Agent tool surface decides which fight your tool has to win
Where you expose a tool decides what it competes with
I had been reluctant to expose Orbit tools via MCP because I was concerned about the context overhead compared to a CLI interface. That had led me to stubbornly push for CLI as the primary tool interface for agents.
This changed during our graph-tool benchmarking session. The experiment had three groups:
graph-only: only graph toolshybrid: graph tools plus ordinary file reading and searchno-graph: only ordinary file reading and search
In v1 and v2, Codex accessed graph tools through exec_command, using commands such as orbit tool run orbit.graph.search. Claude, on the other hand, had graph tools exposed through MCP alongside Read, Grep, and Glob. Across the two trials, hybrid Codex never reached for the graph tools in 60 runs. Not even once.
In v3, we changed things up and gave Codex MCP graph tools instead. Hybrid Codex made 23 graph-tool calls across 30 runs.
Claude had MCP all along, yet used graph tools just once across 60 hybrid runs in v1 and v3 (v2 was skipped for Claude).
| Agent | Surface for graph tools | Hybrid runs (n) | Graph-tool invocations |
|---|---|---|---|
| Codex | exec_command (v1) | 30 | 0 |
| Codex | exec_command (v2) | 30 | 0 |
| Codex | MCP (v3) | 30 | 23 |
| Claude | MCP (v1) | 30 | 1 |
| Claude | MCP (v3) | 30 | 0 |
My suspicion is that familiarity and visibility play a role at each tool selection point.
For Codex, in v1 and v2, the selection point came after exec_command, where it had to choose between orbit tool run orbit.graph.search and rg. Even for humans, we would reach for the familiar rg command over unfamiliar and untested ones.
By exposing the graph tools as MCP in v3, the tool selection point moved one level higher. And the selection option became: orbit.graph.search and exec_command. So Codex was more likely to choose the graph tool this time.
For Claude, the graph MCP tools already sat alongside internal tools such as Read, Grep, and Glob - so there was no room for a better selection point to test.
In summary, Claude's MCP-heavy native tool exposure leaves little room for tools of similar nature; while Codex provides more rooms to work with.
Takeaway
The exposure surface of a tool can determine its likelihood of being used by an agent.
When tools are surfaced through MCP, they gain higher visibility, which can lead to increased usage, as seen with Codex in v3. Conversely, if tools are only accessible through a generic shell interface, they may be overlooked in favor of more familiar commands.