Skip to main content
T3N ADK is Terminal 3’s SDK for building AI agents that act only on scoped, revocable permission a user actually granted — never a standing API key. See What is T3N ADK? and Why T3N ADK? for the full picture.
Just starting? Paste the prompt on Start with an AI Assistant into a new session: it runs the whole Get Started guide for you. The skill file below is for ongoing work.
If you’re building with an AI coding assistant, you don’t need to paste this whole site into its context window. Drop the skill file below into your project instead — it teaches the assistant to build and debug a T3N ADK integration in the right order, with a known-pitfalls table to check before guessing at a fix.

Install it

Save the file below as .claude/skills/t3n-adk-quickstart/SKILL.md in your project (or your global ~/.claude/skills/ directory), then invoke it with /t3n-adk-quickstart or just describe what you’re building — Claude Code will pick it up automatically based on the skill’s description.

Self-service agent credit funding

T3nClient’s createAgent method (via its initialCredits option) and its fundAgent() method (confirmed live on testnet, @terminal3/t3n-sdk 5.16.0+ — see the Changelog) let a developer fund an agent’s T3N credit balance without asking a cluster operator (Terminal 3’s own ops team). Both are called on the authenticated client (t3n.createAgent(...), t3n.fundAgent(...)) from Step 3 above, not imported as standalone functions. The authorization model behind them is easy to get wrong on a first read, and most of the confusion isn’t in the API shape — it’s in why a call gets refused. The two prompts below encode that model as explicit rules so your assistant can reason about a refusal correctly the first time, rather than guessing from the error text alone. Same install mechanism as the skill file above — save the Claude-flavored one as a Claude Code skill (or append it to the skill above), and the OpenAI-flavored one as a system prompt / project instructions file for an OpenAI-family assistant.

FAQ

The Markdown file in the accordion above — not documentation to read, but something you drop straight into your project so the assistant follows it automatically. See Install it for the exact file to save per tool.
Claude Code natively (as a Skill), Codex (saved as AGENTS.md), Cursor (saved as .cursorrules), and Windsurf or any other tool that accepts a plain system-context/rules file — see the tabs under Install it.
No. The skill file only teaches the assistant how to call the SDK correctly — it doesn’t change what the assistant can see at runtime. Real user data (PII, credentials) is substituted inside the TEE via placeholder markers (see ADK Tour), so it never reaches the model’s context window, a chat log, or your application logs. See Placeholders in outbound calls for the mechanism.
Yes, and that limit is also the containment: an agent only ever gets the specific functions and external hosts a user explicitly delegated to it, with no broader standing access to fall back on — so a compromised or misbehaving agent can still only reach what was delegated. The contract still runs, but any other call is simply refused (host/http.egress_denied), not silently allowed. See Member Delegation for how a user grants that scope, and Outbound HTTP calls are authorized by the user, not the contract for what the refusal looks like.
Yes — see This documentation is agent-readable too below. The skill file is the fast path for building; the raw Markdown pages and /llms.txt index are there for open-ended exploration or questions the skill file doesn’t cover.

This documentation is agent-readable too

Every page on this site is available as plain Markdown by appending .md to its URL, and a machine-readable index of every page lives at /llms.txt — point an agent there if you want it to explore the docs directly instead of relying on the skill file above.

SDK & API Reference

If you’re pasting context manually instead of using the skill file, this single page is the highest-value one to include — every confirmed method, host interface, and API tag in one place.