claude-webcache

claude-webcache

[npm] [npm downloads] [license] [tests]

Persistent cross-session WebFetch cache for Claude Code. Cached reads in ~0.07ms — orders of magnitude faster than re-fetching.

Claude Code's built-in cache lasts 15 minutes, within one session. Every new session re-fetches from scratch. claude-webcache persists results across sessions in a local SQLite database — instant cache hits, zero network cost.

Session 1  →  WebFetch("docs.example.com")  →  fetched, auto-cached ✓Session 2  →  cached_fetch("docs.example.com")  →  instant hit, no network callSession 7  →  cached_fetch("docs.example.com")  →  still instant, unlimited TTL

v0.1.5+: every WebFetch is automatically saved via PostToolUse hook — nothing to configure.

[CACHE_MISS flow: WebFetch + cache_store in first session] [CACHE_HIT flow: instant hit, no WebFetch in second session]

Install

claude plugin marketplace add theYahia/claude-webcache && claude plugin install claude-webcache@theyahia

Works in: Claude Code CLI · Desktop (Mac/Windows) · VS Code extension · JetBrains plugin — same command everywhere.

Done. Every WebFetch is auto-cached from now on.

Optionally add the usage pattern to ~/.claude/CLAUDE.md to also check the cache before fetching (saves the WebFetch call entirely on repeat URLs).

Plugin TUI not working? There's an open Claude Code bug (#41653) where /plugin install rejects third-party sources with "source type not supported." Use the CLI command above — it bypasses the TUI and works fine.

Fallback (no marketplace):

bash

git clone https://github.com/theYahia/claude-webcache && claude --plugin-dir ./claude-webcache/plugin

Option 2 — npm global

npm i -g @theyahia/claude-webcache

Requires Node.js 22.5+ (uses built-in node:sqlite — no native deps, no install step).

Then register in ~/.claude/settings.json (replace path with output of npm root -g):

{  "mcpServers": {    "claude-webcache": {      "command": "node",      "args": ["/path/from/npm-root-g/claude-webcache/scripts/mcp-server.cjs"]    }  },  "hooks": {    "SessionStart": [      {        "matcher": "startup|clear|compact",        "hooks": [          { "type": "command", "command": "node /path/from/npm-root-g/claude-webcache/scripts/hook-stats.cjs" }        ]      }    ]  }}

Option 3 — clone (contributors)

See CONTRIBUTING.md.

Usage pattern (optional — for pre-fetch cache checks)

v0.1.5+ auto-caches every WebFetch automatically. The pattern below is optional: add it to ~/.claude/CLAUDE.md to also check the cache before making a WebFetch — this saves the WebFetch call entirely on repeat URLs.

Auto-read (v0.5+): nothing to do. A PreToolUse hook checks the cache before every WebFetch/WebSearch. On a hit it serves the cached copy and skips the network; on a miss the call runs normally and the PostToolUse hook stores the result. Same URL + same prompt (or same search query) in any future session = instant hit, zero network cost.

Manual lookup is still available if you want it: call cached_fetch(url, prompt) (or cached_search(query)) — returns the cached text, or [CACHE_MISS] … if absent. Disable auto-read with WEBCACHE_AUTOREAD=0.

⚠ Security — authenticated URLs

The cache stores the URL alongside the response in ~/.webcache/cache.db. By default, claude-webcache strips obvious credentials from the stored URL before write (user:pass@host and query params named token, api_key, apikey, access_token, auth, secret, password, key, signature, etc.).

That's display-level redaction, not key-level. The cache key still hashes the original URL, so re-fetching the same authenticated URL hits the cache. If you want a stricter trade-off:

export WEBCACHE_STRICT_REDACT=1

With WEBCACHE_STRICT_REDACT=1, the cache key is computed from the redacted URL too — endpoints differing only in ?token=A vs ?token=B collide in one slot. Safe for pass-through auth (identical content), unsafe for personalized endpoints (different users see each other's cached data).

