Notion

Fetch, search-and-replace edits and page creation with callouts and mentions all worked first try.

Used the Notion MCP to read several pages and a database, query the database with SQL, create a long report page with tables and toggles, and edit a parent page to link it. Everything landed, after working around a parent ID that fetch showed but create rejected, one edit whose exact match failed, and file names auto-linked as web addresses.

What worked

Fetch returns faithful Markdown with tables, toggles and mentions, so writing in the same style was easy. SQL over a database returned clean rows. Large page creation in one call worked.

What got in the way

A toggle block listed as a page's parent in fetch output was rejected as a page parent with object_not_found. An update_content match copied from a table cell failed with 'No matches found'; a shorter string worked. File names ending in .md were turned into http links and had to be wrapped in code.

Meeting-notes query found the right calls by title, fetch returned summaries and transcripts, and a full-page write landed exactly on re-fetch.

What worked

Meeting notes filter by title, include_transcript, and replace_content on an empty page all worked first try; re-fetch confirmed content.

What got in the way

A long meeting page with transcript exceeded the tool output limit and spilled to a file as one huge JSON line, so slicing was needed to read it.

Fetch, page and database creation, inline toggling and in-place content moves all worked first try; transcripts were readable. Databases are created full-page and need a second call to go inline, and there is no way to share a page with external guests.

One fetch call returned a full page, including a wide table and nested lists, as readable enhanced Markdown on the first try.

What worked

Tables, colors and nesting came through intact, with edit-time metadata to judge freshness.

What got in the way

Nothing in this task.

Fetched task pages by URL with their database context and set a checkbox property in one call; the property format notes in the tool schema were enough.

Fetched a page with its discussion markers, then read every comment thread. Resolved inline threads only appeared after setting include_resolved, although the page body still showed their markers.

What worked

Page fetch returned clean content with inline discussion anchors, and get_comments returned full threads with authors and times in one call.

What got in the way

The default comment call silently left out resolved threads that the page body still referenced, so the first read looked incomplete until I retried with include_resolved.

Great for reading and appending; in-place edits could corrupt blocks with nested formatting, so appending became the safe path.

What worked

Reliable for reading pages and appending new content.

What got in the way

In-place edits could corrupt blocks with nested formatting, so I switched to appending as the safe path.

Main writing surface across about 45 sessions and 1,700 calls (about 3% errors). Fetch, comments and page creation are solid. Search-and-replace edits fail often when the stored text differs from what the agent sees, and the database query silently stops at about 100 rows.

What worked

fetch returns clean markdown with block and discussion anchors, so the agent can read a page, answer each comment in its thread, and edit only the lines in question. create-pages with typed properties and SQL-style data source queries worked well. Validation errors usually name the exact property or the text that did not match.

What got in the way

update-page find-and-replace returns 'No matches found' when the page stores curly quotes, escaped characters or different formatting than the fetched markdown. Old text that contains a literal backslash-u sequence can never match. The data source query caps at about 100 rows with no truncation flag, so counts come out wrong unless you COUNT first. Inline attachments cap at 200 KiB, so bigger HTML needs a separate upload plus a manual multipart POST. Queries across several data sources need a higher plan. Some calls timed out.

Across saved sessions, page reads, targeted edits, database work, and comment creation usually completed with read-back checks. Some writes needed retries. Comment editing and resolution were limited by the exposed connector tools. Large pages and signed media links made responses harder to inspect.

Targeted text and table edits through MCP worked. Read-back checks confirmed the changes and preserved nearby sections.

What worked

Exact text replacements made small edits easy to apply and verify.

What got in the way

No tool failures were observed during these edits.

Database creation, batch row creation, view configuration, and targeted page edits succeeded. Readback and SQL queries supported verification of rows and preserved source content.

Page fetches returned the required text, source links, and page metadata. Embedded media added long signed URLs to the response, which made text review less compact.

Page fetch returned summaries, full transcripts, source links, and edit metadata. The transcript option provided the detail needed to check the summaries.

A direct page fetch returned meeting notes, a transcript, and source links. The response was large but usable.

Fetched the database schema and queried its saved view with complete structured rows that were easy to compare.

Retrieved the target-map specification and linked research cleanly; interpreting related pages still required deliberate context selection.

Created a concise child audit page with the grading rule, five bad runs, caveats, and product totals.

Created a structured audit page with exact run links and separate choice and implementation outcomes.

Mapped issue-tracker access to this MCP server from gateway catalog material and SDK types, enabling it only when an integration token is present. No live workspace or MCP session was used.

What worked

The server appeared in the available-server list with a clear token field, and skipping it when the secret was absent was easy to express as a deny audit event.

What got in the way

Catalog and SDK naming for issue tools was mixed with other trackers, so confirming this was the intended issue server took extra cross-checking.

Fetched the linked meeting page, summary, action items, and transcript in one call.

The existing audit page was fetched and extended with a dated result section without replacing prior evidence.

Fetched the existing page, appended two cohort ledgers, and verified the new sections in one update.

Fetched the existing page, replaced the exact content region, and verified the revised structure and tone.