Agent Relay: Open-Source Messaging, Shared Sessions and Flows So Claude Code and Codex Can Hand Off Work

Agent Workforce Incorporated released Agent Relay, an Apache-2.0 open-source toolkit for coordinating coding agents such as Claude and Codex across tools, people and machines, announced by Will Washburn on October 23, 2026, with the latest release v14.0.0 published October 13, 2026. The repository, created December 2025 and at about 900 GitHub stars, ships four components — Messaging, Integrations, Shared Sessions and Flows — and installs via `npm install -g agent-relay` or a curl install script, with a no-credit-card Agent Relay Cloud option and a self-hosting guide. Agent Relay targets coordination failures rather than model failures, exposing GitHub, Linear, Notion and Slack as a virtual filesystem instead of separate MCP servers and defining multi-step Flows in TypeScript with deterministic checks and human gates.

Anyone who runs more than one coding agent knows the copy-paste routine: one agent finishes a task in a terminal, you paste the summary into chat so a teammate can paste it into their agent, and half the context is lost on the way. Agent Relay , an open-source project from Agent Workforce, aims to remove that step. Will Washburn announced it on October 23, 2026 we could not open the original post ; the README says it lets teams run coding agents together, with searchable shared sessions, and offers a cloud version that needs no credit card. This post goes through what the AgentWorkforce/relay repository https://github.com/AgentWorkforce/relay?ref=explainx says it does, how to install it, where it sits among other agent coordination approaches, and what to check before putting it in a team workflow. We did not run it ourselves, so everything below comes from the README, release page and public posts. TL;DR: Agent Relay in one table | Question | Answer | |---|---| | What is it? | Open-source toolkit for running coding agents together across tools, people and machines | | Who makes it? | Agent Workforce Incorporated; announced by Will Washburn | | License | Apache-2.0 | | Latest release | v14.0.0, published October 13, 2026 | | Repo activity | About 900 GitHub stars, created December 2025, pushed to on October 23, 2026 | | Four pieces | Messaging, Integrations, Shared Sessions, Flows | | Install | curl -fsSL https://raw.githubusercontent.com/AgentWorkforce/relay/main/install.sh \| bash or npm install -g agent-relay | | Cloud | Agent Relay Cloud, no credit card; self-hosting guide exists | | Named agents | Claude, Codex README | The problem it targets The README opens with a list of complaints that will sound familiar: copy-pasting from Claude Code into Slack so a coworker can paste it into their agent; putting an important rule in a skill or AGENTS.md and watching the agent ignore it; two people's agents building the same thing; wondering whether a coworker's agent made a change on purpose; and typing "babysit this PR until CI passes" for the fifth time that week. None of those are model problems. They are coordination problems between tools, which is why Agent Relay positions itself as infrastructure, not another agent. It "gives engineering teams infrastructure for running coding agents together without replacing the agents and tools they already use." The four pieces Messaging Agents such as Claude and Codex "can talk directly through shared channels, threads, DMs, files, search, and real-time events," and can run on different machines while sharing a workspace. Think Slack-style primitives for agents rather than humans. Integrations as a virtual filesystem GitHub, Linear, Notion, Slack and others are "exposed as a virtual filesystem," so agents use ls , cat , grep and ordinary file writes to work with them. This is a deliberate alternative to wiring each tool through its own MCP server. For background on that protocol, see our MCP guide https://explainx.ai/blog/what-is-mcp-model-context-protocol-guide . The filesystem idea lives in a separate repository, relayfile , which the README links. Shared Sessions This is the feature behind the "read each other's sessions" claim. Agent Relay captures coding agent sessions so teammates and their agents can search previous work, decisions and context. Agents query team history via a Shared Sessions plugin, or you can connect its MCP server with agent-relay mcp --sessions-only after a Cloud login. The README points to two companion projects for how sessions relayhistory and decisions trajectories are captured. Flows "Turn instructions you hope an agent follows into workflows you can enforce." Flows are multi-step workflows defined in TypeScript with deterministic checks, required steps and human gates. The idea is to put rules that matter in code instead of relying on a prompt to be remembered. This sits next to what we describe in loop engineering for coding agents https://explainx.ai/blog/loop-engineering-coding-agents-claude-code-guide-2026 , where verification loops, not longer prompts, make agents reliable. How to try it 1. Cloud route. Go to the Agent Relay Cloud flows page linked from the README, and read the official documentation https://agentrelay.com/docs/introduction?ref=explainx first. It says no credit card is required. 2. CLI route. Install with the curl script or npm install -g agent-relay . The latest release lists standalone binaries, so Node.js is not required for the script install. 