tai

tai is Terminal AI for people who want their terminal to keep behaving like a terminal. It turns natural-language requests into concise shell command proposals, classifies risk, asks before sensitive actions, streams command output, and reports exit status.

It does not expose hidden reasoning or chain-of-thought. Provider outputs are parsed as command proposals, and any model-generated thinking blocks are stripped before user-facing display.

Install

bun add -g @hasna/tai

From a source checkout:

bun install
bun run build

Surfaces

• npm package: @hasna/tai
• CLI binary: tai
• MCP binary: tai-mcp
• SDK export: createTai, classifyCommand, redactSensitiveText, parseCommandProposal, planAgenticCommands

CLI

tai propose "show the largest files in this repo"
tai plan "inspect the repo status, run the relevant checks, and show me the safe next commands"
tai classify "rm -rf dist"
tai run "ls -la" --yes

tai run streams stdout/stderr directly and exits with the child command status. It preserves the current working directory and environment by default, and forwards common process signals to the child.

Provider Routing

Defaults are local-first and configurable:

1. OpenAI-compatible local endpoint, default http://localhost:11434/v1 with TAI_LOCAL_MODEL.
2. Groq fallback when GROQ_API_KEY and TAI_GROQ_MODEL are set.
3. Cerebras fallback when CEREBRAS_API_KEY and TAI_CEREBRAS_MODEL are set.

Model availability changes over time. Verified on 2026-06-26:

• Ollama documents OpenAI-compatible endpoints at https://docs.ollama.com/api/openai-compatibility.
• Groq publishes the current hosted model catalog at https://console.groq.com/docs/models.
• Cerebras publishes the current hosted model catalog at https://inference-docs.cerebras.ai/models/overview.
• Candidate local/open model examples worth evaluating, if available in your runtime, include OpenThinker/OpenThoughts agent models, North Mini Code GGUF, and Devstral Small. Do not assume these IDs are installed locally; set TAI_LOCAL_MODEL.

Example:

export TAI_LOCAL_BASE_URL=http://localhost:11434/v1
export TAI_LOCAL_MODEL=devstral
export TAI_GROQ_MODEL=openai/gpt-oss-120b
export TAI_CEREBRAS_MODEL=qwen-3-coder-480b

Agentic Planning

tai plan uses AI SDK Core for native tool calling and multi-step loops. Verified against the current AI SDK docs on 2026-06-26:

• generateText supports tool calling for agentic automation.
• tool defines executable tools with typed input schemas.
• stopWhen: stepCountIs(n) lets the model call tools over multiple steps.
• @ai-sdk/openai-compatible, @ai-sdk/groq, and @ai-sdk/cerebras provide the local/open-provider routing surface.

The planner gives the model constrained tools to inspect safe terminal context, classify candidate commands, run strictly read-only inspection commands with timeout and redaction, and submit a final multi-command plan. The model never gets a write-capable shell tool. Every final command is reclassified locally before it is shown to the user or returned through MCP.

The read-only inspection tool does not invoke a shell. It uses a small execFile allowlist with a scrubbed environment, so shell substitution and interpreter side effects are refused before execution.

Safety Policy

• Allow: clearly read-only commands such as ls, cat, git status, pwd, find, and simple inspections.
• Confirm: writes, network calls, long-running processes, package managers, process control, and system changes.
• Block unless explicitly overridden: destructive filesystem actions, credential exfiltration, privilege escalation, deploy/publish operations, and force pushes.

See docs/architecture.md for the PTY/shell boundary design.

MCP

Run the stdio MCP server:

tai-mcp

Initial tools:

• tai.propose_command
• tai.plan_commands
• tai.classify_command
• tai.redact

Development

bun install
bun run build
bun run typecheck
bun test

GitHub Status

Canonical public repository: https://github.com/hasna/apps. @hasna/tai is published to npm; releases are cut by maintainers.