Gong records every call your team runs and analyzes each one. What you cannot easily do is read across all of them at once. The questions that matter live in the aggregate: which objection came up in nineteen calls last quarter, which competitor gets named right before a deal stalls, what the six closed-won calls said that the twelve closed-lost ones did not. Answering those means opening calls one at a time, or filing a ticket for whoever owns your warehouse.
Scratch pulls the corpus down to files on your computer. Calls with their full analysis, transcripts, users, and library structure all become local records your agent reads together, about 10x faster than the same agent working over an API, because it reads the files directly instead of making a request per call. Gong is read-only in Scratch and always will be: Gong's API has no update surface for any of this data, so nothing your agent does can reach your call history.
What Scratch pulls from Gong
- Calls, with every analysis block Gong produced: brief, topics, key points, outline, highlights, trackers, and outcome
- Call participants, embedded on the call record
- Call transcripts, rendered as speaker-labeled subtitle files with timestamps
- Public comments left on a call
- Users, with the reporting line between them
- Workspaces, and your library folder tree
- Scorecard definitions
How it works
- Scratch pulls your Gong data into local files. One file per call, one per transcript, plus users, workspaces, and library folders. Calls arrive as Gong returns them, with the analysis blocks intact rather than summarized. This is read-only. Nothing touches your Gong instance.
- Your agent reads across the whole corpus. Open the folder in the agent you already use. It reads every call and transcript together, so a question about patterns across two hundred calls is one prompt instead of two hundred readings.
- You keep the output as files, or export the corpus somewhere. Analyses, summaries, and tallies sit next to the calls they came from. If you want the corpus in a database or a workspace tool, Scratch exports it: calls, transcripts, and analysis text into Supabase, Airtable, or Notion, with the links between calls, hosts, and folders preserved as real relations.
What teams use it for
- Find the objections that recur across a quarter of calls, with the exact quotes attached.
- Track how often a competitor is named, and what gets said in the same breath.
- Compare what closed-won calls covered against what closed-lost calls covered.
- Pull the language customers use for a problem, verbatim, into positioning and site copy.
- Build a searchable transcript corpus in your own database instead of paging an API.
- Feed real call context into an AI workflow without giving that AI access to your Gong instance.
Why pull it into files at all?
Because the interesting questions are about the corpus, not a call. An agent working over the API reads a handful of calls per request and never holds the set. An agent working over files holds all of them, reasons across them, and cites the call it got each finding from. The analysis runs locally, so it is fast, and your Gong instance is never written to.
Exporting is the other half. A transcript corpus in Postgres or Airtable is queryable by everyone on your team, joinable against your CRM, and yours to keep.
What's safe
Nothing writes back, and this is not a roadmap item. Gong's API exposes no way to edit calls, transcripts, users, or analysis, so every table Scratch pulls is read-only by design: write operations are disabled rather than merely discouraged. Your agent reads a copy on your machine and can never alter your call history. You bring your own AI: Scratch holds no AI credentials and runs no model, so you sign into Claude, Claude Code, Codex, Cursor, Copilot, Cline, or Windsurf the way you already do. The workspace is git-backed, so every pull and every analysis is tracked. Scratch is SOC 2 compliant.
Questions Gong users ask
Gong already has AI. Why pull the data out?
Gong's AI answers questions inside Gong, and the answers stay there. Pulling the corpus down gives the agent you already run the raw material: every transcript in one pass, findings cited back to the calls they came from, output kept as files, and the freedom to join call data with whatever else is on your machine, a CRM export, a positioning draft, last quarter's analysis.
Can Scratch change anything in Gong?
No, and it never will. Gong's API has no update path for calls, transcripts, users, or analysis, so read-only is a property of the service rather than a stage Scratch is working through. Create, update, and delete are disabled on every Gong table in Scratch.
Does it pull the AI analysis, or just the transcript?
Both, and they are separate records. The call record carries the analysis blocks Gong generated: brief, topics, key points, outline, highlights, trackers, and outcome, in the shape Gong returns them. The transcript is its own record, one per call, rendered as a speaker-labeled subtitle file so timestamps and turn-taking survive.
What do transcripts look like as files?
Speaker-attributed SubRip, the subtitle format: numbered cues, start and end timestamps, and a Speaker N: label on each turn. That keeps the diarization Gong produced, so an agent can tell who said what and when, and a human can read it without special tooling.
Can I get the corpus into my warehouse?
Yes. Scratch exports Gong tables to Supabase or Postgres, Airtable, and Notion, with foreign keys resolved: a call links to its host, a library folder links to its parent, and the analysis text lands as readable columns.
How does it connect?
An API key from your Gong instance: an access key and secret, generated by a Gong admin under API settings. Scratch uses Gong's own rate limit and retries when it is hit, so a large pull runs unattended.
Do I need to be an admin?
To create the key, yes, or someone with admin access does it once. After that the loop is the same as any Scratch source: pick the tables, pull, point your agent at the folder.
See it on your own calls
Pull a month of calls and ask one question across all of them. Try Scratch free, or book a 30-minute demo on your own call library.