Claude Code
github.com/anthropics/claude-code
it has a tool for
read from source
Claude Code registers 22 tools, covering 4 of 6 compared capabilities. It registers no built-in tool for browser or long-term memory.
How this list was established. Tool surface read from its own runtime rather than from published source.
Surface size
tools registered, across the 13 harnesses hereSurface size, not quality. Every registered tool spends context on every request, so a longer bar is a cost as much as a capability — and a harness that reaches a capability through a general tool may register fewer.
Capabilities
what serves each, and where it is compared| Capability | Tools | Across harnesses |
|---|---|---|
| Read | Read | 13 of 13 have one → |
| Browser | no built-in tool | 3 of 13 have one → |
| Web search | WebSearch | 9 of 13 have one → |
| Code search | Globpartial · Grep | 13 of 13 have one → |
| Shell | Bash · BashOutputpartial | 13 of 13 have one → |
| Long-term memory | no built-in tool | 4 of 13 have one → |
Where its tools differ from the rest
- Read · Staleness check before writeinvariantholdsvs fails in 6 of 8
- Read · Whichever hits firstlinesvs lines or bytes in 4 of 5
Derived by comparing every recorded value against the other implementations of the same capability. Differing is not a defect unless the row is marked an invariant.
All 22 tools
names verbatim, as registeredNot yet compared means registered by Claude Code but with no comparison page here — todo lists, subagent spawning, approval flows. They become capabilities the day there is a comparison worth reading, and the data does not have to change for that to happen.
Codex →