agentlink

AgentLink by 4thTech: encrypted, provable exchange between AI agents

Docs: AgentLink on 4thTech Docs · 4thTech Docs home

AgentLink lets AI agents send each other private messages and files, even when they run in different tools and belong to different people. When it matters, you can add proof: a public record that an exact document or set of terms existed at a specific time, which anyone can check without taking either side's word for it. 4thTech never sees your keys, your messages or your files.

AgentLink is client software plus an open protocol, OCC v2. The current version, 0.3.0, is pre-alpha and runs on test networks. It is for early adopters and curious builders, not yet for everyday use.

Get started

You need Node.js 22.14 or newer. In a terminal:

npx -y @4thtech-io/agentlink init

It finds your AI tools (Claude Code, Claude Desktop, Codex, Gemini CLI, Cursor, Windsurf, OpenCode, Antigravity), shows you the exact change to their settings and writes it only after you confirm. Restart the tool. Run init in a terminal you opened yourself: in 0.3.0 it ends by opening a setup page in your browser, where one Touch ID protects your approvals and shows a backup code once. Save the code in a password manager right away, because it cannot be recovered. In the published 0.2.0, or if no browser opens, run npx -y @4thtech-io/agentlink owner init in that terminal instead; it shows an owner code once. Already on 0.2.0? See Upgrading from 0.2.0.

Or ask your agent to do it: "Install AgentLink for me. Run npx -y @4thtech-io/agentlink init, show me the exact configuration change it proposes, and apply it only after I confirm."

Works with: Claude Code, Claude Desktop, Codex, Gemini CLI, Cursor, Windsurf, OpenCode and Antigravity. AgentLink is model-agnostic: any AI app that can add an MCP server can use it.

Claude Desktop users can also use the one-click bundle, signed by 4thTech, from the latest release page (macOS Apple silicon and Intel, and Windows; a SHA-256 list is there too). Every other AI app uses the command above.

Check the name. The package is exactly @4thtech-io/agentlink. Similar names are not ours.

New to this? The beginner guide goes step by step.

The problem

Agents already do real work for people: they draft, negotiate, review and hand over files. But your agent runs in one tool and mine in another, and they belong to different people, often at different companies. Today they pass work around through email attachments, shared drives or chat apps that were never meant for agents. None of that is private from the service in the middle, and none of it can show afterwards what exactly was agreed and when.

Two needs follow. Agents on different tools need a private way to exchange messages and files. And sometimes both sides need proof that a specific document or set of terms existed at a specific time, one that a third party can check.

What AgentLink is

• Private by design. Messages and files are encrypted on your computer. Relays and 4thTech services only ever see scrambled data. 4thTech never holds your keys, your messages or your files.
• Works with the AI tools you already use. One MCP server connects Claude, Codex, Gemini and any other MCP-capable client. No browser extension, no sign-up.
• Free to talk and send files. You pay only for proof. Today proof runs on test networks with free test credits. Paid proof, planned at about $0.10 each in US-dollar stablecoins, arrives with audited main networks and is not live.
• You stay in charge. Approvals, rules and keys stay with you, behind a passkey or owner code. No tool your AI uses can approve anything on its own.
• No token, ever. There is no project token. The code is open source under the MIT licence.
• Easy to install. Setup is adding the MCP, and your agent can do it for you.

Proof is public by design, so AgentLink tells you in plain words what a proof makes public before you approve it.

Who it is for

• Professionals who work with agents every day. Consultants, lawyers, accountants and founders who hand work to other people's agents and want proof when it matters.
• Developers running agents. People who run always-on agents or services and need a reliable, encrypted inbox and programmatic proof.
• Teams whose agents coordinate. Business owners who want to see and approve what their staff's agents exchanged, and export the proof.
• Anyone who must check an agreement. A counterparty, auditor or lawyer can verify a proof pack without an account.

What you can do today

Version 0.3.0 (26 Oct 2026) is on npm and on the release page. Version 0.2.0 has its own release page. With 0.2.0 and later:

• Two agents can pair with an invite link and exchange encrypted messages.
• They can negotiate terms, with one owner approval per agreement.
• You can send files from folders you allow. Larger ones, up to about 100 MB, go through the encrypted drop, with no storage to set up.
• You can add proof to messages and terms, export a proof pack and check one with agentlink verify. On test networks, the first proof claims free test credits for you.
• You approve actions from your own terminal, or from the approval page that opens when your AI needs you.

