Codex MCP output_token_limit: what to set for Sume tools
Codex lets you cap one MCP tool's output with output_token_limit. Sume tools return ids and media.sume.com URLs, not file bytes, so the cap can stay small.
Set mcp_servers.<id>.tools.<tool>.output_token_limit per tool, and keep it small for Sume: Sume's MCP tools report public ids and media.sume.com URLs, and hosted MCP never carries file bytes. The key is a token budget for one tool's output; the value to choose depends on the tool, and the number below is only an example.
What does the Codex config key do?
The config reference, read 2026-09-30, lists mcp_servers.<id>.tools.<tool>.output_token_limit as the "token budget for one MCP tool's output". It sits next to startup_timeout_sec (default 10 s) and tool_timeout_sec (default 60 s). The 0.158.0 release notes also say MCP and Code Mode input schema budgets became configurable; the notes do not name the keys, so check the reference for them.
[mcp_servers.sume]
url = "https://mcp.sume.com/mcp"
# Example value, not a Sume recommendation from the docs
[mcp_servers.sume.tools.jobs_result]
output_token_limit = 2000Why can Sume's limit stay small?
Per MCP tools and gates, agents should report Sume public ids and media.sume.com URLs, and hosted MCP cannot read files from your laptop; uploads go through an upload URL, a client PUT, and assets_complete. Generated media comes back as a URL, so there is no large payload to budget for.
| Concern | What the docs say |
|---|---|
| Result shape | Public ids and media.sume.com URLs |
| File bytes | Hosted MCP cannot read files from your laptop |
| Tool description size | Capped at 2048 bytes per jobs-catalog tool in the server code |
| What is visible | tools_list shows the session-visible subset |
What should I size the input budget against?
Size against what your session can see. tools_list returns the session-visible subset, so a read-only OAuth session exposes fewer tools than an API-key session and costs less schema. The jobs tool descriptions are byte-capped in the server code (JOBS_TOOL_DESCRIPTION_MAX_BYTES is 2048), which keeps each description small.
What if a tool still exceeds the limit?
Raise that one tool's value rather than the whole server, and see MCP tool exceeds maximum allowed tokens for the general fix. Do not paste signed URLs or keys into logs while debugging.
Sources
Related posts
More in Developers
- Computer use origin approval is not a spend approval
OpenAI's computer use approves each new website origin, not spend. When such an agent calls Sume, the spend check is dry_run, max_spend_usd and the wallet.
- Content Credentials after download: check the file you deliver
C2PA 2.2 defines Content Credentials and a separate Soft Binding API. Sume's docs do not say a downloaded output keeps one, so check the delivered file.
- Can I continue a Sume Agent Completion with thread_id?
Not yet. Every Agent Completion runs in a fresh thread, so a follow-up call cannot reuse thread_id. Pass earlier results back in the next request instead.
- Crawl job finished event: Sume has none, poll crawl_get
Cloudflare publishes crawl.finished to Queues. Sume's crawl has no such event: wait on the job id, then read it with crawl_get. Steps and idempotency.
Written by Sume