LetCode is an AI coding agent: it understands your project, edits code, runs commands, and passes tests — turning "fix this bug" into a replayable, resumable, reusable workflow. Built on a deeply re-engineered agent kernel, with models, tools and capabilities composed entirely through plugins.
LetCode runs as an independent agent process: it reads files, runs commands and calls tools on its own — every step is auditable.
Read files, edit code, run commands, search and call tools — one closed loop from start to finish. No need to babysit it step by step.
loop · tools · filesModels, tools, sandbox, sessions, storage and UI — all swappable and composable. A plugin-based kernel means you can extend anything without touching core source.
plugin · composableSystem prompts, reasoning, tool calls and results are all written to an append-only session log. Resume, fork, search and replay — all supported.
trajectory · resumeRead, write and command execution are approved tier by tier; critical actions always ask you first. Persistent permission whitelists balance safety and speed.
sandbox · permissionClaude, GPT, Gemini, DeepSeek, Qwen, Kimi and GLM all integrated — local models too. Hard reasoning and routine edits are auto-routed to the best model per task.
multi-model · routerLarge tasks are decomposed automatically and multiple sub-agents work in parallel: one edits code, one runs tests, one researches — then results are consolidated.
subagent · workflowFull-text search across session history; @file / @symbol inject context anytime. Resume, fork and replay — pick up where you left off.
memory · contextSkills teach the model workflows, MCP plugs in the external tool ecosystem, /commands define custom shortcuts. Capabilities are plug-and-play.
mcp · skills · commandsOne plugin = one new capability. The SDK exposes every extension point — models, tools, UI — so deep customization is never a black box.
sdk · pluginAn AI coding agent should be local and composable — owned by developers, not by any single vendor.
Swappable models, tools and sandbox. The agent is never locked into one vendor's cloud.
Runs fully local by default — your code never leaves your machine. You decide which model and what context to send.
Everything is a plugin: swap models, tools, UI or sandbox without touching the core. Your workflow, your rules.
Not a patch bolted on later — tiered permissions and isolation are built into the kernel.
Sessions and logs are stored locally in ~/.dsh/sessions by default — no forced cloud sync.
File reads, writes and command execution are approved tier by tier; critical actions ask you first, and whitelists persist.
Commands and scripts run in a controlled environment; out-of-bounds actions are blocked before they can touch your system.
Append-only session logs capture every step — prompts, reasoning, tool calls, results — inspectable anytime.
Three real scenarios showing how LetCode completes the full "understand → edit → verify" loop on its own. Click to switch.
fix the token-expiry bug on the login page. LetCode reads files, locates logic, edits code and runs tests — 14 tool calls in total, fully replayable.A purpose-built agent core with deep engineering: the core keeps only the Agent Loop; everything else is composed from plugins — swap models, tools or sandbox without touching the core.
One kernel, different ways to work: from a full-featured workbench to a minimal baseline — choose per task.
Full toolset: file editing, shell, search, skills, sub-agents and workflows. The default for daily development.
Inherits everything from Standard; the model can orchestrate multi-step tool calls directly in TypeScript for complex pipelines in one pass.
Keeps only the persistent shell and file editor, for model benchmarking and minimal-environment verification with reproducible results.
Runtime checks + a plugin test bench: craft your own presets and package them into new modes to share with your team.
No environment variables, no builds. Install Node.js, run one command and you're off.
The only runtime dependency. Grab the LTS build from nodejs.org or install via Homebrew.
$ brew install nodenpx pulls the agent kernel, boots the Web workbench and opens your browser.
$ npx @deepseek-ai/dsh webIn Settings, paste an API key for any supported model (Claude / GPT / DeepSeek / Qwen etc.), pick a working directory, and just talk.
$ echo 'ANTHROPIC_API_KEY=sk-ant-…' > ~/.dsh/.envConversational input is the default; slash commands and reference symbols are your speed boosters.
npx @deepseek-ai/dsh webStart the Web workbench — opens in your browserdsh --profile headless "task description"One-shot task mode, exits when done — great for CI & automation@file src/app.tsReference a project file in chat for precise context@symbol fetchUserReference a code symbol — jump to definitions & call sites/skillsLoad skill packs that teach the model your workflows/mcpConnect external tools: databases, browsers, internal systems, SaaS/agentsOrchestrate parallel sub-agent tasks — big jobs decompose automatically/shareShare a session with your team for seamless handoffAlso a coding agent — but LetCode stands on the side of composability: swappable models, extensible kernel, auditable trajectories.
Core local capabilities are free. Need hosted sessions and more task quota? Four tiers scale with usage — pay for what you use. Bring-your-own-key users pay model vendors directly; we never take a cut.
Daily development for individuals — about 40-60 full tasks, ideal for getting started with heavy use.
Power users and small teams starting out — three times Starter's quota, with early access to new models.
The workhorse for small teams (≤5 seats): shared context, collaboration and audit out of the box.
Growing teams (≤10 seats): a shared quota pool with transparent, controllable usage.
On-prem deployment, SSO integration, SLA guarantees and a dedicated customer success team. Built for mid-to-large teams with strict data boundaries.
Whether you're building a personal project, iterating with a team or automating pipelines, LetCode has a place for you.
Hand the repetitive work to the agent: fix bugs, add tests, write docs, refactor legacy code.
Sessions are saved per project with shared context; sub-agents work in parallel so review and development advance together.
Headless mode runs and exits, plugging into pipelines: auto-fix tests, auto-open PRs, auto-write changelogs.
Quick answers on models, privacy and environment.
~/.dsh/sessions. Only when you explicitly connect a remote model API is the relevant context sent to that model provider — the standard mechanism for all cloud models.Skills teach the model workflows, MCP connects external tools, /commands define custom shortcuts, and Subagent & Workflow split tasks in parallel. One new capability = one plugin — no core changes.Sessions, agent activity, files, Git changes, terminal, plans — all in one window, no more context-switching between tools.
The desktop app puts chat, execution and review in one interface: watch the agent work while reviewing every change it makes.
macOS / Windows clients: pick the build for your chip. Web needs zero install; CLI suits automation and CI.
Double-click to launch: auto environment check, key setup and browser opening.
macOS 13+ · Node.js 18+ requiredNative Windows client, x64. Install and go on Windows 10/11.
Windows 10+ · x64 · Node.js 18+ requiredZero install: one npx command starts a local service, open it in your browser. Cross-platform.
npx @deepseek-ai/dsh web Quick StartOne-shot task mode: runs, exits, prints to the terminal. Great for CI, scripts and automation.
dsh --profile headless "run the tests" Quick StartPlugins, MCP and community — every capability comes from an open ecosystem you can freely combine.
The community keeps shipping plugins — model adapters, tool integrations, UI themes. One plugin is one more capability.
Connect databases, browsers, internal systems and third-party SaaS via MCP — your whole toolchain, at its disposal.
GitHub discussions, issues and feature requests are all open. Report bugs, propose ideas.
The same capabilities run on macOS, Windows, Linux and the browser — switch between terminal, desktop and Web freely.
Local-first and free to start. Run it now and feel the efficiency of the full Agent Loop.