OpenAnalyst CLI — The Universal AI Agent for Your Terminal

Get Started

Install in seconds

One command. Native binary. No runtime dependencies.

Terminal

curl -fsSL https://raw.githubusercontent.com/OpenAnalystInc/cli/main/install.sh | bash

PowerShell

irm https://raw.githubusercontent.com/OpenAnalystInc/cli/main/install.ps1 | iex

Terminal

git clone https://github.com/OpenAnalystInc/cli.git
cd openanalyst-cli/rust

$ openanalyst login

  Select your LLM provider to authenticate:

  > OpenAnalyst         gpt-oss-120b free model or API key with credits
    Anthropic / Claude  opus, sonnet, haiku — API key
    OpenAI / Codex      gpt-4o, o3, codex-mini — API key
    Google Gemini       gemini-2.5-pro, flash — API key
    xAI / Grok
    OpenRouter          350+ models
    Amazon Bedrock

  How would you like to authenticate?

  > Use free model  gpt-oss-120b — no credits needed
    Use API key       OpenAnalyst API with credits

  Step 1  Connecting to OpenAnalyst... ✓ Connected

  ✓ Free model access configured

  Launch OpenAnalyst now? [Y/n]

  ✓ Login complete

Direct Provider Login — For Claude, Codex, and Gemini you can login directly through your provider account via browser OAuth. No API key needed — just sign in with your existing Anthropic, OpenAI, or Google account. Credentials are stored securely with PKCE and auto-refresh.

Multi-Provider

7 LLM providers. One interface.

Switch models mid-conversation. Session persists across providers. Login via browser API key. No other CLI does this.

OpenAnalyst DEFAULT

OPENANALYST_AUTH_TOKEN

gpt-oss-120b free model or API credits

Anthropic / Claude OAuth

opus, sonnet, haiku

openanalyst loginorANTHROPIC_API_KEY

OpenAI / Codex OAuth

gpt-4o, o3, codex-mini

openanalyst loginorOPENAI_API_KEY

Google Gemini OAuth

gemini-2.5-pro, flash

openanalyst loginorGEMINI_API_KEY

xAI / Grok

grok-3, grok-mini

XAI_API_KEY

OpenRouter

350+ models from any provider

OPENROUTER_API_KEY

Amazon Bedrock

Live discovery from gateway

BEDROCK_API_KEY

Mid-Conversation Provider Switching

> Explain this codebase                     # openanalyst-beta answers

> /model gpt-4o
  Model updated · Provider: OpenAI
  Session persisted (3 messages)              # conversation carries over

> Now fix the bug you found                 # gpt-4o sees full history

> /model gemini-2.5-pro
  Model updated · Provider: Google Gemini
  Session persisted (5 messages)              # 3 providers, 1 conversation

Capabilities

What you can do

The most feature-rich AI agent CLI ever built. Native binary, 7 LLM providers, 65+ commands, 24 tools, voice input, multi-agent orchestration, clipboard paste, /undo revert, and more.

⚡

Smart Per-Action Model Routing NEW

▶

Every prompt is auto-classified into explore, research, code, or write — each routed to the optimal model+effort. Explore tasks go to Haiku (1K thinking), coding goes to Opus (32K thinking). Save 80% on tokens without thinking about it.

/effort code max — per-category control
/route — view/edit the full routing table
/effort high — set all categories globally

★

Knowledge Base — Agentic RAG NEW

▶

/knowledge best Meta Ads strategy for D2C brands

Powered by BGE-M3 1024-dim embeddings on A100 GPU, PostgreSQL pgvector, and Neo4j knowledge graph. The pipeline:

1. Local MOE intent classification — local, zero latency (strategic, procedural, factual, comparative, diagnostic, etc.)
2. API call to hosted AgenticRAG with intent hint
3. Hybrid search — pgvector cosine + PostgreSQL FTS + Neo4j graph expansion
4. RRF fusion — merges results from all sources
5. KnowledgeCard — tabbed, collapsible results with abstracted category labels
6. Feedback — inline thumbs-up/down + /feedback corrections
7. Local cache — instant replay from .openanalyst/knowledge/

No raw course names exposed — results show "Ads Strategy", "AI & Machine Learning", etc.
Set OPENANALYST_API_KEY=oa_your_key to access.

⚡

Switch Models Mid-Conversation

▶

Switch between any LLM provider mid-conversation. /model now updates the orchestrator config and rebuilds the routing table in real-time. Session persists across provider boundaries. Start with Claude, continue with GPT-4o, finish with Gemini — all in one session.

