Documentation

Everything you need to get started with Sakura.

Preview. Sakura is currently in preview. Features, pricing, and APIs may change. Report issues or feedback to support@sakura-ai.dev.

Getting Started

Install

macOS / Linux:

curl -fsSL https://cli.sakura-ai.dev/install | bash

Windows (PowerShell):

irm https://cli.sakura-ai.dev/install.ps1 | iex

Requires Node.js 20+.

Sign up

Create an account at sakura-ai.dev/signup.

Log in

sakura login --github

# or

sakura login --email you@example.com --password yourpassword

Start chatting

sakura-cli

Or equivalently: sakura chat or just sakura with no arguments.

Built-in Tools

Sakura has 11 tools that the AI calls automatically during conversation using structured function calling. You don't need to invoke them manually — just ask Sakura to do something and she'll pick the right tool.

ToolTrust LevelWhat It Does
readtrust working dirRead files or list directories
writetrustedCreate or overwrite files
shelltrustedRun shell commands
globtrust working dirFind files matching patterns
greptrust working dirSearch for regex patterns in code
codetrust read-onlyAnalyze code structure (symbols, exports)
awstrust read-onlyRun AWS CLI commands
introspecttrustedInspect Sakura's capabilities and settings
sessionapproval requiredManage conversation sessions
reportapproval requiredGenerate structured reports
subagentapproval requiredDelegate parallel subtasks

Trust Levels

  • trusted — runs automatically, no approval needed
  • trust working dir — auto-runs within your project directory, asks permission for paths outside it
  • trust read-only — auto-runs read operations, asks permission for mutations (e.g. AWS write commands)
  • approval required — always asks before executing

Tool Output

When Sakura uses a tool, you'll see what's happening:

Reading file: /Users/you/project/src/config.ts, from line 1 to end (using tool: read)

✓ Successfully read 2141 bytes from /Users/you/project/src/config.ts

Smart Model Routing

Sakura automatically routes each request to the cheapest appropriate model to minimize credit usage:

TaskModelCost
Classifications, greetingsGPT-4.1 Nano$0.10/1M tokens
AWS translations, summariesGPT-4.1 Mini$0.40/1M tokens
General conversation, code analysisGPT-5.2$1.50/1M tokens
Code generationGPT-5.2 Pro$3.00/1M tokens
You can override the model with /model in chat (shows an interactive picker) or sakura models to see all available models for your tier.

CLI Commands

CommandPurpose
sakuraStart interactive chat (default)
sakura chatStart interactive chat
sakura ask One-shot question
sakura loginLog in
sakura logoutLog out
sakura accountAccount info and usage
sakura modelsList available AI models
sakura usageDetailed usage + request history
sakura upgradeManage subscription
sakura initCreate .sakura/ config
sakura doctorCheck environment
sakura plan Generate a plan
sakura applyExecute a saved plan
sakura undoReverse last code patch
sakura code plan Generate code patch
sakura code apply Apply code patch
sakura aws AWS CLI passthrough
sakura mcp create [dir]Build MCP knowledge pack
sakura mcp listList knowledge packs
sakura --versionPrint version
sakura-cli is an alias for sakura — both start interactive chat when run with no arguments.

Slash Commands

Use these inside sakura chat:

CommandPurpose
/helpShow all commands
/read [s:e]Read a file or directory
/search Search project files
/statusGit status
/diff [file]Git diff
/logRecent commits
/branchList branches
/cd Change directory
/plan Multi-step plan
/toolBuilt-in tools
/parallel Parallel subtasks
/plannerRead-only planning mode
/image Attach image
/modelSelect model (interactive picker)
/modelsList available models with pricing
/run Run shell command
/trustTrust commands for session
/reset-trustRevoke session trust
/contextContext window usage
/compactSummarize history
/newFresh conversation
/chat resumeResume previous session
/quitExit
You can also say "start over" or "new conversation" naturally and Sakura will reset.

Safety & Approvals

Sakura uses a trust-based safety model:

  • Read-only operations run immediately (file reads, searches, AWS describe/list)
  • Mutating operations require your approval:
- [y] Allow once

- [n] Deny

- [t] Trust for session

AWS commands are classified automatically — read-only commands like describe-instances run freely, while create-bucket or terminate-instances require confirmation.

Code changes are generated as diffs, shown to you, and only applied after approval. Use sakura undo to reverse the last patch.

Error Recovery

When a command fails, Sakura will automatically try to diagnose and fix the issue — up to 2 attempts with escalating strategies. If it can't recover, it'll explain what went wrong and suggest what you can do manually.

Context & Conversations

Context Window

Use /context to see how much of the context window you've used. Sakura auto-compacts conversation history when it gets too large.

Sessions

Conversations are saved automatically. Use /chat resume to pick up where you left off, or /new to start fresh.

Context Switching

Sakura uses recent conversation history for intent classification, so you can switch topics naturally without getting stuck in a mode. If you mention "EC2" in passing, she won't suddenly try to run AWS commands.