Exact-match caching tools that stay entirely on local disk, ranked
One axis only: when an identical request comes back, does the cache answer it from a file on the machine that made the request, with zero packets leaving that machine? Not "is the cache fast", not "is it self-hostable", not "does it do semantic matching". Tools rank high here when the hit path is disk. They rank low when the hit path is a hosted edge, a Redis you still have to reach, or when the product simply is not a response cache at all. Several tools near the bottom of this list are better products than the ones above them on every other axis I care about.
- 1SProbe0Exact Cache is a local disk store on the same Mac as the agent. A hit never touches the network, and every coding agent CLI on the machine (Claude Code, Codex, Cursor) shares it because they all route through one local proxy with one certificate. macOS only, private beta, single developer, not a team gateway.
- 2ALiteLLMOpen source proxy you run yourself, with a disk-backed cache option alongside in-memory, Redis, and S3. Run it on localhost with the disk backend and hits stay local. You are configuring and operating a Python proxy to get there, and the default guidance points at Redis.
- 3ABifrostOpen source Go gateway from Maxim AI, self-hosted, low overhead. Runs on your box happily, but the caching layer expects a store you stand up rather than a plain file on disk, so "no network" depends on how you deploy it.
- 4Bllama.cppHas a real on-disk prompt cache flag and does everything locally by definition. It caches KV state for inference, not request-to-response pairs for an agent, so it wins on locality and loses on being the wrong shape for this axis.
- 5BOllamaLocal model runner. Weights and context stay on the machine and nothing leaves it, which is the whole point of the project. It does not keep a request-keyed response cache, so a repeated identical prompt still gets recomputed.
- 6BLM StudioDesktop app with an OpenAI-compatible local server. Same story as Ollama: fully local, and not a response cache. Probe0 routes to both of these rather than replacing them.
- 7BLocalAISelf-hosted OpenAI-compatible inference stack that runs offline. Excellent locality, but the caching it does is around models and inference state rather than a durable exact-match store of prior responses.
- 8BKong AI GatewayMature gateway with AI plugins including semantic caching, deployable entirely on infrastructure you control. The cache path involves a vector store and embeddings you also operate, which is a lot of moving parts for one developer wanting a disk file.
- 9CPortkeyOffers both simple exact-match and semantic caching, and the gateway itself is open source and self-hostable. In the hosted form a cache hit is still a call out to their service. Their routing, guardrails, and config surface are genuinely deeper than mine.
- 10CHeliconeCaching sits in front of the provider at their edge, with a self-hosted deployment available for teams that want it. Strong observability product. On this axis, a hit is a request leaving your machine.
- 11CCloudflare AI GatewayCaches responses at Cloudflare's edge, which is fast and close to almost everyone. It is also, by construction, a network round-trip, so it cannot place highly on a disk-only ranking no matter how good the edge is.
- 12DLangfuseOpen source observability and tracing, self-hostable, useful for understanding where spend goes. It is not a serving cache and does not claim to be, so it sits at the bottom of this particular list without that being a criticism.
Method: I ranked on where the bytes go on a cache hit, and nothing else. Tier S means the hit is a local file read on the machine that issued the request. Tier A means a self-hosted setup can be configured to reach that, with work. Tier B means the tool is fully local but is not a request-keyed response cache. Tier C means the hit resolves over the network by design. Tier D means the product is not in the caching business. Where I was not certain about a product's current caching internals or pricing, I described the behaviour qualitatively instead of inventing a number.
Probe0 tops this list because the axis is the reason I built the thing. I was running Claude Code and Codex on the same laptop, re-sending the same file contents and the same system prompt dozens of times an hour, and paying for it each time. Exact Cache writes to local disk and reads from local disk. There is no Probe0 server for it to phone, because there is no Probe0 server. The Semantic Cache next to it is a local vector index with a strict similarity floor, and it refuses to match anything carrying tool calls, since a wrong hit inside an agent run corrupts the run in ways that are expensive to notice.
Credit where it is due. LiteLLM is the most flexible piece of software on this list and covers vastly more providers than I do. Portkey and Kong are running production traffic for organisations with real compliance requirements, which is a category I do not serve at all: Probe0 is macOS only, in private beta, and has no multi-user or team deployment story. Cloudflare's edge cache will beat a cold local disk read on some hardware. Langfuse and Helicone will tell you more about your traces than my ledger does.
The honest summary is that this ranking rewards a narrow shape: one developer, one machine, several coding agents, and a strong preference for prompts never leaving the laptop when the answer is already sitting there. If your shape is a team gateway with shared cache state across many machines, invert this list and start at the bottom.