█

Real Permission Enforcement NEW

▶

5-tier permission system: read-only, workspace-write, prompt, full, allow. Permission dialogs are real — the worker thread blocks on a oneshot channel until you Allow or Deny. Deny on timeout. Fail-safe, not fail-open. /permissions prompt to switch modes at runtime.

█

Full Terminal UI

▶

Default mode is a full-screen terminal UI with scrollable chat (mouse wheel + PageUp/PageDown from any mode), interactive sidebar with 5 sections (Agents, Files, Plans, Routing, Activity — press Tab to focus, j/k to navigate, Enter to activate), per-category model selection from all configured providers (Enter on a Routing row cycles through every available model from Claude, OpenAI, Gemini, xAI, and more), permission mode cycling (Ctrl+P cycles Default → Plan → Accept Edits → Danger with color-coded input border and right-side badge), rich diff rendering in tool call cards, animated status bar, vim-mode input, permission dialogs as real modals, input queue (type while streaming), slash command autocomplete, input history, voice input (Space bar), dynamic provider detection (auto-selects model from API keys — ANTHROPIC_API_KEY→claude-sonnet-4-6, OPENAI_API_KEY→gpt-4o), input box right-side badges showing [mode] [agent] [branch], and session auto-save.

⚙

Multi-Agent Orchestrator

▶

Spawn parallel sub-agents for complex tasks. Each agent has its own conversation runtime, tool permissions, and lifecycle. The orchestrator bridges sync execution to the async TUI via channels. Watch agents work in real-time in the sidebar panel.

✎

60+ Slash Commands

▶

Session management, git integration, AI planning, multimedia generation, web scraping, code analysis, plugin management, MCP server control, autonomous agents, and more. All fully wired — no stubs. Every command works with every provider. Type /help to see them all.

♬

Multimedia AI + Voice Input NEW

▶

/image — Generate images (Gemini Imagen, DALL-E, Stability AI)
/voice — Transcribe audio files (Whisper, Gemini)
/speak — Text-to-speech (OpenAI TTS)
/vision — Analyze images (Gemini, GPT-4o, Claude)
/diagram — Generate Mermaid diagrams

Live Voice Input — Press Space to activate microphone. Blue VU meter shows audio level. Transcription via Whisper or Gemini. Talk to your code. NEW

Paste & Drag-Drop — Paste file paths, image paths, audio paths, or multi-line text directly into the input box. File types auto-detected: [image], [audio], [video], [document], [file].

⚡

/openanalyst — Autonomous Agent NEW

▶

Inspired by Andrej Karpathy’s agent philosophy: simple loop, good model, basic tools, verifiable criteria. The autonomous agent runs a think→act→observe→verify loop without user interaction.

/openanalyst fix all failing tests --criteria "npm test"
/oa refactor auth to async --goal "all async" --criteria "npm run build"
/oa add caching --max-turns 20

Optional params: --goal (description), --criteria (shell command to verify), --max-turns (default 30). Criteria commands have 60s timeout. Turn budget enforced as hard limit.

⎇

MCP Protocol + LSP Integration NEW

▶

MCP (Model Context Protocol) — Connect external tool servers via Stdio and HTTP/SSE transports. JSON-RPC 2.0 compliant, proper Content-Length framing, multi-line JSON support. Tools auto-discovered and registered with mcp__ prefix. Execution dispatched to live server processes. /mcp add and /mcp remove manage servers directly.

LSP — Full Language Server Protocol client for diagnostics, go-to-definition, find-references. Tested with real mock servers.

Smart Context Injection — Prompts auto-analyzed for file paths and code keywords. Relevant files injected as context before the API call, reducing tool call round trips.

⎇

Git Integration

▶

Commit, branch, PR, issue creation, diff review — all from slash commands. /commit generates a message and commits. /pr drafts a pull request. /diff-review gets AI-powered code review of your changes.

★

22 Built-in Tools + Plugin System

▶

Bash, file read/write/edit, glob/grep search, web search/fetch, sub-agents, notebooks, REPL, PowerShell, and more. Plus a full plugin system — install, enable, disable, uninstall custom tools with /plugins install <path>. All backed by real PluginManager.

