Why another coding agent
Every coding agent runs one model in one loop and asks it to review its own work. localcode is built around what that loop cannot do.
Reviewers that do not share a brain. /debate hands your change to up to three reviewers, each a separate agent on its own model, none able to see the others. Nothing ships until all of them say yes. Put Claude, a local Qwen and GPT on the panel and you get disagreement instead of a model agreeing with itself.
Work that runs when you said, not when you sent it. tomorrow at 9am run the full test suite and summarise only the failures is booked, the time parsed by localcode rather than guessed at by the model. Korean works too. A time that passes while localcode is closed is reported as missed, not run late.
Plans checked before a token is spent. Orchestrate makes a multi-agent plan structured data. Every agent named, every stage reference and every count is validated whole, and a plan that would not work is refused with the reason, not narrated and abandoned halfway. A plan that worked is saved and run again by name, and a run that stopped is resumed without paying again for the agent turns that finished.
Any model, hosted or on your own machine, in one conversation. Bedrock, the Anthropic API and anything OpenAI-compatible (LM Studio, vLLM, llama.cpp, Ollama), with switching mid-conversation. Local runtimes get both spellings of the reasoning field, the context window the server actually loaded, and failover to another profile.
You can see what the model will see. /context reports the next request piece by piece (source, trust class, token estimate) without a model call. Another conversation reaches this one only through #name, as a tool result the model asked for, never spliced into your prompt.
On the same model
Twenty-five SWE-bench Verified instances, one model, four agents. Claude Code resolved 19; localcode resolved 19 with Smart Agent off and 21 with it on. That run used a version in which Smart Agent also brought the specialists and delegation that now come with Orchestrate. That spread is inside run-to-run variance, so the benchmark does not rank them. What it shows is that pointing an agent at your own model, with reviewers and schedules and validated plans on top, gives nothing up.
It reads what you already have
CLAUDE.md and AGENTS.md with @path imports; skills and commands from .localcode, .claude, .opencode and opencode's own ~/.config/opencode, all merged, with .localcode winning a shared name; your Claude Code MCP servers via localcode mcp import-claude or by listing .mcp.json; an opencode.json as it is, when you list it under include. One daemon serves a terminal client, a browser client and a native window at once. Run it on the box with the GPU, attach from a laptop.
Install
One static binary, no root, nothing left outside $HOME:
curl -fsSL https://raw.githubusercontent.com/dennis2lee/localcode/main/scripts/install.sh | sh
The Windows MSI, the macOS .app and the Debian packages are on the releases page. Every other way, and each platform's first-run steps, is in INSTALL.md.
First run
-
Write ~/.localcode/config.json (on Windows, %USERPROFILE%\.localcode\config.json). It needs one provider, one profile, and default_profile. Use the block for the model you have.
Anthropic API key. Run localcode login anthropic first. It saves the key and prints this file.
{
"providers": { "anthropic": { "type": "anthropic" } },
"profiles": { "main": { "provider": "anthropic", "model": "claude-sonnet-5" } },
"default_profile": "main"
}
Amazon Bedrock. Needs AWS credentials and model access in the region. With credentials from aws configure or aws sso login, the block below is complete. localcode login bedrock signs in through IAM Identity Center without the AWS CLI and saves an AWS profile: add "profile": "localcode-bedrock" to the provider then. See MODELS.md.
{
"providers": { "bedrock": { "type": "bedrock", "region": "us-west-2" } },
"profiles": { "main": { "provider": "bedrock", "model": "us.anthropic.claude-opus-4-6-v1", "max_tokens": 8192 } },
"default_profile": "main"
}
A local server. LM Studio listens on http://localhost:1234/v1, Ollama on http://localhost:11434/v1. model is the name the server lists. A model that reads images also needs "vision": true on its profile.
{
"providers": { "local": { "type": "openai-compat", "base_url": "http://localhost:1234/v1" } },
"profiles": { "main": { "provider": "local", "model": "qwen3-30b-a3b" } },
"default_profile": "main"
}
-
Start localcode in the project it should work on: cd ~/src/your-project && localcode. The directory it starts in is the workspace, where the model reads and edits files.
-
Type a message and press Enter. A file write or a shell command asks first: y allows once, s for this session, a always, n denies. git commands run without asking, except one that makes git run another program, such as git rebase --exec.
-
Open http://127.0.0.1:4096 for the browser client on the same conversations. Keep the terminal running.
localcode installs a newer release when it starts. Set "auto_update": false in config.json to stay on the installed version. config.example.json is the full key reference: every key, with a note on each.
Read on
| |
|---|
| Where localcode differs | Fourteen capabilities, the use case for each, and what is deliberately the same as elsewhere · 한국어 |
| Coding agents on one model | The SWE-bench run above: setup, per-instance results, and the one instance every agent failed |
| INSTALL.md | Installing on each platform, updating, uninstalling, and building from source |
| MODELS.md | Setting up each provider, with verified model IDs |
| FEATURES.md | Everything localcode ships, one line each |
| USAGE.md | The manual: configuration, commands, sessions, agents |
| API.md | The HTTP and SSE API the front ends use |
| CHANGELOG.md | What changed, and what the reviews of it found |
MIT. Packages are unsigned; the installer verifies SHA-256. The native window is experimental: releases ship it in the Windows MSI, the macOS .app opens the terminal client in Terminal, and Linux has the daemon, terminal and browser clients.
