Cortex

AI-first multi-model terminal. One interface, every model, no switching.

 ██████╗ ██████╗ ██████╗ ████████╗███████╗██╗  ██╗
██╔════╝██╔═══██╗██╔══██╗╚══██╔══╝██╔════╝╚██╗██╔╝
██║     ██║   ██║██████╔╝   ██║   █████╗   ╚███╔╝
██║     ██║   ██║██╔══██╗   ██║   ██╔══╝   ██╔██╗
╚██████╗╚██████╔╝██║  ██║   ██║   ███████╗██╔╝ ██╗
 ╚═════╝ ╚═════╝ ╚═╝  ╚═╝   ╚═╝   ╚══════╝╚═╝  ╚═╝

AI first terminal intertwined with shell second by Orellius.ai Built by Orel

Screenshots


Welcome	Context-aware AI

ASCII welcome, unified black UI, session tabs	Reads cwd, git status, CLAUDE.md, file tree
Local model (Nemotron)	Shell + AI in one flow

Local models are context-aware too	`!` prefix runs commands inline
Slash commands + markdown	Rich markdown rendering

`/help`, tables, code blocks, LaTeX	Headers, bullets, inline code, bold
Markdown preview sidebar	Settings (9 pages)

`/preview` opens rich .md viewer, click any .md file to open	Models, Providers (26+), Routing, Budget, Permissions, MCP, Appearance
Project launcher

Cmd+K to switch projects, Browse for any directory, Recents tab

How Cortex compares

Feature	iTerm2	Warp	Cursor	Claude Code	Cortex
Terminal (real PTY)	Yes	Yes	No (IDE)	Yes	Yes
AI built-in	No	1 model	1-2 models	1 (Claude)	Multiple, routed
Multi-model routing	No	No	No	No	Yes (complexity scored)
Local models (Ollama)	No	No	No	No	Yes (free, private)
MCP tool bridge	No	No	No	Yes	Yes (shared across all models)
Cost tracking	No	No	No	No	Yes (budget cap + local fallback)
Streaming responses	N/A	Yes	Yes	Yes	Yes
Context-aware (git, cwd)	No	Partial	Yes	Yes	Yes
Split panes	Yes	No	Yes	No	Yes
Open source	No	No	No	Yes	Yes (Apache 2.0)
Native desktop	Yes	Yes	Yes	CLI	Yes (Tauri)
Price	Free	Free/$15	$20/mo	Subscription	Free

Supported Providers (26+)

Cortex supports multiple AI providers through a unified routing layer. Most cloud providers use OpenAI-compatible SSE endpoints, making integration straightforward.

Cloud Providers

Provider	Models	Pricing (cheapest)	Auth	Streaming	Tools	Free Tier
OpenAI	GPT-5, GPT-4.1, o3, o4-mini	$0.25/1M in	API Key	SSE	Yes	Trial credits
Anthropic	Claude Opus 4.6, Sonnet 4.6, Haiku 4.5	$1.00/1M in	API Key	SSE	Yes	No
Google Gemini	2.5 Pro, 2.5 Flash, 2.5 Flash Lite	$0.10/1M in	API Key	SSE	Yes	Yes
Mistral	Large 3, Small 3, Codestral, Nemo	$0.02/1M in	API Key	SSE	Yes	Rate-limited
DeepSeek	V3, V4, R1 (reasoning)	$0.14/1M in	API Key	SSE	Yes	No
Groq	Llama 4 Scout, gpt-oss-20B/120B	$0.05/1M in	API Key	SSE	Yes	Yes
xAI (Grok)	Grok 4, Grok 4.1 Fast	$0.20/1M in	API Key	SSE	Yes	$25 credit
Together AI	200+ models (Llama, Qwen, Mistral)	$0.10/1M in	API Key	SSE	Yes	$1 credit
Cerebras	Llama 8B/70B/405B (fastest inference)	$0.10/1M in	API Key	SSE	Limited	24M tok/day
Cohere	Command R+, Command A	$0.15/1M in	API Key	SSE	Yes	1K req/mo
Perplexity	Sonar, Sonar Pro (search-augmented)	$1.00/1M in	API Key	SSE	No	No
Fireworks AI	100+ open models, fast inference	$0.20/1M in	API Key	SSE	Yes	No
Amazon Bedrock	Claude, Nova, Llama, Mistral	$0.035/1M in	AWS IAM	NDJSON	Yes	No
Azure OpenAI	GPT-4.1, o3, o4-mini	$1.10/1M in	API Key	SSE	Yes	No
Alibaba (Qwen)	Qwen Max, Plus, Turbo	$0.05/1M in	API Key	SSE	Yes	Yes (29 models)
AI21 (Jamba)	Jamba 1.6 Mini, Large 1.7 (256K ctx)	$0.25/1M in	API Key	SSE	Yes	Trial credits
Reka AI	Flash 3, Core (multimodal)	$0.20/1M in	API Key	SSE	Limited	No
Replicate	1000+ community models	Per GPU-sec	API Key	SSE	Rare	No

