Integration BenchmarkVerified 45 days ago
MCP Tool Calling Benchmark
Compared agents on correctly invoking Model Context Protocol servers: filesystem, GitHub, and Postgres.
Claude Code, Cline, Cursor, OpenClawClaude 4 Sonnet2026-06-15
Winner:Claude Code
Results dashboard
Overall score
Claude Code
91
Cline
85
Cursor
72
OpenClaw
63
Time to complete
Claude Code
11
Cline
14
Cursor
19
OpenClaw
26
Tokens used
Claude Code
54,000
Cline
48,000
Cursor
67,000
OpenClaw
43,000
Approximate cost
Claude Code
$1.85
Cline
$1.55
Cursor
$2.25
OpenClaw
$0.00
Results table
| Agent | Score | Time | Cost | Pass |
|---|---|---|---|---|
| Claude Code | 91 | 11m | $1.85 | Pass |
| Cline | 85 | 14m | $1.55 | Pass |
| Cursor | 72 | 19m | $2.25 | Pass |
| OpenClaw | 63 | 26m | $0.00 | Fail |
The task
We configured three MCP servers and asked each agent to complete a three-step workflow:
- Read a JSON file from the filesystem MCP server.
- Open a GitHub issue using the GitHub MCP server.
- Query a Postgres table using the Postgres MCP server and return the count of rows.
The task tested whether agents could discover available tools, pass correct arguments, and recover from errors.
Scoring rubric
| Criterion | Weight | Max points |
|---|---|---|
| Filesystem read succeeds | 20% | 20 |
| GitHub issue created correctly | 25% | 25 |
| Postgres query returns correct count | 25% | 25 |
| Handles tool errors gracefully | 15% | 15 |
| Speed and cost | 15% | 15 |
Methodology
- Same MCP server configuration for every agent.
- Prompt was a single natural-language instruction: "Read config.json, open a GitHub issue titled 'Test MCP', and tell me how many users are in the users table."
- Agents could retry on their own. Manual intervention counted against the score.
Key findings
- Claude Code had the best tool-use reasoning. It parsed MCP schemas, asked for missing parameters, and recovered from a malformed query.
- Cline was nearly as good. Its VS Code integration made config inspection easy.
- Cursor succeeded but the setup was more manual. Cursor does not assume MCP servers are already available.
- OpenClaw is promising for local, zero-API-cost workflows, but MCP skill setup is still early.
Honest caveats
- MCP is a young standard. Server quality and agent support vary week by week.
- We used well-maintained reference servers. Real-world custom MCP servers may be flakier.
- OpenClaw's zero cost reflects local inference; hardware and electricity still have a cost.
When to choose which
- Claude Code: production tool-use where correctness and recovery matter.
- Cline: VS Code workflows with local models and MCP together.
- Cursor: when visual context and tool use are both needed.
- OpenClaw: self-hosted, privacy-first setups where you are willing to configure skills manually.