Claude Code Integration

Launch Claude Code with Treasure AI as the backend.

Treasure AI customers can use Claude Code with their existing AI Suite credits — no separate Claude subscription required.

Features

• Use Your Own TD Subscription - Run Claude Code using AI Suite credits
• Full Claude Code Features - Extended thinking, plan mode, subagents, and Opus model support
• TD Documentation Search - AI-powered search across Treasure AI documentation
• TD Skills Integration - Enhanced with td-skills for writing SQL queries, workflows, and building user segments on CDP

Architecture

When you run tdx claude:

1. Starts a local pass-through proxy to Treasure AI
2. Launches Claude Code with td-skills and td-docs MCP pre-configured
3. The proxy forwards Claude Code API requests (/v1/messages) to Treasure AI, which provides Anthropic-compatible API access to Claude models

Client identification

TD-proxied LLM requests use X-TD-Client: tdx/claude. TD API requests made by child tdx commands inherit the same identity through TDX_CLIENT. The User-Agent retains tdx/<version> and adds client=tdx/claude in its comment; X-TD-Client-Version remains the tdx version.

With --no-proxy, only child TD CLI requests receive this identity; external Claude providers are left unchanged. Parent startup requests and OAuth token refresh requests are outside this attribution. Explicit user settings can override the inherited environment.

Usage

tdx claude [options] [-- claude-options]

Options

Option	Description
-c, --continue	Forwarded to Claude Code to continue the most recent conversation
-r, --resume [session]	Forwarded to Claude Code to resume a conversation or open its selector
--no-proxy	Use your own Claude Code account (Anthropic/AWS Bedrock/Vertex AI, etc.)
--disable-1h-cache	Skip the one-hour cache default (Claude models only)
--refresh	Force a full td-docs MCP and td-skills refresh before launching
--subagent-model <m>	Model for Claude Code sub-agents: inherit (default), auto, or a model

Claude Code options are forwarded without reordering. To resume a known session ID, pass it directly after --resume:

tdx claude --model="efficient[1m]" --resume <session-id>

Setup

tdx claude sets up the td-docs MCP server and td-skills plugins for you, and skips any step that's already done — so the first launch on a machine does the real work and later launches start immediately. Skills refresh whenever tdx is upgraded, and at least every two weeks.

--refresh forces an update, which you shouldn't normally need.

Prompt caching

In proxy mode, Claude models use a one-hour prompt cache for the main conversation by default. This reduces rewrites after idle gaps, but writes cost 2× base input instead of 1.25× for five minutes.

To skip this default:

tdx claude --disable-1h-cache

The flag leaves prompt caching and explicit TTL settings intact. --no-proxy, sub-agent caching, and non-Claude providers' cache policies are unchanged.

Sub-agent model

Claude Code often delegates work to sub-agents, for example the built-in Explore agent that searches your codebase, or custom agents defined in .claude/agents/. Each sub-agent can run on a different model from your main conversation, and by default Claude Code picks those models on its own, so a session you started on a cheaper model can still spend Opus-tier AI Suite credits in the background.

What tdx claude does by default

• You picked a model with --model (for example tdx claude --model sonnet): every sub-agent runs on that same model. Your whole session, including Explore and Plan, stays on the model you chose.
• You didn't pick a model: nothing changes. Claude Code runs the main conversation on its default model and routes sub-agents the way it always has.

You don't need to do anything to get this behavior.

Changing it

Use --subagent-model when you want something else:

# Main conversation on the default model, sub-agents on a cheaper one
tdx claude --subagent-model efficient

# Main conversation on sonnet, sub-agents pinned to a different model
tdx claude --model sonnet --subagent-model haiku

# Let Claude Code choose sub-agent models on its own (no tdx involvement)
tdx claude --subagent-model auto

--subagent-model	Sub-agents run on
inherit (default)	The model you chose with --model. If you chose none, the default leaves Claude Code's routing as is, while an explicit --subagent-model inherit makes every sub-agent follow the main conversation's model.
auto	Whatever Claude Code decides: each agent's own model setting, and its built-in defaults for Explore.
a model name	That model, for every sub-agent. Any name --model accepts works: haiku, efficient, efficient[1m], claude-sonnet-5.

Details

• tdx claude implements this through Claude Code's CLAUDE_CODE_SUBAGENT_MODEL and CLAUDE_CODE_SUBAGENT_MODEL_FORCE environment variables. If you already export either of them in your shell and don't pass --subagent-model, tdx leaves your values alone.
• ANTHROPIC_MODEL in your environment counts as a chosen model, the same as --model.
• With --no-proxy, tdx only sets these when you pass --subagent-model explicitly.
• Applying the model to the built-in Explore and Plan agents needs Claude Code 2.1.257 or later. Older versions still honor it for custom agents.

Examples

# Launch Claude Code with Treasure AI
tdx claude

# Continue most recent conversation
tdx claude -c

# Run sub-agents on a cheaper model
tdx claude --subagent-model efficient

# Resume a conversation (shows selector)
tdx claude -r

# Start with an initial prompt
tdx claude -- "Explain this code"

# Use your own Claude Code account with TD skills
tdx claude --no-proxy

# Force a td-docs MCP and td-skills refresh
tdx claude --refresh

Example Conversations

Once inside Claude Code, you can ask questions like:

> How to set up data ingestion to TD from Snowflake?

> How to add web access log collector in my web page?

> Analyze mydb.user_events data

> Write a workflow to periodically import data from S3

> Create a SQL query to find top customers by revenue

> Build a user segment for customers who purchased in the last 30 days

Requirements

Install Claude Code

See the Claude Code Quick Start for installation instructions.

curl -fsSL https://claude.ai/install.sh | bash

Minimum Version

Claude Code 2.0.20 or later is required for td-skills support. Update with:

claude update

Other Requirements

• Valid TD API key with LLM access

For usage across Claude, Codex, and OpenCode, run tdx usage.

Discover models

Use tdx ai models to discover model IDs and aliases for --model or --subagent-model. Filter by v1_messages for Claude Code-compatible models:

tdx ai models --feature v1_messages
tdx ai models --feature v1_messages --provider anthropic

Subcommands

statusline

Output credit usage for Claude Code's status bar. This is automatically configured when running tdx claude. See Claude Code Statusline Documentation for more details.

tdx claude statusline

Reads session info from stdin (JSON with session_id and model) and outputs a formatted statusline.

When using --profile, the statusline shows profile-specific today/month credits:

tdx --profile myprofile claude

Output:

🤖 tdx - Sonnet 4.5 | 🪙 credit: 0.12 session / 1.23 today / 10.45 month | 🧠 22,396 (11%)

The statusline shows:

• Current model name
• Session / Today / Month credits (estimates, filtered by profile when specified)
• Context memory usage

For users with a treasure.ai or treasure-data.com email, USD cost replaces credits in the same compact display:

💎 tdx - Sonnet 4.5 | 🪙 $12.28 session / $125.85 today / $1069.24 month | 🧠 22,396 (11%)

Costs use tracked proxy token usage and model pricing, following Treasure Work. The email is cached when starting tdx claude; restart it to refresh an existing account cache. Daily and monthly totals remain filtered by profile when specified.

Learn More

For generic Claude Code usage, refer to the official documentation:

• Claude Code Documentation
• Claude Code in Action (Tutorial)