Local Providers (free, private)

Provider	Endpoint	Streaming	Auth	Notes
Ollama (built-in)	localhost:11434	NDJSON/SSE	None	100+ models, vision, tools
LM Studio	localhost:1234	SSE	None	GUI-based, OpenAI + Anthropic compat
llama.cpp	localhost:8080	SSE	None	CPU-first, used by Ollama/LM Studio
vLLM	localhost:8000	SSE	None	Production-grade, multi-GPU
LocalAI	localhost:8080	SSE	None	35+ backends, MCP native
Jan.ai	localhost:1337	SSE	None	GUI, also routes to cloud APIs

What it is

Cortex is a native desktop terminal that puts Claude and local Ollama models alongside a real shell in one window. You type in a single input. The router decides which model to call based on what you wrote. Responses stream token-by-token. All models share the same MCP tool bridge.

There is no chat bubble UI, no sidebar with a robot avatar. Just a terminal.

Features

Multi-model AI

Three tiers run simultaneously. The router picks the right one automatically.

Provider	Models	When used
Claude (CLI)	sonnet (default), opus, haiku	Code, debugging, implementation
Claude Sonnet	sonnet	Research, explanation, comparison
Ollama	any local model (qwen, nemotron, deepseek, llama)	Simple queries, local-only, budget cap fallback

Force a provider with a prefix:

c: implement a binary search in Rust
s: explain the difference between Arc and Rc
l: what does grep do

Automatic routing uses deterministic complexity scoring (no ML). Keywords like implement, fix, refactor score high and route to Claude. Questions like what is or explain route to Sonnet. Short lookups route to local.

Streaming

Responses stream token-by-token from both Claude CLI and Ollama. Claude streams via stdout pipe reading, Ollama via NDJSON SSE. No waiting for a complete response before rendering.

Context awareness

Before every AI query, Cortex reads:

Current working directory (where you launched Cortex from)
Git branch and git status --short
Top-level project files
Stack detection (Rust, TypeScript, Python, Tauri)
CLAUDE.md if present in the project root

This context is injected automatically. You do not need to paste file contents into prompts.

Conversation memory

The last 20 messages per session are persisted to SQLite and included as history on every query. Conversations survive tab switches and app restarts.

Shell execution

Prefix any message with ! to run it as a shell command inside the AI chat:

!ls -la
!git status
!cargo check

Output appears inline, not in a separate terminal pane.

Slash commands

Command	Action
`/clear`	Clear current conversation
`/help`	Show available commands
`/model`	Show routing info
`/settings`	Open settings overlay
`/search`	Activate search
`/palette`	Open command palette
`/budget`	Show today's spend and daily cap

LaTeX math rendering

Model responses containing LaTeX notation (\[...\], $...$, $$...$$) render as formatted equations via KaTeX.

Verification gate

Local model outputs pass through a verification layer. If the output fails consistency checks (garbage, hallucination signals, AI-speak), a warning is prepended. The gate is smart enough to allow short answers when you ask short questions.

Budget metering

Claude CLI calls are cost-estimated and logged to SQLite. When the daily budget cap is reached, the router falls back to Ollama for all queries. Default daily cap: $5.00. Local models and CLI subscriptions are free.

MCP Bridge (Phase 6)

Cortex runs a Model Context Protocol bridge that starts configured MCP servers on launch via JSON-RPC over stdio. All models share the same tool registry. MCP servers are configured in ~/.cortex/mcp.toml and can be imported from your Claude Code settings.json.

