Run local models
LocalBox
Launch local GGUF models through llama.cpp and connect them to Claude Code, Codex, or LocalPilot.
Start here if: you need a local model running reliably.
Explore LocalBoxYour local AI toolkit
From local model to useful coding agent.
LocalStack brings four focused tools into one practical workflow: run a model, tune it for your machine, put it to work, and keep what it learns.
Local-first · No usage telemetry · Your data stays yours
How it fits together
One path. Four clear jobs.
Use only what you need today. Add the next piece when you are ready.
- 01RunLocalBox serves the model
- 02TuneLocalBench fits it to your hardware
- 03CodeLocalPilot gives it tools, research, and a loop
- 04RememberLocalMind keeps reviewed memory and docs
Privacy by design
Your machine. Your models. Your data.
LocalX does not collect your prompts, code, model activity, memories, benchmark results, or usage analytics. The local path stays local, with no LocalX account required.
Choose a hosted provider or download? That network action is explicit and goes only to the service you selected—not through LocalX.
- No usage telemetryNo product analytics or silent reporting back to us.
- Local by defaultModels, memory, logs, reports, and configuration stay under your control. Cross-device sync is opt-in and end-to-end encrypted.
- Fully yoursInspect it, move it, back it up, or delete it whenever you choose.
Choose your starting point
What do you need right now?
Each project does one job and works on its own. Together, they form the LocalX toolchain.
Tune your hardware
LocalBench
Find fast, stable runtime settings for your machine and export ready-to-use launcher profiles.
Use it when: the model runs, but not as well as it could.
Explore LocalBenchBuild with an agent
LocalPilot
Drive any supported model through a provider-neutral coding agent with tools, permissions, and harness mode. It runs deep research with every web fetch disclosed and auditable, can ask before guessing at unclear requests, and another agent can coach it over MCP.
Use it when: you want the model to work across a codebase.
Explore LocalPilotKeep useful context
LocalMind
Turn reviewed sessions and project docs into searchable local memory. Manage it in a local web UI, serve it to agents over MCP, and sync it across devices — opt-in and end-to-end encrypted.
Use it when: your agent keeps relearning the same lessons.
Explore LocalMindA practical first run
Start small. Add the rest later.
Get a model running first. Tune it when you need more speed. Add the coding harness when you are ready to build.
Open the LocalBox setup guide- 1
Run it
Install LocalBox and launch a GGUF model.
- 2
Tune it
Let LocalBench find a stable profile for your machine.
- 3
Code with it
Point LocalPilot at the model and start working.
Why the harness matters
Same model. A much better loop.
The model stayed fixed. Every arm — raw inference, Claude Code, LocalPilot, and LocalPilot with reviewed memory — drove the same local model. This compares harnesses, not models: the Claude Code arm is Claude Code running that same local model, not a hosted frontier model.
25% 92%LocalPilot harness vs raw: +67 points
92% vs 88%LocalPilot without learning vs Claude Code
95% vs 88%LocalPilot with reviewed learning vs Claude Code
See the full comparison and methodology
How to read this
The benchmark covers 225 Aider-polyglot exercises in six languages, each graded in a network-isolated container. A 600-second timeout counts as unsolved.
LocalPilot's full run scored 92%, four points above Claude Code at 88%, before learning was enabled. Reviewed learning raised LocalPilot to 95%, extending that lead to seven points. Raw inference scored 25%.
The Claude Code arm is Claude Code driving the same local model through a compatible proxy — a harness-vs-harness comparison on one shared model, not a comparison against a hosted frontier model. Read the difference between harnesses, not the absolute score. This is one model and quant, and public-corpus absolutes may be affected by contamination.