[BUG] Per-Turn Tool Call Limit Regression Breaks Agentic MCP/SSH Workflows in Claude Desktop #33969
Description
Preflight Checklist
- I have searched existing issues and this hasn't been reported yet
- This is a single bug report (please file separate reports for different bugs)
- I am using the latest version of Claude Code
What's Wrong?
The per-turn tool call limit in Claude Desktop has been significantly reduced without announcement. Sessions that previously ran 60–80+ tool calls autonomously over 45+ minutes now hit "Claude reached its tool-use limit for this turn" after approximately 20 tool calls, requiring a manual "Continue" click to resume. This breaks agentic MCP/SSH workflows where Claude is given a well-specified task and expected to execute autonomously.
This is not a rate limit or usage cap issue — it's a per-turn ceiling on tool calls within a single response. I am on the Max plan.
Example Session (March 13, 2026)
Task: Fix multiple issues in a HubSpot API integration (company name resolution via two-hop engagement→contact→company, future-dated item filtering, email volume caps, activity type filter UI).
Claude chained approximately 20+ tool calls (RunRemoteCommand and Used mcp-ssh integration) across:
Debugging a toObjectId string/number type mismatch in v4 batch API responses
Adding company associations via v4 batch
Filtering future-dated meetings/tasks
Adjusting per-type fetch limits
Adding a type filter dropdown to the Activity tab UI
Fixing prop destructuring
Session was interrupted at the final commit step with: "Claude reached its tool-use limit for this turn."
What Should Happen?
Claude should be able to execute extended multi-step tool call chains (60–80+) without interruption, as it did previously. At minimum, Max plan users paying $100–$200/month should have a significantly higher or configurable per-turn limit. An "auto-continue" option should exist so sessions resume without manual intervention.
Suggested Improvements
Restore previous behavior: The per-turn tool call limit should return to its previous level, or be removed for Max plan subscribers
Configurable per-turn limit: Let users set their own threshold (or "unlimited") in Claude Desktop settings
Auto-continue option: If a limit must exist, offer an "auto-continue" toggle so the session resumes without manual intervention
Plan-tiered limits: Scale the per-turn cap based on plan tier (e.g., Free: 10, Pro: 30, Max 5x: 100, Max 20x: unlimited)
Announce changes: Any modification to tool call limits should be communicated via changelog or support docs
Error Messages/Logs
No error in the traditional sense — a UI banner appears:
"Claude reached its tool-use limit for this turn"
With a "Continue" button. The underlying API mechanism is stop_reason="pause_turn" when the server-side sampling loop hits its iteration limit (documented default: 10 iterations).Steps to Reproduce
Open claude.ai with MCP SSH tools connected
Give Claude a multi-step coding task (e.g., debug and fix multiple issues in an API integration)
Let Claude work autonomously — it chains RunRemoteCommand and SSH tool calls
After ~20 tool calls, session pauses with "Claude reached its tool-use limit for this turn"
Claude Model
Opus
Is this a regression?
Yes, this worked in a previous version
Last Working Version
No response
Claude Code Version
N/A — this bug is in Claude Desktop (latest) interface, not Claude Code CLI.
Platform
Other
Operating System
macOS
Terminal/Shell
Other
Additional Information
Previous behavior: Identical workflows (multi-step code builds via MCP SSH) ran 60–80+ tool calls over 45+ minutes without this interruption. This was consistent and repeatable across many sessions over several months.
Current behavior: The limit is hit at approximately 20 tool calls per turn. The "Continue" button resumes the session but breaks flow, momentum, and sometimes causes Claude to lose context on what it was doing.
Setup: Claude Desktop on macOS, MCP SSH integration via @idletoaster/ssh-mcp-server, connecting to remote Linux servers via Tailscale. Max plan subscriber.
Why this matters for agentic workflows: The value proposition of MCP tool integration is that Claude can execute complex multi-step tasks independently. A manual checkpoint every ~20 calls turns autonomous execution into a babysitting exercise. This is especially impactful for users who invest in thorough upfront specification and expect Claude to execute the full plan without interruption.
Note: This issue is filed against claude-code as there is no public claude-desktop issues repo. This bug occurs in Claude Desktop, not Claude Code CLI. Please route to the appropriate team.