One service vs. many connectors
We tested Claude Code on coding tasks adapted from EnterpriseRAG-Bench, comparing Unblocked MCP with separate MCP connectors. Each task requires company knowledge from docs, conversations, tickets, or meetings. With no context outside of the code, the task is unsolvable.
Claude Code~14 calls
Claude Code~3 calls
Unblocked MCP- What changes with Unblocked
- Unblocked assembles relevant company knowledge for Claude Code through a single MCP server. With separate MCP connectors, Claude Code searches each connected tool and tries to assemble that knowledge itself.
- How we tested
- Both setups used Claude Code with Claude Opus 5.5, the same API account and sandbox, and only the task prompt. Tasks covered basic lookups, conflicting documents, and questions whose answers were missing.
Inside the dataset
107 coding tasks over the same 511,962 documents from nine enterprise tools. Each task is decided by facts in one or more of them.
| Source | Tasks | Share of tasks |
|---|---|---|
| Confluence | 46 | 43% |
| GitHub | 38 | 36% |
| Linear | 16 | 15% |
| Jira | 15 | 14% |
| Slack | 11 | 10% |
| Google Drive | 9 | 8% |
| Gmail | 4 | 4% |
| Fireflies | 3 | 3% |
| HubSpot | 3 | 3% |
28 tasks need two or more sources, so the counts add up to more than 107.
Methodology
- How the tasks were built
- We turned EnterpriseRAG-Bench questions into 107 code changes. Each is a repo at a fixed commit, a change request, and hidden tests that check the facts that decide the change. Those facts live only in the company’s docs, tickets, threads, meetings, email and CRM. Claude Code with no outside context solved none of them.
- How the setups differ
- Every setup runs Claude Opus 5.5 on the same tasks, prompts, sandbox and API account, with sub-agents off. Only the route to company knowledge changes. The MCP connectors are synthetic: nine of them, one per tool, using each vendor’s own tool names and schemas over the same corpus.
- How it’s scored
- A run is solved when every hidden test passes. Each setup ran each task three times. Cost is list-price Claude usage.
Built on the EnterpriseRAG-Bench corpus from Onyx, 1,279 graded runs, . Unblocked built the coding tasks; Onyx did not review this work.