3. SDK. npm install @agent-relay/sdk for programmatic use. 4. Self-hosting. Each primitive has self-hosting options; start with the self-hosting guide in the repo's docs folder. The team offers help at the contact address on the README. A cautious first test: run two agents, say Claude Code and Codex, in a throwaway repository. Let one finish a small task, then ask the other to read its session and continue. If the handoff works, only then connect a real repo. What a team workflow could look like Based on the four pieces in the README, here is a plausible setup for a small team. It is our illustration of how the parts combine, not a documented recipe. 1. A developer asks Claude Code in the terminal to investigate a failing test. The session is captured as shared history. 2. A teammate working in the Codex Mac app asks their agent, "what did the investigation find?" The agent searches team sessions through the plugin and reads the earlier reasoning instead of starting from scratch. 3. A flow defines the release path: run tests, require a review step, then a human gate before anything is merged. Because it is TypeScript with deterministic checks, a skipped step fails the flow rather than depending on the agent to remember. 4. Messaging channels let a third agent that watches CI post updates, replacing the manual "babysit this PR until CI passes" instruction. Each step reduces a manual handoff. Each also adds a dependency, so the more your process leans on shared sessions and flows, the more the reliability and security of Relay itself matters. What people will ask next Does it replace Claude Code or Codex? No. The README says it works without replacing the agents and tools you already use. Is this the same as agent-to-agent protocols? It overlaps with the idea behind them, but it is a concrete product with a message store, a plugin and a CLI rather than a protocol specification. Our coverage of the AWS agent registry https://explainx.ai/blog/aws-agent-registry-runtime-tool-discovery-mcp-a2a-agentcore-sample-guide-2026 explains how protocol-level discovery differs. Can I run it fully offline or on my own infrastructure? The README says self-hosting options exist for each primitive and points to a guide, and the team offers setup help. Whether every feature works without the cloud login is something to test, since the documented sessions flow mentions it. What does it cost? The cloud version needs no credit card to try, per the README. We found no published pricing for paid tiers. Who is behind it? The README says "Copyright 2026 Agent Workforce Incorporated." It names no individual founders, so see the website for team details. Where it fits among agent coordination options | Approach | What it solves | Limits | |---|---|---| | Skills and AGENTS.md | Teach an agent your conventions; see agent skills explained https://explainx.ai/blog/what-are-agent-skills-complete-guide | Advisory; an agent can ignore them | | Markdown wiki or shared memory | Manual cross-tool handoff; see our shared memory and handoff guide for Claude Code and Codex https://explainx.ai/blog/ai-memory-akitaonrails-shared-memory-claude-code-codex-handoff-markdown-wiki-guide-2026 | You maintain it | | MCP servers | Tool access for one agent | Not a team session history | | Agent Relay | Cross-agent messaging, searchable team sessions, enforced flows | Newer; cloud-first for some features | Relay does not replace these. The honest pitch is that it adds a team layer: who did what, in which session, with an enforceable process around it. What to check before relying on it 1. Where session data goes. Capturing sessions means storing prompts, code and possibly secrets. The documented sessions flow uses a Cloud login. Confirm retention, access controls and self-hosted alternatives before capturing anything sensitive. 2. Permissions. An agent that can write to Slack, GitHub or Linear through a virtual filesystem is an agent with real credentials. Scope tokens narrowly and review what each flow can write. A runtime guard such as AgentBeam https://agentbeam.com/?ref=explainx , the agent security platform from the explainx.ai team, can add a policy layer where agents take actions. 3. Prompt injection across agents. If agents read each other's sessions and messages, text from one agent or document can steer another. Treat inbound messages as untrusted input, especially from tickets and external repositories. 4. Maturity. The repository is under a year old, has about 900 stars, and ships major versions quickly v14.0.0 in October . Pin versions and read the changelog. 5. Claims on social media. Posts promoting it are endorsements, not tests. We found no independent benchmark or review yet. Why this matters As teams adopt several agents at once, the bottleneck shifts from writing code to coordinating who is doing what. Tools that give agents shared memory and enforceable processes are an early answer, and an Apache-2.0 option with self-hosting matters for teams that cannot send code to a vendor. Whether Agent Relay becomes the standard depends on how well it handles security and how many agent vendors expose hooks for it. For now it is a promising, young project worth a sandbox trial. Details are accurate as of October 23, 2026, based on the project's README, release page and public posts. Features and licensing can change.