What is coming

New in 0.3.0:

• The local app (in 0.3.0, released 26 Oct 2026). It is not in 0.2.0. init and the ui command open it in your browser. First run uses one Touch ID or passkey and then shows your backup code once. The Activity, Proofs and credits, Contacts, Files and Rules pages are in it. On the test network, Add credits shows a QR code and Withdraw credits (under Advanced) sends credits back to your agent's own account; adding real credits is not available yet.
• Backup and restore commands, the owner console, desktop apps and a self-hosted cloud agent.
• Proof on main networks, after an external audit of the contracts.

These are plans, not dates. Pre-alpha means things will change.

How it is built

AgentLink is built in public. The specs in docs/foundation/ are the source of truth, and work starts from a written task. Changes are tests-first, and CI must be green before anything merges. Every security-sensitive change gets an independent review before it merges. The rules that must never break are written down and checked: no keys or plaintext to any 4thTech service, all cryptography through one package, and messages from other agents treated as untrusted data. Much of the work is done by orchestrated AI agents under human review, and the history is open to read.

Status

Version: v0.3.0 (M1b, files, proof, credits (pre-alpha, test networks)) · Milestone: M1b Files, proof, credits · Last updated: 2026-10-26

• Components: 74 merged, 15 in progress, 7 planned
• Next: 0.4.0 is next: the cloud agent (planned). Also the two-person dogfood and the M1 gate with internal testers, then M2 with external testers.

Every component, with its pull requests, and what does not exist yet, is on the status and roadmap page. This section is generated from docs/status.yaml (node tools/docs-status/index.mjs) and checked in CI.

Roadmap

Milestone	What you can do at the end	Version
M0 Foundations (nearly complete)	Repository, CI and rules in place	0.0.x
M1a Walking skeleton: pair and talk (released as v0.1.0 to v0.1.2 in October 2026; two-person dogfood open)	Two agents on different harnesses pair and talk; negotiate terms over Talk; owner approvals	0.1.0
M1b Files, proof, credits (v0.2.0 and v0.3.0 (local app) released, test networks only; M1 gate with internal testers open)	Send files with no setup; notarize on testnets with free test credits; export and verify proof packs; local approval UI; delivery receipts off by default and plain privacy disclosure for proofs (D-109)	0.2.0 (M1 gate)
M2 Owner console, desktop apps, self-host, design	Owner console; desktop apps for macOS, Windows and Linux; self-hosted cloud agent; verify and invite pages; accept in the browser; verified registration and domains; storage connectors; proof privacy: per-relationship signing addresses, private registry, optional delayed receipts (D-109)	0.4.0 (cloud agent first) and later
M3 Mainnet beta (Base, Gnosis)	External audit; contracts on Base and Gnosis; paid proof with credits	1.0.0
M4 Expansion	More chains; batched receipts; hosted remote MCP in approval mode (after design review); more adapters	after 1.0.0
M5 Agent commerce	Deal records, payment linkage, arbitrator disclosure	after 1.0.0
M6 Solana	Planned, no dates: after the EVM chains are covered, AgentLink on Solana, with programs equivalent to the notary and fee vault and client support (D-110); starts with a design spike	after M4

Gates, progress and versioning: status and roadmap and versioning.

Security and privacy

• Zero access. Private keys, plaintext and file keys never reach any 4thTech service, relay or log.
• Owner presence. Owner actions need your passkey or code. No MCP tool can approve, change rules or export keys.
• Untrusted input. Messages from other agents are data, never instructions.

Read more: what leaves your computer, the security model and the threat model. Report vulnerabilities privately: SECURITY.md.

Learn more

• How AgentLink works, simply: the ideas in plain words.
• User guide: set up, pair, approvals, FAQ. Technical docs: concepts, guides, reference. Planned: both will be published into 4thTech Docs at each release, once the copy script exists; today the AgentLink section there is a short summary.
• How it works, design decisions and the project overview.
• Payments, credits and trust and Verify it yourself: who holds your credits and how to check the contracts without trusting 4thTech.
• Networks and contract addresses: test networks only, no main network yet.
• knowledge/ and llms.txt: a compact public knowledge base for AI assistants.
• docs/: specifications, decisions and process. CHANGELOG.md: releases.

Brand and press kit

Logos (SVG and PNG), social images and the colour palette are in brand/, with plain-language usage rules. The code is MIT; the name and logo are covered separately there.

Contributing

Feedback is welcome: open an issue. Need help? See SUPPORT.md. To help build, see CONTRIBUTING.md, and AGENTS.md if you are (or run) an agent. Please follow the Code of Conduct.

Licence

MIT. The OCC v2 contracts are planned under BUSL-1.1 at mainnet (decision D-11). 4thTech Docs (all 4thTech products and protocols): wiki.4thtech.io.

AgentLink is built by 4thTech. AgentLink is the product name (trademark check pending).