Terminal

Full PTY via portable-pty (real shell, not a subprocess wrapper)
xterm.js v6 with DOM renderer
Split panes: Cmd+D (vertical), Cmd+Shift+D (horizontal)
Each pane gets its own PTY process

Tab management

Session numbering: 1# AI, 2# Shell, 3# AI
Double-click tab to rename
Drag to reorder
Cmd+Shift+T to reopen closed tabs (10-tab recovery stack)
Session state persists across quit and relaunch

Search

Cmd+F activates search in both AI chat (text filter with opacity dimming) and shell (xterm addon-search with match highlighting).

Command palette

Cmd+Shift+P opens a fuzzy command palette listing all available actions with keyboard navigation.

Global hotkey

Ctrl+`` toggles the window from anywhere on the system (quake-style drop-down).

Settings (9 pages)

Sidebar layout with lucide-react icons. Opened with Cmd+, or /settings.

Tab	Contents
Models	Per-role model assignment, auto-optimize, CLI detection
Providers	API keys, Ollama endpoint, connection testing
Routing	Complexity scoring thresholds, override rules
Budget	Daily cap, spend history, cost breakdown
Permissions	Safe / Ask / Auto / Bypass modes
MCP Servers	Add/remove/toggle, import from Claude Code
Appearance	Font size, font family, cursor style, accent color, opacity
Shortcuts	Keyboard shortcut reference
About	Version, stack, config paths, links

Toast notifications

Update notifications with action buttons (Update / Skip). Used for model updates, MCP server changes, and system alerts.

Keyboard shortcuts

Shortcut	Action
`Cmd+T`	New AI tab
`Cmd+W`	Close tab
`Cmd+1` to `Cmd+9`	Switch to tab N
`Cmd+D`	Split pane vertically
`Cmd+Shift+D`	Split pane horizontally
`Cmd+F`	Search in current pane
`Cmd+Shift+P`	Command palette
`Cmd+,`	Open settings
`Cmd+K`	Project launcher
`Cmd+Shift+H`	Paste history
`Cmd+Shift+T`	Reopen closed tab
Ctrl+``	Toggle window (global)
`Enter`	Submit query
`Shift+Enter`	New line in input

Installation

Prerequisites

macOS 13+ (Ventura or later)
Rust 1.77.2+
Node.js 20+
pnpm 10+
Tauri CLI v2

For AI features (optional but recommended):

Claude Code CLI for Claude provider
Ollama for local model provider

Setting up AI providers

Claude (via CLI subscription)

# Install Claude Code CLI
npm install -g @anthropic-ai/claude-code

# Login (opens browser for OAuth)
claude login

# Verify
claude --version

Cortex detects Claude automatically. No API key needed. Uses your existing Claude Pro/Max subscription.

Ollama (local, free)

# Install from ollama.ai or via Homebrew
brew install ollama

# Start the server
ollama serve

# Pull a model
ollama pull qwen3:8b        # fast, good for quick queries
ollama pull nemotron:12b     # strong reasoning
ollama pull deepseek-r1:32b  # heavy, best quality

Cortex connects to http://localhost:11434 by default. Change the endpoint in Settings > Providers.

Project directory for launcher

The project launcher (Cmd+K) scans for git repos. Set CORTEX_PROJECTS_DIR to your projects folder:

export CORTEX_PROJECTS_DIR=~/Projects

Or it auto-detects ~/Projects, ~/Desktop/Projects, ~/Developer, or ~/dev.

Build from source

git clone https://github.com/Orellius/cortex-terminal
cd cortex
pnpm install
cargo tauri dev

Production build:

cargo tauri build

The .dmg installer outputs to src-tauri/target/release/bundle/dmg/.

Configuration

All config lives in ~/.cortex/. Created automatically on first launch.

`~/.cortex/config.toml`

claude_model = "sonnet"
ollama_model = "nemotron-cascade-2"
ollama_endpoint = "http://localhost:11434"
daily_budget_usd = 5.0
permission_mode = "ask"

`~/.cortex/mcp.toml`

[[servers]]
name = "filesystem"
command = "npx"
args = ["-y", "@modelcontextprotocol/server-filesystem", "~/"]
enabled = true

`~/.cortex/identity.md`

Optional. Write your name, role, preferred coding style, or project context. This is prepended to every AI query as identity injection. All models read this file and respond as "Cortex".

