cloudparts

AGNTCY

Open infrastructure for agents from different vendors to find, verify and message each other: a schema for describing agents, a federated directory, verifiable agent identities and an encrypted messaging layer that carries A2A and MCP.

Linux Foundation project
What it is
AI agents
Type
Open source
Made by
AGNTCY project (Linux Foundation), started by Cisco’s Outshift
Foundation
The Linux Foundation, joined July 2025. A series of LF Projects, LLC, donated by Cisco
License
Apache 2.0
Facts checked
28 October 2026

What it is

AGNTCY turns Cisco’s idea of an “open, interoperable Internet of Agents” into a set of components. The Open Agentic Schema Framework (OASF) describes an agent, MCP server or Agent Skill as a JSON record with a name, version, authors, skills and domains from fixed taxonomies, locators saying where to get it, such as a container image, and modules for A2A cards or MCP server details; OASF is modelled on the Open Cybersecurity Schema Framework (OCSF), and its schema server is a derivative of OCSF’s. The Agent Directory Service stores these records in an OCI registry such as Zot, addressed by a content identifier (CID), lets publishers sign them with Sigstore, and announces them over a distributed hash table, so directories run by different organisations can find each other’s agents by skill without one central registry. dirctl imports existing MCP registry entries, A2A agent cards and SKILL.md folders as records.

Two further layers handle trust and traffic. Identity gives each agent or MCP server an identifier backed by a W3C Verifiable Credential, the Agent Badge, issued either through an existing identity provider such as Okta or Duo or as a decentralised identifier (DID). SLIM (Secure Low-Latency Interactive Messaging), written in Rust, is a messaging overlay: agents connect outbound to SLIM nodes and are reached by name, payloads are end-to-end encrypted with Message Layer Security (MLS) even across shared routing nodes, and the same overlay carries A2A, MCP and OpenTelemetry traffic, point to point or to groups. Around these sit SDKs for observability and evaluation of multi-agent workflows and SHADI, a hardened runtime with gated secrets and OS sandboxing. The code is Apache 2.0, and the directory and SLIM protocols are written up as IETF Internet-Drafts.

Why teams pick it

  • Works with the protocols agents already use rather than replacing them: the directory lists MCP servers, A2A agents and Agent Skills, and SLIM carries A2A and MCP
  • Federated discovery: each organisation can run its own directory, and records are content-addressed and signed, so a record found through another peer can be verified
  • Encryption from agent to agent with MLS, so messages stay private even when they cross routing nodes run by someone else
  • Vendor-neutral governance under the Linux Foundation, with Cisco, Dell Technologies, Google Cloud, Oracle and Red Hat as formative members

Watch out for

  • Young and fast-moving: the directory reached v1.8.0 in October 2026, and its specifications are Internet-Drafts, not published standards
  • Many separate parts (schema, directory, identity, messaging, runtime, observability), each with its own CLI, SDKs and deployment, so adopting more than one is real work
  • It overlaps with registries and gateways in the Agentic AI Foundation, where MCP and A2A themselves now live, and with the official MCP Registry; AGNTCY is not part of that foundation
  • Directory records are only as trustworthy as their signatures and ownership claims; an unsigned record’s name is a label anyone can pick

How it fits with other technologies

Publish an agent record to the Agent Directory, then find it by skill · bash

# record.json is an OASF record: name, version, authors, skills, domains, locators, modules
dirctl validate record.json --url https://schema.oasf.outshift.com
CID=$(dirctl push record.json -o raw)   # store it; prints the content identifier
dirctl sign "$CID"                      # keyless Sigstore signature, stored as an OCI referrer
dirctl routing publish "$CID"           # announce it to other directories over the DHT

# Turn MCP Registry entries into records (review before pushing)
dirctl import --type=mcp-registry --url=https://registry.modelcontextprotocol.io/v0.1 \
  --filter search=github --limit 10 --dry-run --output-dir=./out

# Consumers: search this directory, or the peer-to-peer network, and pull a record
dirctl search --skill "*text_completion*" --trusted -o json
dirctl routing search "agent for summarizing financial reports" --limit 25
dirctl pull "$CID"

The main components#

Each lives in its own repository under github.com/agntcy and can be used without the others.

OASF

schema

The Open Agentic Schema Framework: a data model for describing agents, MCP servers and skills, with closed taxonomies of skills and domains and extension modules for A2A and MCP. Inspired by OCSF, with a schema server derived from OCSF’s.

Agent Directory

discovery

A federated registry: records are stored by CID in an OCI registry, signed with Sigstore, and published to a DHT that maps skills, domains, modules and locators to records. A runtime component can also discover agents running in Docker or Kubernetes.

Identity

trust

Identifiers for agents, MCP servers and multi-agent systems, each backed by a Verifiable Credential (the Agent Badge), issued through an identity provider or as a DID.

SLIM

messaging

Secure Low-Latency Interactive Messaging: a Rust data plane with MLS end-to-end encryption, name-based routing, point-to-point and group sessions, SLIMRPC, and SDKs for Python, Go, Java, Kotlin, .NET and Node.js.

See it in a real stack

No architecture uses AGNTCY yet. Drop it into the playground and wire it up yourself.Open playground →
Where it came from2025–2026

By early 2025, as Cisco’s incubation group Outshift put it, AI agents were “built in isolated silos, unable to collaborate across different frameworks and platforms”: there was no standard way for agents from different vendors to find each other, verify who they were talking to or exchange messages securely without custom integrations. Outshift conceived AGNTCY in January 2025 and launched it on March 6, 2025, with LangChain and Galileo as fellow core maintainers.

On July 29, 2025 the Linux Foundation announced that it would host AGNTCY, with Cisco, Dell Technologies, Google Cloud, Oracle and Red Hat as formative members and more than 65 supporting companies. The announcement stressed that AGNTCY works with A2A, which had recently moved to the Linux Foundation, and with MCP, making A2A agents and MCP servers discoverable through its directories.

In December 2025 the Linux Foundation formed the Agentic AI Foundation around MCP, goose and AGENTS.md, and A2A later joined it. AGNTCY did not: in October 2026 it is still a series of LF Projects, LLC hosted directly by the Linux Foundation, and is absent from the Agentic AI Foundation’s list of ten projects.

Why it’s built this way

Build on A2A and MCP rather than replace them
Agents already spoke those protocols, so AGNTCY adds the missing pieces around them: a directory that indexes MCP servers and A2A cards, identities for them, and a secure transport that carries their messages.
A federated directory instead of one registry
Organisations want to publish and control their own agent catalogues, so directories announce records to each other over a DHT and records are content-addressed and signed for verification.
Reuse OCSF’s schema machinery
OCSF had already solved versioned, extensible schemas with taxonomies and a schema server in security, so OASF copied its approach and server for describing agents.

Timeline

  1. 2025

    Cisco’s Outshift launches AGNTCY on March 6, with LangChain and Galileo

  2. 2025

    The Linux Foundation announces it will host AGNTCY, on July 29

  3. 2026

    Agent Directory v1.8.0 released on October 15; the project is still hosted directly by the Linux Foundation

Sources11 · facts checked 28 October 2026

Improve this page on GitHub