tools run on a separate local stdio MCP server. The hosted HTTP MCP endpoint cannot read your repository files.

​
Start the local host

Use an authorized checkout with dependencies installed and Node.js 22 or later. Initialize the ledger in the customer repository, then configure your MCP client to launch:
Set the executable to node and supply the remaining values as arguments in clients that use a command/argument form. The root is fixed at startup. Tool arguments cannot move it; configured paths and imports must remain inside that boundary, including through symlinks.

​
Start from manual context

Call context_status, then context_manual. manual_only is a complete configuration. Record target pages, your offering, assets and approved competitors through the repository’s manual inputs. Manual statements remain judgments with their own source label. Use context_focus to select focus pages. It considers manual targets before saved search clicks. For one page, context_gsc_page reads existing fact rows without making a Google request.

​
Add dated facts

context_gsc_import accepts a supported connector handoff inside the repository. Preserve its retrieval identity, aggregation and date window. Missing rows in a truncated result never mean zero impressions. A direct GSC refresh needs its own read-only Google access and obeys bounded request limits. context_profile_build fetches a bounded public site profile. Unreadable pages stay explicit. context_profile_check checks whether citations in the authored profile resolve to manual inputs or fact rows. Use the facts in a campaign brief, keeping observations separate from your agent’s conclusions. The concepts guide explains ownership, and troubleshooting covers missing grants and path refusals. Read local context results for partial refreshes, nullable metrics and missing-file responses.