"Relay agents" usually means software that forwards work between AI agents, or between an agent and an API the agent must call. The same word covers four different shapes: message relays, task routers, HTTP gateways, and identity brokers.
On 2026-10-05 search results mix those shapes. Agent Relay (agentrelay.com) sells channels and durable delivery between coding agents. Relaycast (github.com/agentworkforce/relaycast) calls itself headless Slack for agents. Other repos use "agent relay" for a local control plane or an open agent registry.
Each relay kind fixes a different failure. Match the product to the broken step, not the search keyword.
Empyre Relay is OAuth and identity for agents, not an agent-to-agent chat bus.
Four relay meanings in agent stacks
| Kind | What it relays | Without it |
|---|---|---|
| Message / coordination relay | Events, channel posts, DMs, and typed actions between agents | Each agent runtime stays isolated; handoffs live in ad hoc logs or copy-paste |
| Task routing relay | A human task into runs on one or more CLI agents, with streamed logs | You SSH to each machine to launch Codex, Claude, or similar tools separately |
| API / LLM gateway relay | HTTP calls to models or vendor APIs under policy, keys, and quotas | Agents hold long-lived API keys in env vars the model can read |
| Identity / OAuth relay | User consent, scoped tokens, refresh, revoke, and introspection for agents | You paste root credentials into prompts, or build OAuth yourself in every app |
Message relays between agents
Products in this bucket treat agents like teammates that need shared context. They add channels, direct messages, durable delivery, and listeners so one runtime can react when another posts.
Agent Relay at agentrelay.com was read 2026-10-05. The page title is "Turn coding agents into a working team." Its JSON-LD describes channels, DMs, durable delivery, event listeners, and typed actions for any agent runtime. The homepage lists Y Combinator among investors on that date.
Relaycast's README at github.com/agentworkforce/relaycast, read 2026-10-05, opens with "Headless Slack for agents" and documents shared channels, threads, DMs, reactions, files, search, and realtime events via the `@relaycast/sdk` package.
These tools do not replace login for your end users. They coordinate builders. If your failure mode is "Codex and Claude never see each other's status," a message relay fits. If the failure is "this agent cannot call Gmail with the owner's consent," you need identity relay instead.
Task routing relays for coding agents
A second "agent relay" pattern is a control plane that accepts a task, launches configured CLI agents, streams logs, and stores structured results. The relay does not review code itself; it orchestrates runs you already trust on specific hosts.
The README at github.com/skyzer/agent-relay, read 2026-10-05, describes Agent Relay as a local-first control plane for your own agent fleet. Its documented flow pastes a GitHub PR link, runs enabled agents on a relay machine and optional remote hosts, and compares outputs in one UI.
Pick this when your agents are already installed on machines you control and you want one place to start and observe them. It does not mint OAuth tokens for third-party SaaS, and it is not a federated directory for arbitrary agents on the public internet.
Open registry and A2A-oriented relays
Another repository named Agent Relay lives at github.com/agent-relay/agent-relay. Read 2026-10-05, its README frames open infrastructure for an agent economy.
That includes a federated registry, agent cards, and discovery so one agent can request skills from another.
That vision overlaps with protocols such as A2A and MCP in prose, but the code path is different from a hosted Slack-like bus or from OAuth. It targets discoverability and paid skills on open rails, not hosted consent for a single app's users.
If you need a searchable network of agents, read that repository's docs. If you need your product's customers to sign in once and let an agent act with scoped access, registry work will not solve the token problem.
API and LLM gateway relays
Searchers sometimes say "API relay" when they mean any proxy between an agent and a vendor HTTP API. That layer can enforce rate limits, redact secrets, or centralize billing for model calls.
Gateways solve credential sprawl and outbound policy. They do not, by themselves, prove which human approved access to a third-party account. When the risk is key leakage into prompts, start with architecture that never loads root keys on the agent host.
For OAuth brokers versus HTTP gateways, read API relay for AI agents on this site. That page is about pattern choice, not a product catalog.
Identity and OAuth relays for agents
Identity relay means a hosted authorization server built for unattended software. A human approves once at consent. The agent receives short-lived, scoped tokens, refresh rotation, and endpoints that fail closed on revoke and introspect.
Without that layer, teams paste API keys into agent context, store refresh tokens in plain config, or reimplement PKCE and rotation in every service. Any of those paths breaks the moment the model logs a secret or a token outlives the owner's intent.