Architecture

cortex/
├── src/                        # React 19 frontend
│   ├── components/
│   │   ├── ai/                 # AiChatView, AiMessage, AiChatInput
│   │   ├── settings/           # 9-page settings (sidebar layout)
│   │   ├── TabBar.tsx          # Tab management, drag reorder, rename
│   │   ├── StatusBar.tsx       # Git branch, cost, Claude usage
│   │   ├── CommandPalette.tsx  # Fuzzy action search
│   │   └── Toast.tsx           # Notification system
│   ├── hooks/                  # useTerminal, useTabs, useKeyboard
│   └── types.ts
│
└── src-tauri/                  # Rust backend
    └── src/
        ├── ai/
        │   ├── router.rs       # Complexity-scored query routing
        │   ├── providers.rs    # Claude CLI + Ollama streaming
        │   ├── brain.rs        # Context injection (cwd, git, CLAUDE.md)
        │   ├── mcp.rs          # MCP bridge (JSON-RPC stdio)
        │   ├── verification.rs # Output verification gate
        │   ├── database.rs     # SQLite persistence
        │   └── config.rs       # config.toml + mcp.toml
        ├── commands/           # Tauri IPC handlers
        ├── pty.rs              # PTY manager (one thread per pane)
        └── lib.rs              # App setup, state registration

How it works

You type a query in the terminal input
The router scores it (0-10) based on word signals
Score 5+ goes to Claude, 3-4 to Sonnet, 0-2 to local Ollama
The provider streams tokens back via cortex:ai:stream events
Context (cwd, git, CLAUDE.md) is injected into every prompt automatically
MCP tools are discovered on startup and injected into system prompts

The frontend never talks to AI providers directly. Every call goes through Rust. The backend owns all provider state, cost tracking, and verification.

Stack

Layer	Technology
Desktop framework	Tauri 2.0
Backend	Rust
Frontend	React 19 + TypeScript
Terminal	xterm.js v6 (DOM renderer)
Database	SQLite (rusqlite, bundled)
Icons	lucide-react
Math	KaTeX
Build	Vite 8 + Cargo
Package manager	pnpm

Contributing

See CONTRIBUTING.md for guidelines. Good first issues are labeled and ready to pick up.

If Cortex is useful to you, give it a star. It helps others find the project.

Attribution

Cortex Terminal is created by Orellius Labs.

Attribution is required under the Apache 2.0 License (Section 4). If you fork or redistribute this software:

You must retain the NOTICE file with original attribution
You must retain the LICENSE file
You must state any changes you made to the source code
You must not use the "Cortex" or "Orellius" names, logos, or trademarks without permission (Section 6)
The attribution credit is compiled into the binary and verified at runtime. Removal constitutes a license violation.

Forks are welcome. Rebranding as your own product without attribution is not.

License

Apache License 2.0. See LICENSE and NOTICE.

Built by Orel / @Orellius on X / orellius.ai

Name		Name	Last commit message	Last commit date
Latest commit History 106 Commits
.github/workflows		.github/workflows
cli		cli
docs/screenshots		docs/screenshots
src-tauri		src-tauri
src		src
.gitignore		.gitignore
CLAUDE.md		CLAUDE.md
CONTRIBUTING.md		CONTRIBUTING.md
LICENSE		LICENSE
NOTICE		NOTICE
README.md		README.md
index.html		index.html
package.json		package.json
pnpm-lock.yaml		pnpm-lock.yaml
sqlite_mcp_server.db		sqlite_mcp_server.db
tsconfig.json		tsconfig.json
vite.config.ts		vite.config.ts

Folders and files

Latest commit

History

Repository files navigation

Cortex

Screenshots

How Cortex compares

Supported Providers (26+)

Cloud Providers

Local Providers (free, private)

What it is

Features

Multi-model AI

Streaming

Context awareness

Conversation memory

Shell execution

Slash commands

LaTeX math rendering

Verification gate

Budget metering

MCP Bridge (Phase 6)

Terminal

Tab management

Search

Command palette

Global hotkey

Settings (9 pages)

Toast notifications

Keyboard shortcuts

Installation

Prerequisites

Setting up AI providers

Build from source

Configuration

~/.cortex/config.toml

~/.cortex/mcp.toml

~/.cortex/identity.md

Architecture

How it works

Stack

Contributing

Attribution

License

About

Topics

Resources

License

Contributing

Uh oh!

Stars

Watchers

Forks

Releases 1

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

`~/.cortex/config.toml`

`~/.cortex/mcp.toml`

`~/.cortex/identity.md`

Packages