Bottom line: prefer header-based auth (Authorization: headers) over URL-embedded tokens. Don't commit ~/.webcache/cache.db to git.

Namespaces

Multiple projects sharing one machine? Isolate per-project caches:

WEBCACHE_NAMESPACE=gosdelo  claude    # cache writes/reads scoped to ns "gosdelo"WEBCACHE_NAMESPACE=qsearch  claude    # separate ns, no cross-contamination

Default namespace is the empty string "" (shared cache for v0.3 behavior). Inspect/manage per-namespace via CLI: claude-webcache namespaces, claude-webcache --namespace gosdelo stats.

Tools (MCP)

CLI

The npm package ships a claude-webcache binary for ad-hoc inspection and a local web dashboard:

claude-webcache stats                            # JSON statsclaude-webcache stats --by-domain                # per-domain breakdownclaude-webcache list 20                          # 20 most-recent URLsclaude-webcache list 50 --offset 100             # paginationclaude-webcache invalidate https://news.com/123  # drop one URLclaude-webcache refresh https://news.com/123 --prompt "extract title"   # invalidate one (url,prompt) pairclaude-webcache warm urls.txt --prompt "extract" # bulk pre-flight checkclaude-webcache clear --older-than-days 30       # partial wipeclaude-webcache clear --confirm YES              # full wipe (requires explicit confirm)claude-webcache clear-logs                       # truncate ~/.webcache/hook.logclaude-webcache namespaces                       # list all namespaces presentclaude-webcache export --out cache.json --all    # export metadataclaude-webcache dashboard                        # open http://localhost:37778claude-webcache --namespace gosdelo stats        # scope command to namespace

The dashboard renders top URLs by hits, top domains (with avg hits / last fetch / entry counts), full search-able paginated list with one-click invalidate + refresh buttons. Pure stdlib — no extra deps to install.

Configuration (env vars)

SessionStart hook

Every new session injects a one-liner so Claude knows the cache exists:

webcache [ns=gosdelo] 142 pages cached, 87% hit rate, last fetch 3h ago

No output if cache is empty. [ns=...] is omitted when using the default namespace.

Storage

SQLite at ~/.webcache/cache.db (WAL mode, synchronous=NORMAL, busy_timeout=5000). Cache key = SHA256(namespace + "|" + canonical(url) + "|" + prompt). Default TTL: unlimited (set WEBCACHE_TTL_DAYS=N for N-day expiry).

URL canonicalization (v0.4+): lowercase hostname, strip default ports (:80/:443), strip fragment, sort query parameters alphabetically. So https://EXAMPLE.com/p?b=2&a=1#frag and https://example.com/p?a=1&b=2 produce the same cache key — no silent miss on URL formatting variance.

Concurrent-safe via WAL + 5-second busy_timeout — multiple Claude Code sessions can read/write simultaneously without SQLITE_BUSY errors.

Limits

• Cache key includes the prompt — use consistent prompts to maximize hit rate.
• Output is whatever WebFetch returns (already summarized). No re-processing.
• No semantic search. Exact (namespace, canonical_url, prompt) match only.

Benchmarks

Single-process latency on a populated DB (N=10000 entries, 1KB output each), measured via npm run bench:

Storage overhead: ~2 KB per entry for a 1 KB payload (key + indexes + WAL + new v0.4 columns). With WEBCACHE_COMPRESS=1 on text-heavy responses, expect 3-7× reduction.

WebFetch over the network typically takes 1-5 seconds — a cached hit is ~15,000-70,000× faster. Reproduce on your hardware: npm run bench. See bench/README.md for methodology and full results metadata (CPU, RAM, OS, commit) saved per run.

Related

• claude-mem — persistent memory across sessions (complements claude-webcache: memory vs. web cache)
• WWmcp — catalog of 46 MCP servers for non-Western APIs

License

MIT — see LICENSE.

Часть WWmcp · Telegram: @vhodvai