Local-first agent runtime for MCP workflows with explicit trust controls and replayable runs.
LocalAgent helps you get from curiosity to a working local agent workflow without fighting provider setup, unsafe defaults, or opaque failures.
It is built for the hard part of local agents: connecting on-machine LLMs to MCP tools in a way that is guided, inspectable, and operationally clear.
Most friction in local agents is operational.
LocalAgent narrows that gap with:
- guided startup and provider auto-detection
- interactive TUI chat for local agent workflows
- explicit trust controls with approvals and audit trails
- replayable artifacts and inspectable event logs
- safe defaults with shell and write access disabled unless explicitly enabled
- built-in eval workflows and reviewable run outputs
- a beginner-friendly path without hiding advanced controls
Start a supported local provider, then launch LocalAgent in the project directory you want to work in.
# 1) Install from the repo root
cargo install --path . --force
# 2) Launch LocalAgent in the workspace you want to work in
localagentIf your provider starts after LocalAgent is already open, press R in the startup screen to refresh provider detection.
- Ollama
- LM Studio
- llama.cpp server
localagent --provider ollama --model llama3.2 --prompt "Summarize src/main.rs" runlocalagent --provider ollama --model llama3.2 chat --tuilocalagent doctor --provider ollama
localagent doctor --provider lmstudio
localagent doctor --provider llamacpplocalagent --trust on --provider ollama --model llama3.2 chat --tuiEnable shell and write tools only when you intentionally want side effects.
Global flags come before subcommands. See the examples above and the CLI reference for full command syntax.
LocalAgent is designed to make side effects explicit.
- shell and write access are disabled unless explicitly enabled
--allow-shell-in-workdiris a narrower shell mode than--allow-shell- trust mode can enforce policy and approvals
- persistent runs remain inspectable through artifacts and logs
The goal is not to remove every restriction. It is to make local agents usable without hiding risk.
Before running LocalAgent, start a supported local provider and make sure a model is available.
See: Provider setup
cargo build --releaseBinary output:
- Windows:
target/release/localagent.exe - Linux/macOS:
target/release/localagent
cargo install --path . --forcePrebuilt binaries are available in GitHub Releases.
For full install, updates, Windows troubleshooting, and verification steps, see:
Issues, feedback, and contributions are welcome.
Start here:
MIT