⚠

Production-Grade Reliability NEW

▶

Stream timeout — 30s per-event timeout detects hung connections and provider stalls
Retry with backoff — Exponential backoff (1s→2s→4s→8s, max 32s) on transient errors
Zero unsafe code — unsafe_code = "forbid" enforced at workspace level
340+ tests — Unit, integration, edge-case, and mock-server tests across all crates
No panics — All unwrap() in critical paths replaced with proper error handling
Crash recovery — Panic handler restores terminal and saves crash marker. Session auto-saves every 60s
Format-on-save resilience — Detects external file changes between consecutive edits (formatters, linters)
Bracketed paste — Multi-line paste handled as single event, no corruption
Zero silent failures — All async channel sends log on failure instead of silently dropping events
Hook system — 9 events: PreToolUse, PostToolUse, CwdChanged, FileChanged, SessionStart, SessionEnd, TaskCreated, Notification, Stop — all with allow/deny/warn. Manage via /hooks
Custom keybindings — All keys remappable via .openanalyst/keybindings.json

⚙

Cross-Platform Native Binary

▶

Single native binary. No Node.js, no Python, no Docker. Native on macOS (Intel + Apple Silicon), Linux (x64 + ARM), and Windows. Fast startup, low memory. ~18MB binary.

Reference

59 slash commands

All commands work identically regardless of which LLM provider you use. Every command is fully functional — zero stubs.

Command	Description
/help	Show all available commands
/status	Model, permissions, tokens, cost, session info
/cost	Cumulative token usage and cost breakdown
/model [name]	Show live models or switch mid-session
/clear	Start a fresh session
/compact	Smart compact — collapses tool calls, shows token budget %, suggests /clear when needed
/memory NEW	Show all memory files — instruction files, project memory, user memory with types
/effort [category] [level] NEW	Set thinking effort globally or per-category (explore/research/code/write)
/route [category] [tier] NEW	View or edit the per-action model routing table
/session [list\|switch]	List or switch saved sessions
/export [file]	Export conversation transcript
/version	Show CLI version and build info
/resume <path>	Load a saved session
/login NEW	Add or switch provider API keys from REPL
/logout NEW	Clear all saved credentials
/context NEW	Show context window usage, token counts, model
/vim NEW	Toggle vim keybinding mode
/whoami NEW	Show all logged-in providers and their status
/knowledge <query> NEW	Search the OpenAnalyst knowledge base
/explore <url> NEW	Smart-explore a GitHub repo from its history
/ask <question> NEW	Quick question — no tools, fast response (alias: /btw)
/user-prompt <msg> NEW	Inject a message with full tool access (alias: /up)

Command	Description
/diff	Show git diff of workspace changes
/commit	Generate commit message and commit
/commit-push-pr [ctx]	Commit, push, and open a PR
/pr [context]	Draft or create a pull request
/issue [context]	Draft or create a GitHub issue
/branch [list\|create\|switch]	Manage git branches
/worktree [list\|add\|remove]	Manage git worktrees
/teleport <symbol>	Jump to a file or symbol
/diff-review [file] NEW	AI-powered review of git diff

Command	Description
/image <prompt> NEW	Generate image (Gemini Imagen / DALL-E / Stability)
/voice <file> NEW	Transcribe audio/video (Whisper / Gemini)
/speak <text> NEW	Text-to-speech audio (OpenAI TTS)
/vision <image> [prompt] NEW	Analyze image (Gemini / GPT-4o / Claude)
/diagram <desc> NEW	Generate Mermaid diagram
/translate <lang> <text> NEW	Translate text to any language
/tokens [text] NEW	Estimate token count
/scrape <url> [selector] NEW	Fetch URL and extract text
/json <url> NEW	Fetch JSON API and pretty-print

Command	Description
/bughunter [scope]	Scan codebase for likely bugs
/ultraplan [task]	Deep multi-step planning with reasoning
/debug-tool-call	Replay last tool call with debug output
/think [prompt] NEW	Force extended thinking for next response
/changelog [since] NEW	Generate changelog from git log via AI
/doctor NEW	Diagnose installation, provider keys, MCP, workspace
/openanalyst <task> NEW	Autonomous agent — think→act→observe→verify loop
/swarm <task> NEW	Spawn a swarm of parallel agents for complex tasks

