OxideSens AI

OxideSens is OxideTerm’s built-in AI assistant that helps you work more efficiently in the terminal. It features 40+ autonomous tools, MCP server integration, RAG knowledge base, and supports both sidebar and inline chat modes.

Bring Your Own Key

OxideSens follows a BYOK (Bring Your Own Key) model. You provide your own API key for any supported AI provider. Keys are securely stored in your OS keychain — never in config files.

Supported Providers

• OpenAI — GPT-5.4, GPT-5.4 mini, GPT-5.4 nano, etc.
• Anthropic — Claude Opus 4.6, Claude Sonnet 4.6, Claude Haiku 4.5, etc.
• Google — Gemini 3 Flash, Gemini 2.5 Pro, Gemini 2.5 Flash, etc.
• DeepSeek — V3.2, etc.
• Ollama — any local model (Llama, Mistral, Qwen, etc.)
• OneAPI — self-hosted API gateway
• Any /v1/chat/completions compatible endpoint

Models are fetched dynamically from provider APIs — new models are available automatically without app updates.

Setup

1. Open Settings → AI tab
2. Enable AI Capabilities
3. Read and confirm the privacy statement
4. Enter your API base URL (e.g., https://api.openai.com/v1)
5. Enter your API key — stored in OS keychain

Sidebar Chat

The sidebar chat panel provides a persistent conversational interface with full history:

• Ask questions about commands, scripts, or system administration
• Get explanations of error messages and log analysis
• Request command suggestions for specific tasks
• Multi-source context — IDE files, SFTP paths, Git status automatically available as context
• Full chat history with session persistence

Terminal Context

OxideSens captures terminal buffer from the active pane or all split panes simultaneously. It can analyze:

• Recent command output
• Error messages and stack traces
• Log files visible in the terminal
• Running process output

Context is injected into a structured prompt with environment information (local OS, remote OS if detected via SSH).

Inline Panel

Press ⌘I (macOS) or Ctrl+Shift+I (Windows/Linux) to open the inline panel — a lightweight, floating command assistant:

• VS Code-style floating panel — appears at the cursor position, 520px wide
• AI-suggested commands — inserted via bracketed paste (safe, no auto-execution)
• Output analysis — select error text, press ⌘I, ask “what went wrong?”
• One-click actions: Insert (paste into terminal), Execute (run immediately), Copy, Regenerate

Inline Panel Shortcuts

Key	Action
Enter	Send question; with result: execute command
Tab	Insert AI suggestion into terminal
Esc	Close panel

Smart Positioning

The panel follows the terminal cursor position:

• Prefers display below the cursor
• Auto-switches to above when space is insufficient
• Horizontal auto-adjustment for screen edges

Note

The inline panel is designed for quick, lightweight queries without tool invocation. For 40+ autonomous tools, MCP server extensions, or RAG knowledge base search, use the sidebar chat.

40+ Autonomous Tools

OxideSens can autonomously invoke tools without manual triggering:

Category	Examples
File Operations	Create, read, write, move, delete files and directories
Process Management	List processes, kill processes, check resource usage
Network Diagnostics	Check ports, DNS resolution, connectivity tests
TUI Interaction	Send keys to running TUI apps (vim, htop, yazi)
Text Processing	Search, replace, extract, transform text content
System Info	Disk usage, memory, uptime, OS details

Tools are invoked through a structured function-calling interface with the AI model.

MCP Integration

Connect external Model Context Protocol servers for third-party tool integration:

• stdio transport — launch a local MCP server process
• SSE transport — connect to a remote MCP server via Server-Sent Events
• Configure MCP servers in OxideTerm settings
• MCP tools appear alongside built-in tools in the sidebar chat

This allows extending OxideSens with domain-specific capabilities — database queries, cloud infrastructure management, documentation search, and more.

RAG Knowledge Base

Import your own documents into scoped knowledge collections:

How It Works

1. Import: Add Markdown (.md) or plain text (.txt) files
2. Chunking: Markdown-aware chunking preserves heading hierarchy — sections stay semantically coherent
3. Indexing: Dual index — BM25 keyword index + vector cosine similarity
4. Search: Hybrid retrieval fuses both scores via Reciprocal Rank Fusion (RRF)

Collection Scoping

• Global collections — available across all connections
• Per-connection collections — scoped to a specific SSH connection for project-specific documentation

CJK Support

The tokenizer uses bigram segmentation for Chinese, Japanese, and Korean content, ensuring accurate search results without requiring a full NLP pipeline.

Use Cases

• Import project documentation for context-aware code assistance
• Add runbooks for instant operational guidance
• Store internal API references for command generation

AI Agent Mode

OxideSens includes an Agent Mode that enables multi-round autonomous task execution — the AI plans a sequence of tool calls, executes them, observes results, and iterates until the goal is reached.

How It Works

1. You give a high-level goal: “set up a Python virtual environment and install the requirements”
2. The agent decomposes it into steps, choosing tools at each step
3. After each tool call, the agent observes the result before deciding the next action
4. The loop continues until the goal is complete or a blocking decision requires your input

Tool Approval Modes

To maintain control over autonomous execution, OxideSens provides three approval stances:

Mode	Behavior
Supervised	Confirm every tool call before execution
Balanced	Auto-approve low-risk tools (read-only); confirm writes and destructive actions
Autonomous	Auto-approve all tools; agent runs until completion

Default is Balanced. Switch via the toolbar in the AI sidebar.

Per-Tool Auto-Approval

For fine-grained control, you can whitelist specific tools for automatic approval while leaving others requiring confirmation. In Settings → AI → Tool Approvals, each tool can be individually toggled.

Tool Call Logging

Every tool call the agent makes is logged in the sidebar conversation with:

• Tool name and input parameters
• Result or error (collapsed by default, expandable)
• Elapsed time

This gives full auditability of what the agent did during a session.

Tip

Use Supervised mode when exploring a new task. Once you trust the agent’s behavior for a specific workflow, switch to Autonomous to let it run unattended.

Privacy & Security

• All API keys stored in OS keychain (macOS Keychain, Windows Credential Manager, Linux Secret Service)
• On macOS, key reads gated behind Touch ID via LAContext — cached after first auth per session; no entitlements or code-signing required
• Terminal buffer data is sent to the AI provider only when you explicitly request it (click the Context button and send a message)
• No telemetry or data collection — OxideTerm never phones home
• Local models via Ollama keep everything on your machine — zero network traffic for AI queries
• Streaming SSE for real-time responses — tokens appear as they’re generated