Description
Gap
10.4.7 requires consent when a local MCP server is installed. It comes from SEP-1024, which targets one-click installation through links and configuration shared by others. Coding agents also start servers that a repository declares in project-level files (.mcp.json, .vscode/mcp.json, .cursor/mcp.json and similar). That is not an installation step, so 10.4.7 does not clearly apply, and the prompt the user answers is about trusting the folder, not about the programs it declares. 10.1.2 (allow-listed servers) covers this only where an allow-list is configured and applied to project-declared entries. Appendix C treats repository content as untrusted context (AC.3.3, AC.3.4), but not as configuration that starts processes.
Evidence
Between late September and 9 October 2026 I measured eight AI coding clients against a repository declaring one harmless canary server, each run from a clean VM snapshot. No client started it before the trust answer. But in three clients, accepting the folder-trust prompt started it with no separate question; in a fourth, it started without a question once the chat asked for a tool; and one client's non-interactive mode started it with no trust prompt at all. One client asked per server and showed the full command, which suggests the controls below are practical.
Method and data: https://paolocostanzo.github.io/trusting-a-repo-means-trusting-its-mcp-servers/
Proposed requirements (C10.4), kept atomic like the rest of the chapter:
- 10.4.x Verify that MCP clients do not start an MCP server declared in project or workspace configuration until the user has approved that server specifically, including in non-interactive modes; approving the folder is not approval of the servers it declares. (L2, like 10.4.7)
- 10.4.y Verify that MCP clients show the full command and arguments of a project-declared MCP server when asking for approval, and ask again when the declaration changes. (L2; this extends the command transparency of SEP-1024 to project-declared servers)
A cross-reference from Appendix C (AC.3) would point readers working on coding agents to them.
Related: #1201 (same chapter).
Activity
jmanico commentedon Oct 24, 2026
Good catch. My preference is to broaden 10.4.7 before adding two new requirements. The current wording ties consent to installation:
A project can start a program that is already on your machine. No installation step, but we are still executing code! That deserves explicit approval specific to the server.
The related controls help, but answer different questions: 10.1.2 covers which servers are allowed, and 10.1.3 limits what they can access through sandboxing. 10.4.7 is the right place to clarify consent.
There is useful support for this in the docs. MCP security guidance already requires showing the complete command and allowing cancellation for one-click configuration. VS Code documents that workspace MCP servers inherit folder trust, including after configuration changes. So the distinction you are raising matters.
I'd suggest this as a starting point for 10.4.7, still L2, in 1.02-dev:
For non-interactive automation, approval configured beforehand by an authorized user or administrator should work. We don't need a person clicking a dialog on every automated run. The repository must not be able to approve its own server by supplying an approval setting. User and managed approval settings already exist in clients, so there is a practical path here.
One piece still needs an explicit decision: changes to the approved launch configuration. Swapping the command or arguments must not inherit approval just because the server name stayed the same. That is distinct from #1201, which concerns tool definitions returned by the server. I wouldn't claim the wording above fully resolves that part yet.
Let's start by broadening 10.4.7 and work through that remaining change-approval question here. Thanks for bringing concrete evidence to this. 🙏
ottosulin commentedon Oct 24, 2026
Thanks @paolocostanzo for pointing this out, this is a real gap. Also agreed with Jim, let's try to rather work this into existing requirements.
And yes exactly we should not always require approvals for automated runs when an agent is using MCPs unattended. Related to this exact concern you're asking about, I posted a suggested revised version in #1201 - please check that. TL;DR: we should watch for any changes that get fed to the context.
paolocostanzo commentedon Oct 25, 2026
Thanks @jmanico @ottosulin. Agreed on broadening 10.4.7, and Jim's text works for me. I'd add his point that the project can't approve its own server, plus one sentence for the open change-approval question:
Otto, the revised 10.4.8 covers changes that reach the model context; the launch configuration never does, so I think it needs its own clause. Pre-approval by a user or administrator for unattended runs fits, as long as it is bound to the configuration and not only to the server name.
Working directory and environment matter as much as the command:
./serverresolves against the working directory, andNODE_OPTIONSorLD_PRELOADchange what runs without touching the command line. Comparing variables as declared, before interpolation, keeps secrets and host values out of the comparison. Changes to the code the command resolves to (an unpinnednpx pkg@latest, or a repository script edited by a later commit) are a different problem, closer to 10.1.1 and 10.1.3, and I'd leave them out of scope. I haven't measured re-approval after configuration changes; the closest data point is the VS Code documentation Jim cited.If you'd rather keep 10.4.7 to the launch decision, the last sentence works as a separate 10.4.22 (L2), mirroring 10.4.10 for endpoints. Happy to open the PR once we agree, as with #1204.
ottosulin commentedon Nov 4, 2026
@paolocostanzo in your revised version you have two requirements.
Broken down, and pulling the definition to glossary entry it could look like this:
Although that 10.4.7 could be broken into sentences like Jim's suggestion, making it a bit more readable.
The new part:
Our 1.01 basis for MCP requirements were pretty spec MUSTs + few picks from SHOULDs (I'd remember). To that point, this new part is not part of MCP spec's requirements - there closest language quickly found being (not exactly about installation, but could be said to implicitly cover that also?):
@jmanico / @RicoKomenda thoughts?
jmanico commentedon Nov 8, 2026
Thanks @paolocostanzo and @ottosulin. Let's go with Otto's split:
Plus the Launch Configuration glossary entry as Otto drafted it. Feel free to break 10.4.7 into shorter sentences in the PR if it reads better.
10.4.22 goes beyond the spec's MUSTs, but it applies the same logic as 10.4.10 does for endpoint changes, and the VS Code documentation already shows workspace servers inheriting trust across configuration changes in a shipping client. Pre-approval by a user or administrator for unattended runs still satisfies both, as long as it is bound to the configuration and does not come from the repository itself.
@paolocostanzo, please open the PR against
1.02-dev, with the Appendix B rows (revised 10.4.7, new 10.4.22), the glossary entry, and the AC.3 cross-reference you suggested. @RicoKomenda, shout if you see a problem.jmanico commentedon Nov 8, 2026
@paolocostanzo, whenever you're ready, please open the PR against
1.02-dev: the revised 10.4.7, the new 10.4.22, the Launch Configuration glossary entry, the Appendix B rows, and the AC.3 cross-reference. Thanks! 🙏paolocostanzo commentedon Nov 9, 2026
Thanks @jmanico @ottosulin. Opened #1218 as agreed, with 10.4.7 split into three sentences and no wording change. I left one open question in the PR: whether to add "complete" to the command and arguments, as in Jim's first draft 🚀
[1.02] Revise 10.4.7 for project-declared MCP servers and add 10.4.22 (…
[1.02] 10.4.7: generalize per review, keep folder trust and install e…