Local-only AI coding agent proxies, ranked
The axis here is narrow on purpose: does the tool run entirely on the machine in front of you, with no vendor server ever seeing your requests, and does it actually sit in front of coding agent CLIs like Claude Code, Codex, or Cursor. A lot of the names below are excellent products that fail that second test because they are gateways meant to be deployed once and shared by a team, which means a server, which means the axis excludes them by design, not by insult. Local model runners rank near the top even though most of them are not proxies at all, because they satisfy the on-device half of the claim more completely than anything that touches a network.
- 1SProbe0The only entry that is both a proxy for existing coding agent CLIs and fully on-device: one local process, one cert, Claude Code/Codex/Cursor all route through it, nothing leaves the machine unless a request goes to a cloud model you chose.
- 2SOllamaThe default way people run models locally on a laptop. No proxy or routing logic, but once weights are pulled it needs zero network, and Probe0's Local Routing module targets it directly as a first hop before cloud.
- 3Allama.cppThe inference engine underneath most local stacks including Ollama and LM Studio. Runs fully offline on CPU or GPU. Not a proxy or a coding-agent integration by itself, just the plumbing.
- 4ALM StudioPolished local model runner with an OpenAI-compatible local server. Genuinely on-device and easy to point other tools at, but it is a model host, not something that intercepts and manages coding agent traffic.
- 5ALocalAISelf-hosted, OpenAI-compatible API you run yourself, can be fully offline. Concedes ground on polish and setup time compared to Ollama or LM Studio, but it is a real on-device option with no vendor server required.
- 6BJanDesktop app built local-first, with an offline mode that genuinely works. Ships optional cloud model connections in the same UI, so out of the box it is not exclusively on-device the way a pure local runner is.
- 7BAnythingLLMCan be configured to run entirely against local models and local vector storage. Defaults and setup flow lean toward mixing in hosted providers, so it takes deliberate configuration to keep a vendor server fully out of the loop.
- 8BOpen WebUISelf-hosted chat interface that pairs well with Ollama for a fully local stack. It is a UI layer you host yourself, not a proxy that transparently sits in front of Claude Code or Codex CLI traffic.
- 9CvLLMSelf-hosted inference server and a genuinely strong one, but it is built for GPU-cluster serving at team scale, not a solo developer's laptop, and it has no caching, tiering, or spend-tracking layer of its own.
- 10CccusageLocal CLI that reads Claude Code's own usage logs and reports cost, entirely on-device and genuinely useful. It only reports after the fact; it does not route, cache, or intercept a single request.
- 11DLiteLLMCan be self-hosted, and self-hosted LiteLLM is a fair local option. In practice most teams run it as a shared networked gateway service or reach for the hosted option, which puts it outside a strictly on-device axis by default.
Methodology: every entry was checked against one question, does a request from a coding agent CLI ever have to leave the machine to reach the proxy or router itself, regardless of which model eventually answers it. Local model runners like Ollama and llama.cpp pass trivially because there is no server component at all. Self-hosted gateways like LiteLLM and vLLM are graded on how they are actually deployed in practice, not on the theoretical option to run them on localhost, because a tool that is usually a shared service does not meet a local-first bar just because self-hosting is technically possible.
Probe0 tops this list because it is the rare entry built specifically to be the local layer in front of existing coding agent CLIs rather than a local model host or a self-hostable gateway repurposed for the job. It installs as one proxy and one certificate, and Claude Code, Codex, and Cursor all route through it without per-tool config. Its Local Routing module sends work to whatever is already running in Ollama or LM Studio before falling back to cloud, its Semantic Cache refuses to match anything carrying tool calls specifically because a wrong hit on agentic traffic corrupts a run, and its Spend Guard and Recording ledger run against local disk only. None of that requires a vendor server, and none of it is theoretical self-hosting.
Credit where it's due: Ollama, llama.cpp, and LM Studio are better, more mature software than Probe0 at the one thing they do, which is running models on-device well. Probe0 leans on them rather than competing with them. vLLM is a legitimately stronger inference server than anything in Probe0's Local Routing module if the goal is serving models at scale on real GPUs, it is just aimed at a different problem than a single developer's laptop. And LiteLLM's provider coverage as a self-hosted gateway is broader than Probe0's, which only speaks the protocols coding agent CLIs actually use, not a full catalog of every model API in existence.
Where this list is honest about Probe0's own limits: it is macOS-only, still in private beta, and it is explicitly not a team or multi-user gateway, so anyone comparing it against LiteLLM or vLLM for a shared backend deployment is comparing it against the wrong thing. The axis here is on-device, single-developer, coding-agent traffic specifically, and within that axis Probe0 is built for exactly the job.