Empyre Relay at relay.empyre.dev is in this bucket. It is OAuth for AI agents: hosted consent, mandatory S256 PKCE, one-time authorization codes, refresh-token rotation with family revocation, and fail-closed revoke and introspect. The platform API serves /relay/oauth/* at api.empyre.dev (authorize, token, refresh, introspect, revoke, and scoped GET /relay/oauth/userinfo added 2026-08-06).
Relay has been feature-frozen since 2026-07-10 for everything except bug and security fixes. The published npm contract is `@empyre/relay-sdk` version **1.0.0** on registry.npmjs.org, read 2026-10-05. DeelFlow (deelflow.dev) is documented as an external integration example in Empyre's Relay materials; verify its live behavior on that site before you copy patterns.
Install paths, MCP, and "plugin" search intent are covered in Relay plugin for AI agents. Endpoint and SDK mapping live in Relay API. Token grants and error cases are in Agent OAuth token exchange and OAuth for AI agents.
How to choose a relay kind
| Your problem | Start with |
|---|---|
| Coding agents ignore each other's progress | Message relay (Agent Relay, Relaycast, or self-hosted equivalent) |
| You run CLI agents on several machines and want one launch UI | Task routing relay (for example skyzer/agent-relay) |
| You want a public directory of agent skills and A2A-style discovery | Registry-oriented relay (read agent-relay/agent-relay docs) |
| You must call third-party APIs as the end user, without root keys on the agent | Identity / OAuth relay (Empyre Relay or another OAuth broker) |
| You must hide model API keys and enforce outbound HTTP policy only | API or LLM gateway relay (see api-relay article for pattern) |
What Empyre Relay is not
Empyre Relay does not host channels where your Codex and Claude instances chat. It does not replace Relaycast or Agent Relay for multi-agent coordination inside a build farm.
Empyre Relay also does not operate your company after launch. That work lives on empyre.dev: eight executives on schedules, deploy, marketing, support, and budget gates. Relay is a separate product on the same account, like Vault for signing without key export.
Confusing identity relay with a message bus produces integrations that authorize nobody while still missing team coordination, or the opposite. Name the failure you are fixing before you pick a vendor keyword.
Where Empyre the company builder fits
Empyre on empyre.dev turns a plain-language brief into a live software company with eight AI executives that keep operating it. That is not a coding-agent seat and not an agent chat relay.
Generated products ship working sign-up and sign-in on the first deploy with no auth provider to register for, per Empyre's product file.
The company scaffold in `backend/tools/scaffold.py` seeds a sign-in form with "Continue with Relay" at `/api/auth/relay` and adds `src/app/api/auth/relay/route.ts` when the repo is created.
Third-party apps register OAuth clients at relay.empyre.dev and use `@empyre/relay-sdk` against api.empyre.dev.
If your search was only "relay agents" meaning team coordination, stay with the message-relay products above. If you needed delegated API access for agents, use the Relay articles linked in the identity section. If you needed a company operated after deploy, start with AI website builders vs AI business operators instead.
Common questions
Are relay agents the same as agent-to-agent (A2A) messaging?
Often, but not always. Message relays implement coordination between agents you already run. A2A also names open protocols and registries for discovery. Identity relay is a different layer: it answers who may act and with which scopes, not which channel gets the next post.
Is Empyre Relay an agent chat bus?
No. Empyre Relay is OAuth and identity for AI agents at relay.empyre.dev. Agent Relay at agentrelay.com and Relaycast are examples of message-style relays, verified read 2026-10-05 from their live site and README respectively.
What is the difference between Relay API and this page?
"Relay agents" overloads four relay kinds in the market. The Relay API article maps Empyre OAuth endpoints and SDK calls only.
Does Empyre Relay replace API keys for OpenAI or other model vendors?
Not by itself. Relay issues OAuth tokens for identities and delegated third-party access you wire to consent. Model gateways are a separate relay kind. A stack can use both: identity relay for user-owned APIs, and a gateway for model keys.
What was verified on 2026-10-05?
agentrelay.com homepage title and JSON-LD; investor strip listing Y Combinator; Relaycast README tagline; skyzer/agent-relay and agent-relay/agent-relay README positioning; @empyre/relay-sdk 1.0.0 on registry.npmjs.org. No search volume, ranking, or user-count figures appear here.
Try Empyre free for 3 days
Describe a business in plain words and watch eight AI agents build and deploy it. Starter is free for the first 3 days.
Related
Last updated 2026-10-05. Competitor descriptions reflect each product's publicly documented capabilities at that date; they change often, so check the source before relying on a detail.