Eugene LevitinFounderDocka
A startup that tests whether AI agents finish real jobs in SaaS products. One monday.com job run in Copilot Studio, Oct 5, 2026.
GitHub, mondaycom/mcp, published Read the issue (opens GitHub, mondaycom/mcp in a new tab)
Observation
At Docka, a monday.com job in Copilot Studio failed in all 3 runs with GPT-5.5 Chat on the MCP URL path, because Copilot Studio kept only the first 70 tools.
- Takeaway
- Test each MCP server through the same path and model your agent will use in production. Count the tools the server lists, because tools past the 70th never reach the model.
- What happened
- The same monday.com job passed 6 of 6 runs through the certified connector and 3 of 3 with Claude Sonnet 4.6 on the MCP URL path, but failed 0 of 3 with GPT-5.5 Chat there and created nothing. Copilot Studio's transcripts showed it kept only the first 70 tools the server listed, so the search tool, now 85th, never reached the model, and GPT-5.5 Chat asked the user for GraphQL variables and never sent the call.
- Public engineering write-up (GitHub issue in monday.com's MCP repository)
- Run counts per path and model stated
- Transcripts quoted by the founder
Deployment details
A startup that tests whether AI agents finish real jobs in SaaS products. One monday.com job run in Copilot Studio, Oct 5, 2026.
- Public engineering write-up (GitHub issue in monday.com's MCP repository)
- Run counts per path and model stated
- Transcripts quoted by the founder
- Team
- Docka founder
- Duration
- Runs on Sep 23 and Oct 5, 2026
- Workload
- A monday.com board job run by a Copilot Studio agent
- Product version
- Copilot Studio with the monday.com MCP URL and the certified connector
- Model
- GPT-5.5 Chat and Claude Sonnet 4.6
- Measurement
- 0 of 3 runs with GPT-5.5 Chat on the MCP URL path; 3 of 3 with Claude Sonnet 4.6; 6 of 6 through the certified connector

