Claude Code saves MCP text over 50,000 characters: size Sume reads
Claude Code writes long MCP text results to a file, and Sume caps a tool answer at 256 KiB. Read job waves in batches of 20 ids so each answer stays small.
Claude Code saves successful MCP text results longer than 50,000 characters to a file, whatever MAX_MCP_OUTPUT_TOKENS says. Sume separately caps one tool answer at 256 KiB. If you read a big wave of finished jobs in one call, you can hit either limit, so read waves in groups of 20 ids or fewer and keep each result small.
The limits side by side
The Claude Code figures come from its MCP page. The 20-id ceiling for both jobs_wait and jobs_result is on Sume's jobs and results page. The 256 KiB answer budget is part of the server's work bounds in the Sume repo: oversize output returns an explicit error, not a silent cut.
| Where | Limit | What happens past it |
|---|---|---|
| Claude Code | Warning above 10,000 tokens | Warning shown |
| Claude Code | 25,000 tokens default maximum (MAX_MCP_OUTPUT_TOKENS) | Raise it with the variable if you need more |
| Claude Code | 50,000 characters of text | Saved to a file regardless of the variable |
| Sume hosted MCP | 256 KiB serialized per tool answer | Explicit oversize error, no silent truncation |
| Sume jobs_wait / jobs_result | 20 job ids per call | More ids need more calls |
Why the file path is not a Sume problem
When Claude Code stores a long answer in a file, the agent has to open the file to see it. The job itself is unaffected. Results are durable, so a wave you cannot read in one answer can be read again with jobs_result.
A practical pattern
Wait on up to 20 ids with include_results: true if the results are small, such as transcripts of short clips. Sume names any result that does not fit in results_omitted.job_ids; read those with one batch jobs_result call. For media jobs, the result is mostly URLs, so it stays small anyway.
An oversize Sume answer does not mean the related job failed. Do not resubmit paid work because a result was too large.
- Group ids in tens or twenties rather than sending 60 at once.
- Quote
media.sume.comURLs, not signed links, in the final report. - If the answer lands in a Claude Code file, read the file before deciding anything about the job.
The tradeoff
Batching ids saves calls but concentrates output. If one answer in a batch is large, it can push the whole batch past a limit. Smaller groups cost one more call and remove the surprise.
Sources
Related posts
More in Integrations
- Cloudflare Worker: keep the image model id in KV and swap it live
A Worker that proxies POST /v1/images and reads the Sume model id from Workers KV, so replacing gpt-image-1 is a wrangler command, not a deploy.
- Cloudflare Workers cron (UTC) or a Sume schedule with a timezone?
Worker cron triggers run on UTC and take up to 15 minutes to propagate. Sume schedules take an IANA timezone and a skip-or-reject overlap rule.
- Codex startup_timeout_sec 10 and tool_timeout_sec 60 for Sume
Codex defaults to a 10 s startup and a 60 s tool timeout per MCP server. Sume's jobs_wait holds at most 55 s, so the tool default only just fits.
- Cursor MCP auth block with CLIENT_ID and CLIENT_SECRET: skip for Sume
Cursor's mcp.json supports an auth block with CLIENT_ID and CLIENT_SECRET for servers that need a fixed client. Sume's hosted OAuth does not; use just the URL.
Written by Sume