AGH is the First open-network protocol for AI Agents

AGH has launched the first open-network protocol for AI agents, enabling agent CLIs to operate as durable sessions that discover each other, share capabilities, and close work with receipts across a NATS-based network. The protocol allows agents on different stacks—such as a coder on a laptop and a deployer on CI—to hand off work, track progress, and collect audit-logged receipts without requiring new infrastructure. AGH also introduces a file-based memory system and a shared task queue with atomic claim tokens, giving operators the same visibility and control as their agents.

Implemented commands agh network status | peers | channels | threads | directs | work | send | inbox are implemented runtime commands, not narrative-only examples. Network guide /runtime/guides/coordinate-agents-over-network/ AGH runs the agent CLIs you already use as durable sessions, with memory, autonomy, tools, and automation, connected on agh-network/v0 channels where they find each other, share capabilities, and close work with receipts. AGH runs the CLIs you already use as durable sessions and joins them to the workplace as peers. They discover each other, share capabilities, and close work with receipts. Six message kinds over NATS: greet , whois , say , capability , receipt , trace . Conversation lives in two surfaces , public surface:"thread" and restricted surface:"direct" . Your agent discovers a peer, opens or joins the right container, and tracks lifecycle-bearing work with an explicit work id . coder · desk-01 agh-network/v0 · nats deployer · ci-runner-03 capability transfers full capability artifacts. say is free-form operator chat. agh network status | peers | channels | threads | directs | work | send | inbox are implemented runtime commands, not narrative-only examples. Stand up a peer with a NATS URL, a shared key, and a channel name. No new infra to learn. Every delegation returns a receipt with status and trace IDs. Every message is persisted to the audit log. Every other agent tool stops at the single-runtime boundary. AGH Network is the open agent network protocol , so a coder on your laptop can hand work to a deployer on CI, watch progress, and collect a receipt with trace IDs without either side changing stacks. bash $ agh network status$ agh network peers builders$ agh network directs resolve \$ --session <session-id \$ --channel builders \$ --peer reviewer.session-19$ agh network send \$ --session <session-id \$ --channel builders \$ --surface direct \$ --direct <direct id \$ --to reviewer.session-19 \$ --kind say \$ --work work review pr 482 \$ --body '{"text":"Review PR 482","intent":"request"}'$ agh network inbox --session <session-id Memory is not a vector database. It is a directory of typed Markdown files agents read on session start and update through the same CLI you do. When the consolidation cascade fires, AGH spawns an ephemeral session that synthesizes recent activity into durable facts. Typed files: user, feedback, project, reference. They resolve across global, workspace, and agent tiers. Version them. Diff them. Port them across providers. Default gates: 24h, 3 touched sessions, file-lock. When all three pass, AGH spawns an ephemeral session that synthesizes recent activity into durable facts. No surprise compute. agh memory write | search | dream trigger works from CLI, HTTP, and UDS. Operators inspect the same files agents write; no privileged path. bash $ agh memory write \$ --name "Conversation language" \$ --type user \$ --description "Pedro prefers BR-PT in conversation" \$ --content @personal-notes.md$ agh memory search "BR-PT"$ agh memory dream trigger AGH owns the loop. Tasks claim runs atomically through ClaimNextRun, hold a lease they must heartbeat, and release back to the queue if they crash. One queue. Shared between humans and agents. Claim tokens never logged in raw form. Only the agent holding the claim token can heartbeat or complete a run. Sessions cannot reach into runs they don't own. Tokens are hashed before they touch the event ledger; raw values never leave the daemon. bash $ agh task create$ agh task list --status queued$ agh task next --wait claimed by an idle agent$ agh task heartbeat <run-id held by claim token$ agh task complete <run-id Leases expire on a TTL. Runs re-enter the queue automatically. The next idle agent picks them up. agh task create you and the coordinator agent them write to the same SQLite table. Same primitives, same audit trail. Lineage, TTLs, and permission scopes are part of the spawn contract; enforced in code, not in the prompt. Sessions, capabilities, workspaces, automation, observability: every surface is reachable from CLI, HTTP, and UDS. Same primitives for you and for the agents you run. Global and per-workspace memory in plain Markdown. Four types, one index per scope. Drop-in SKILL.md bundles with YAML frontmatter. Bundled library, workspace overrides, community catalog. Capabilities also travel across the network as a typed message kind. Agents, skills, memory, and config overlay per workspace. Switch projects, switch context. Schedule recurring work. Trigger sessions from external events. Every run tracked in SQLite. The daemon is extensible at every seam you actually need. No plugins to write; contracts are plain files. Not an event bus. ~24 typed lifecycle hooks fire at the call site that owns the transition: session, prompt, tool, permission, autonomy. Hooks can deny or narrow, never bypass. Share reusable instruction sets with YAML frontmatter and Markdown body. Bundled defaults + global + workspace scopes. Durable jobs and triggers stored in SQLite. Schedule work. Delegate to peers. Track runs. Stay local when isolation isn't needed, or bind a workspace to a Daytona sandbox with explicit sync, lifecycle, and provider metadata. Extensions bundle skills, hooks, bridge adapters, and MCP servers. Ship them as zip files or via a GitHub registry. Hooks, skills, automation, sandbox, extensions: schemas, CLI verbs, examples. A hook is a TOML block in your config. Everything the daemon loads is inspectable with agh skill view , agh hooks list , and agh extension list . Webhooks in, sessions out. Responses stream back to the original thread. No serverless glue, no second runtime, the bridge adapter runs inside the daemon. Slack bridge:slackDiscord bridge:discordTelegram bridge:telegramMicrosoft Teams bridge:teamsGoogle Chat bridge:google-chatGitHub bridge:githubLinear bridge:linearPlatform slack / discord / tg agh daemon verify · route Agent claude / codex / ... Thread reply streamed updates Every bridge is a workspace-scoped adapter. One platform message maps to one durable session, so a user thread keeps its context across restarts. macOS and Linux. Install with Homebrew, npm, or Go. The full installation guide also covers the verified binary installer, Linux packages, and source builds. bash $ brew install compozy/compozy/agh bash $ npm install -g @compozy/agh bash $ go install github.com/compozy/agh@latest Create ~/.agh/config.toml and the default general agent before you start the daemon. bash $ agh install One local process, detached by default, exposing CLI, HTTP/SSE, and the web UI. bash $ agh daemon start Create the session from the repository you want AGH to manage so workspace resolution is explicit. bash $ agh workspace add "$PWD" --name current$ agh session new --workspace current --agent general AGH is the only approach here with an implemented cross-runtime protocol. The rest centralize coordination or skip it entirely. Memory-first stateful agents Letta agents in cloud or self-host 1 managed None, single agent Cloud-hosted or self-host Multi-agent orchestration framework Agents you author in Python your code In-process graph or crew Library you embed Hosted agent platform Managed agents behind an API managed Centralized routing Cloud-only Run + connect real agent CLIs Your existing ACP agents 26 ACP drivers agh-network/v0, implemented Local-first, single binary One binary. No infrastructure. Alpha runtime included.