Skip to content

Commit 543d35b

Browse files
committed
docs: add CLAUDE_CODE_ATTRIBUTION_HEADER setting for local models
Critical for KV cache reuse -- without it, Claude Code prepends a changing header that forces full prompt re-processing on every request.
1 parent 3e44282 commit 543d35b

1 file changed

Lines changed: 14 additions & 7 deletions

File tree

docs-site/src/content/docs/integrations/local-llms.mdx

Lines changed: 14 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -89,21 +89,28 @@ cclocal 8124 # connect to localhost:8124
8989
cclocal 8124 --resume abc123 # with extra args
9090
```
9191

92-
:::danger[Disable Telemetry]
93-
Add this to `~/.claude/settings.json`:
92+
:::danger[Required Settings for Local Models]
93+
Add these to `~/.claude/settings.json`:
9494

9595
```json
9696
{
9797
"env": {
98-
"CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1"
98+
"CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1",
99+
"CLAUDE_CODE_ATTRIBUTION_HEADER": "0"
99100
}
100101
}
101102
```
102103

103-
Without this, Claude Code sends telemetry requests to
104-
your local server, which returns 404s and retries
105-
aggressively -- causing ephemeral port exhaustion on
106-
macOS and system-wide network failures.
104+
- **Attribution header** (`CLAUDE_CODE_ATTRIBUTION_HEADER`):
105+
Claude Code prepends an attribution header that
106+
changes on every request, invalidating the KV cache
107+
and forcing full prompt re-processing each time.
108+
Disabling it is critical for usable performance.
109+
- **Telemetry** (`CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC`):
110+
Without this, Claude Code sends telemetry requests
111+
to your local server, which returns 404s and retries
112+
aggressively -- causing ephemeral port exhaustion on
113+
macOS and system-wide network failures.
107114
:::
108115

109116
</TabItem>

0 commit comments

Comments
 (0)