Skip to main content
Sturm TechnologiesSturm
Explore Fast Series
runtime/local-first/MCP

LocalAgent

Run an AI coding agent on your own machine, with your own local models. No cloud account, no code leaving your computer.

LocalAgent by Sturm Technologies is an open-source agent runtime that connects on-machine LLMs (Ollama, LM Studio, llama.cpp) to tools through MCP. It handles the hard operational part: explicit trust controls before anything writes or runs, replayable runs, and logs you can actually inspect afterward.

View on GitHubReleasesBack to Projects
  • Public repo
  • Rust
  • MCP workflows
  • Local-first
  • Alpha

Alpha-stage developer tooling · MIT License · Built in Rust

Why it exists

The friction is operational, not the model

Most local-agent friction is not the model. It is setup, trust, tool access, recovery, and knowing what actually happened after a run. LocalAgent focuses on those operational edges: provider detection, guided startup, explicit approvals, replayable artifacts, and logs that can be inspected after the fact.

Guided startup

Detects supported local providers and gives a clearer path from setup to a working run.

Explicit trust controls

Shell and write access stay gated. Side effects require intentional configuration and approval paths.

Replayable runs

Artifacts, event logs, and tool outputs are preserved so runs can be inspected instead of guessed at.

Built-in evaluation workflows

LocalAgent includes evaluation and validation paths to make agent behavior easier to test and improve.

Supported local providers

Built around local and OpenAI-compatible providers

On startup LocalAgent probes for on-machine model servers, so you know what is reachable before a run begins.

localagent doctorprobing endpoints
  • Ollamallama3.1:8b · localhost:11434reachable
  • LM StudioOpenAI-compatible · localhost:1234reachable
  • llama.cpp serverOpenAI-compatible endpointnot running

Safety model

Useful without hiding risk

The goal is not to remove every restriction. The goal is to make local agents useful without hiding risk. LocalAgent keeps dangerous capabilities explicit, routes tool access through gates, and preserves evidence from runs so the operator can review what happened.

  • Shell and write access disabled unless explicitly enabled
  • Narrower workdir-scoped shell mode available
  • Approval and audit paths for trusted operation
  • Persistent run artifacts and inspectable logs
  • Conservative validation and repair behavior

Developer workflow

Install, run, and check your provider

localagent
# install from a cloned repo
cargo install --path . --force
localagent

# verify a provider is reachable
localagent doctor --provider ollama
localagent doctor --provider lmstudio
localagent doctor --provider llamacpp
Read the install guide

Current release focus

Reliable, inspectable coding-agent runs

Recent LocalAgent work focuses on making coding-agent runs more reliable and inspectable: structured planning, replayable tool-result artifacts, read-only LSP tools, bounded validator-driven repair, and ordered Rust validation outcomes. The latest tagged release is v0.6.0-alpha.1, an alpha prerelease.

Release feed on GitHub

Good fit for

Who gets the most out of it

  • Developers experimenting with local coding agents
  • Builders testing MCP tool workflows
  • People who want inspectable agent runs
  • Local-first AI workflows where trust and reproducibility matter

Not trying to be

Scope, stated plainly

LocalAgent is not a hosted SaaS agent, not an unrestricted autonomous background worker, and not a replacement for judgment. It is a runtime for controlled local workflows where the operator can see, approve, replay, and debug what happened.

Run local agents with more control

LocalAgent is public, actively developed, and built around practical trust boundaries for local AI tooling.

View LocalAgent on GitHubBrowse all projects