Command	Description
/config [section]	Inspect config (env, hooks, model, plugins)
/memory	Show loaded OPENANALYST.md files
/init	Create starter OPENANALYST.md
/permissions [mode]	Show or switch permission mode
/agents	List configured agents
/skills	List available skills
/plugin [list\|install\|enable\|disable]	Manage plugins
/mcp [list\|restart\|add] NEW	Manage MCP servers
/add-dir <path> NEW	Add directory tree to conversation context
Hooks
PreToolUse NEW	Runs before each tool execution — can allow, deny, or warn
PostToolUse NEW	Runs after each tool — format-on-save resilience built in
CwdChanged NEW	Fires when working directory changes (direnv, monorepo navigation)
FileChanged NEW	Fires when a file is modified externally after write/edit
SessionEnd NEW	Fires on session close for cleanup, logging, or archival
TaskCreated NEW	Fires when a task is created for external tracking systems
SessionStart NEW	Fires when a session begins — initialize environment, logging
Notification NEW	Fires when the agent wants to notify the user externally
Stop NEW	Fires when the agent stops execution — post-run cleanup
/hooks [list\|add\|remove\|test] NEW	Manage hooks interactively from the TUI

Command	Description
/dev install NEW	Install Playwright for browser automation
/dev open <url>	Navigate browser to URL
/dev screenshot	Capture screenshot of current page
/dev snap	Snapshot accessibility tree
/dev click <ref>	Click element by a11y reference
/dev type <ref> <text>	Type text into element
/dev test <desc>	Generate Playwright test via AI
/dev codegen	Record browser actions as test code
/dev stop	Close browser session
/dev status	Show Playwright version and state

Tools

22 built-in tools

The AI agent uses these automatically during conversations. Works with every provider. Plus MCP protocol for unlimited external tools.

Bash

Execute shell commands and scripts

ReadFile

Read files with line numbers and offsets

WriteFile

Create new files with content

EditFile

Precise string replacements

GlobSearch

Find files by pattern

GrepSearch

Search file contents with regex

WebSearch

Search the web for answers

WebFetch

Fetch content from URLs

Agent

Spawn sub-agents for parallel tasks

TodoWrite

Track tasks within a session

NotebookEdit

Edit Jupyter notebook cells

Skill

Invoke registered skills

ToolSearch

Search for specialized tools

REPL

Execute code in subprocess

PowerShell

Execute PowerShell on Windows

Config

Get/set settings programmatically

Sleep

Wait between operations

StructuredOutput

Return structured JSON

SendMessage

Send messages during execution

Architecture

Modular, testable, zero-dependency binary. Built with ecosystem crates, not reinvented.

src/modules/

api/                 # Multi-provider API client (7 providers)
commands/            # 59 slash commands
events/              # Shared TUI ↔ backend event types
orchestrator/        # Multi-agent lifecycle + channel bridge
tui/                 # Full-screen TUI application
tui-widgets/         # Markdown, tool cards, input, spinner
runtime/             # Conversation engine, session, MCP
tools/               # 19 built-in tool implementations
plugins/             # Plugin system (install, enable, hooks)
openanalyst-cli/     # Binary entry point
openanalyst-agent/   # Headless autonomous agent runner
server/              # HTTP/SSE server (axum)
lsp/                 # Language Server Protocol integration
compat-harness/      # Upstream manifest extraction

Ecosystem Crates

tui-markdown — Markdown rendering
Vim Engine — Built-in modal editor
tui-tree-widget — File tree
throbber-widgets-tui — Spinner
syntect-tui — Syntax highlighting

Key Metrics

14 crates in workspace
7 LLM providers
59 slash commands
22 built-in tools
340+ tests
MIT license

Config

Project configuration

OPENANALYST.md	Project-specific AI instructions (auto-detected in parent dirs)
.openanalyst.json	Shared project defaults (permissions, model, tools)
.openanalyst/settings.json	Project-level settings
.openanalyst/settings.local.json	Machine-local overrides (gitignored)
~/.openanalyst/.env	Provider API keys and environment config (auto-created)
~/.openanalyst/credentials.json	Saved provider API keys
.openanalyst/sessions/	Saved conversation sessions
.openanalyst/skills/	Custom skill definitions
.openanalyst/commands/	Custom slash commands

Legal

Credits & Trademarks

OpenAnalyst CLI is an independent, open-source product by OpenAnalyst Inc.

All provider trademarks belong to their respective owners:
Claude — Anthropic, PBC · GPT, DALL-E, Whisper, Codex — OpenAI, Inc · Gemini, Imagen — Google LLC · Grok — xAI Corp · Bedrock — Amazon Web Services, Inc · OpenRouter — OpenRouter, Inc

Use of these providers' APIs is subject to each provider's Terms of Service. OpenAnalyst CLI facilitates access to these APIs — it does not claim ownership of any provider's services, models, or intellectual property. Licensed under MIT.