OpenCode 2.0.28 auto-accept per project: paid Sume tools
OpenCode 2.0.28 lets you override Auto-accept permissions for a single project. Use Sume's read-only OAuth scope and spend gates before turning it on.

OpenCode 2.0.28 (October 13, 2026) lets you override the default for one project: "Set Settings -> Projects -> Auto-accept permissions to override the default for one project", or press mod+shift+a during a session. If a project with Sume's hosted MCP server turns that on, paid tool calls will no longer wait for you. Decide the gates before you flip it.
Sume's server is https://mcp.sume.com/mcp.
What does the 2.0.28 setting do?
According to the OpenCode changelog, the per-project setting overrides the default for that project, and the keyboard shortcut toggles it during a session. The changelog does not list which tools it covers, so test it on a harmless MCP call first.
The practical reading: one repo can be set to accept, and another can keep prompting.
| Version | Date | Relevant item |
|---|---|---|
| 2.0.28 | 2026-10-13 | Per-project Auto-accept permissions override |
| 2.0.27 | 2026-10-12 | Built-in worktree tools; confirmation before opencode pair --remote exposes a server |
| 2.0.26 | 2026-10-08 | Google Cloud credential support for Vertex |
Which Sume tools cost money?
The tools and gates page labels generation tools (image, video, audio, upscaling, background removal) as paid and requiring an idempotency_key. Discovery, account, job-read and crawl tools are read operations. dry_run=true previews cost in credits without submitting a job, and max_spend_credits caps spend when included in the request.
What is the safest way to auto-accept in a project?
Let the credential do the limiting, since auto-accept removes the prompt but not the scope. With OAuth, a token holding only mcp:read sees read-only tools, so an auto-accepting project cannot start a render. Add mcp:write only in a project where you want that.
There is no mcp:paid scope, so for a write-enabled project, paid submits are still decided by wallet and admission. That is a balance check, not an approval. Plan the cap yourself.
- Research or planning projects: OAuth, Read only, auto-accept on.
- Production projects: Write on, auto-accept off, or accept with
max_spend_creditsin every paid call. - Never paste the API key into the project file; use OAuth where a browser is available.
How do I verify what an auto-accepting project can do?
Ask the agent to run mcp_health and tools_list right after connecting. If generation tools are absent, the session is read-only. Then run a single dry_run=true call on a write-enabled project and check the estimate. Restricted calls without the Write scope return insufficient_scope, which is the sign the scope gate is working.
Why not just auto-accept everywhere?
Because the default is shared. A per-project override is the new part of 2.0.28; it exists so that one trusted project can differ from the rest. Turning it on globally would also cover projects where Sume is connected with Write and a funded wallet.
Another reason is retries. When a tool call is accepted without a prompt and the agent retries after an error, nobody sees the second submission. Sume's idempotency_key deduplicates a retry of the same request, but only if the agent reuses the key. Tell the agent in the project instructions to keep the key from the first attempt and to read the job list before it resubmits.
Finally, 2.0.27 added built-in worktree tools, so several agents in several worktrees can run in parallel under one account. Parallel auto-accepting agents share one wallet, which is one more reason to set max_spend_credits on each paid call.
Sources
Related posts
More in Integrations
- How to add an MCP server to ChatGPT with developer mode
Turn on ChatGPT developer mode, create an app for the server's URL, and sign in with OAuth. The steps, with Sume's hosted MCP server as the example.
- How to add subtitles to a video in Python
Add subtitles to a video in Python with Requests: POST the video URL to Sume's /v1/video-captions, poll the job, then read the captioned video_url.
- Add Sume to Claude as a custom connector (remote MCP)
Add Sume's hosted MCP server to Claude under Customize > Connectors, see what Sume's OAuth consent grants, and decide whether to allow paid tools.
- Airflow HTTP sensor: wait for an AI video job to finish
Submit an AI video job with Airflow's HttpOperator, then wait with an HttpSensor in reschedule mode that passes once the job's status is completed.
Written by Sume