Xcode relaxes its AI agents but still blocks external connections
Settings found in Xcode 27.0 beta allow the editor to control its AI agents with fewer prompts. But the first connection of an external agent to Xcode tools remains blocked by an approval dialog.
Xcode 27.0 beta gives more leeway to its internal AI agents but does not yet resolve fully autonomous access of an external agent to the editor’s tools. According to a check dated June 16, 2026 on build 27A5194q, the setting IDEChatAgenticChatSkipPermissions allows Xcode to launch a Claude agent with the equivalent of a “no permissions” mode, while the connection of a third-party agent still requires approval on first use. This issue matters for developers already automating their workflows in Xcode, but also for security: the editor keeps a safeguard where one might have expected a wide-open door.
Un raccourci pour l’agent interne, pas pour tout le monde
The first observation is simple: Xcode can now pass to its Claude agent an indicator that bypasses the agent’s own permission checks. The command defaults write com.apple.dt.Xcode IDEChatAgenticChatSkipPermissions -bool YES activates this behavior, and the launched process then includes the option --dangerously-skip-permissions. In short, Xcode lets its own agent act without asking permission at each step. The parallel with a “dangerously” free start is not subtle, but it is transparent.
However, this switch does not solve the opposite case: having Xcode controlled by an external terminal agent without the application asking for approval at each launch. On this point, the note is clear: the lock remains in place, including in 27.0 beta.
Ce que Xcode lance vraiment sous le capot
Documentation from binary exploration shows that Xcode assembles a more compact invocation than in 26.3. The command revealed by xcrun mcpbridge run-agent claude --dry-run in 27.0 notably displays --verbose, a temporary MCP configuration file, and session parameters like CLAUDE_CODE_ENABLE_TELEMETRY=0 and DISABLE_TELEMETRY=1. Xcode also provides ANTHROPIC_MODEL=best, CLAUDE_CONFIG_DIR, and MCP_XCODE_PID.
The role of MCP_XCODE_PID is practical: it allows the mcpbridge bridge to find the correct Xcode instance and access its tool service via stdio. Without this variable, the connection stops. Meanwhile, the temporary MCP file only serves to connect native Xcode tools, not third-party servers, which remain defined in the agent’s .claude.json file.
Le vrai point dur : l’approbation à la première connexion
Xcode 27.0 adds a promising setting, IDEIntelligenceAllowApprovingUntrustedAgentsForSession. The name suggests it could remember an approval during the current session. In practice, tests described in the source did not eliminate the dialog box that appears on the first connection of an external agent. Control thus remains with the user, at least for this initial “do you really want to allow this agent to use Xcode tools?”.
The setting IDEAllowUnauthenticatedAgents still acts as a main switch for external MCP access. Disabled, the MCP server shuts down. Enabled, it allows connection but does not automatically remove the approval request. Another key, IDEChatInternalAllowUntrustedAgentsWithoutUserInteraction, seems aimed at this bypass, but it only works in Apple internal builds. In public or beta versions, it remains ineffective [to be verified].
27.0 élargit la surface IA, sans changer le verrou de sécurité
Between 26.5 and 27.0 beta, the AI-related surface expanded from 138 to 163 IDE keys spread across 7 then 10 frameworks, according to strings extracted from binaries. Most changes concern model providers, agent protocols, or MCP configuration. Permission and approval settings remained identical between the two versions. In other words, Apple rearranged a lot but not the lock.
Among the reported additions are IDEChatEnableBYOACP with the new IDEIntelligenceACP framework, keys related to Gemini such as IDEChatGeminiCLIAPIKey and IDEIntelligenceGemini, as well as new settings for model selector or MCP configurations. However, some names spotted in an older table are no longer confirmed in 27.0 nor in the final 26.5; the source itself flags them as potentially speculative [to be verified].
Ce que les développeurs doivent retenir
For daily use, the message is quite clear: Xcode becomes more flexible for its own agents, but it does not yet switch to “free control” mode for external connections. Teams experimenting with Claude Code, Gemini, or third-party MCP servers can advance automation but will still have to deal with the initial approval window. The good news is that the settings are visible. The less good news is that they do not all do what one might hope. AI tools love shortcuts; Xcode, however, continues to read the “warning” sign before opening the door.
- 27.0 beta confirms Xcode’s move to more AI keys.
- IDEChatAgenticChatSkipPermissions triggers a launch without internal control.
- The first connection of an external agent remains subject to approval.
- The mcpbridge bridge depends on MCP_XCODE_PID to target the correct instance.
- Permission keys have not changed between 26.5 and 27.0 beta.
- New settings mainly concern models, MCP, and providers.
By the numbers
- June 16, 2026 — date of re-check on Xcode 27.0 beta (build 27A5194q).
- 138 IDE keys — surface spotted in 26.5 release, according to binary strings.
- 163 IDE keys — surface spotted in 27.0 beta, according to binary strings.
- 10 frameworks — where IDE keys live in 27.0 beta.
- 7 frameworks — number of IDE frameworks in 26.5 release.