sdk/

Native SDKs

This directory holds the SDKs that let an agent join an Agent Mesh from inside its own process, without a agentmesh-node sidecar. Three languages are in scope:

• meshnet/ — github.com/google/agentmesh/sdk/meshnet, Go package modeled on tailscale.com/tsnet, built on go-libp2p with in-process A2A listeners (private P2P and public external HTTP ingress), WhoIs, outbound mesh *http.Client, local progressive-discovery MCPHandler(), and zero-secret OpenAI /v1 InferenceHandler().
• js/ — agentmesh-p2p, TypeScript for Node.js, built on js-libp2p.
• python/ — agentmesh-p2p (import agent_mesh), built on py-libp2p.

The motivation is issue #480. The sidecar works, but it makes every agent deployment manage a second process, a local port and a token, which does not fit serverless platforms and splits tracing between the agent and the network layer. An earlier attempt at a Python package (sam-mcp-python) only wrapped the sidecar's HTTP API and was removed in favour of this work.

All three SDKs speak the mesh protocol directly. There is no SDK-specific endpoint anywhere in the mesh; a member built with an SDK is, to every other member and to the control plane, a peer like an agentmesh-node.

The Go SDK (sdk/meshnet) embeds a full Go member runtime (tailscale.com/tsnet style): Server.ListenA2A announces a2a://<name> (and optional external HTTP ingress) to the router directory unless NoAnnounce: true, answers system://mesh.catalog, and exposes WhoIs, HTTPClient(), Discover(), AttenuateToken(), MCPHandler(), and InferenceHandler() in-process; see sdk/meshnet/README.md.

In JavaScript (sdk/js) and Python (sdk/python), an SDK member is an agent, not a service provider. It calls what the mesh offers: MCP tools, inference and A2A services that agentmesh-nodes (or sdk/meshnet Go servers) publish and the discovery table names. Inbound, it accepts one thing: A2A requests for itself, as a2a://<name>, reached by peer ID through a router. It publishes no service, writes no directory record and reports no catalog; a tool, a model or a named agent in JS/Python that others should find by name runs behind an agentmesh-node. The reasons are in Agents, not services.

What exists today

Milestones 1 to 5 are implemented and tested in both languages:

A member built this way is on the mesh and uses it in both directions: it finds a service in the directory and calls it through the router, verifying the provider on every call, and it answers A2A requests for its own agent from a agentmesh-node or another SDK member that names it by peer ID, authorizing every caller with the same Datalog the Go node evaluates. It follows the control plane while it runs: a ban reaches it as a gossip event within a second and again with the next pull, and a key rotation makes it refresh its credential under the new key.

Facts about the libp2p implementations that the SDKs work around, each pinned by a test:

• py-libp2p's TLS transport advertises early muxer negotiation in ALPN but cannot complete it (Python's ssl has no ALPN select callback), and go-libp2p then refuses the mux upgrade. The Python host sets no ALPN muxer list, so the muxer is negotiated with multistream-select (sdk/python/src/agent_mesh/host.py).
• js-libp2p's @libp2p/tls reads the libp2p extension from extensions[0] of the peer certificate. py-libp2p's default certificate puts BasicConstraints and KeyUsage first, so the Python host uses a certificate template with only the libp2p extension.
• A router grants a relay reservation only after the auth handshake on the same connection, and drops it with the connection. The Python SDK carries the relay protocol itself (sdk/python/src/agent_mesh/relay.py, sdk/python/proto/circuit.proto) on top of py-libp2p's raw-connection upgrade, so the session decides when to reserve, renew and, after a dropped connection, authenticate and reserve again; py-libp2p's own RelayDiscovery reserves again only after the expiry has passed, when the router has already dropped the slot.
• go-libp2p's relay leaves private addresses out of a reservation, so a router on loopback (every test) grants a reservation that lists no address. js-libp2p falls back to the connection's address; the Python SDK does the same.
• js-libp2p reserves a relay slot when it starts listening on <relay>/p2p-circuit, and a router refuses that before the auth handshake. The JS SDK starts the listener after the handshake, through the transport manager, which is not on the public Libp2p interface (libp2p/js-libp2p#3645).
• go-libp2p's relay grants a reservation for one hour and drops it when that passes, and with the connection it was made on; a member that still advertises the relayed address is then unreachable (NO_RESERVATION). js-libp2p's listener renews on its own, but only on a connection the router still holds the admission of, and does not reserve again on a static relay after the connection dropped (libp2p/js-libp2p#3601); the JS SDK runs the handshake again on every new connection to a router and, when the relayed address is gone, authenticates and reserves again within thirty seconds. The Python SDK renews two minutes before the expiry the router returned, and within thirty seconds of the connection to that router going, running the auth handshake again first.
• py-libp2p's yamux gives a stream's backlog slot (256 per connection) back when the stream closes since 0.8, but not for the trio.Cancelled of a caller's deadline while the SYN waits on a stalled connection. Every stream the SDK opens goes through open_stream, which bounds new_stream with the caller's deadline and hangs up on a connection that cannot open a stream in time, so the call fails instead of parking and the slots go with the connection.
• py-libp2p's yamux drops a reset stream's bookkeeping once its buffer is drained, and a bounded read then reports the end of the stream. The Python /libp2p-http client reads the muxed stream's reset flag on EOF, so a body cut by a reset is an error and not a body that ended.
• py-libp2p's WebSocket transport resolves a /dns4/<host>/tcp/443/wss address to its IP before dialing and names the IP in the TLS SNI and the Host header, which a TLS-terminating edge answers with 403; the Python host dials such addresses by name (libp2p/py-libp2p#1549). It also dials wss with certificate verification off unless given a TLS context (libp2p/py-libp2p#1550); the host gives it the system roots.
• A member keeps no DHT. Both SDKs ask a router they hold over /mesh/directory/1.0.0 (discovery.ts, agent_mesh/discovery.py), one DirectoryRequest per stream, and the router answers from its own records and the routers closest to the key. sdk/testdata/service_keys.json pins the keys both SDKs derive against api/directory_keys.go.
• go-libp2p-http is plain HTTP/1.1 on a stream, one request per stream, with Host set to the peer ID. The JS SDK bridges the stream to a Node duplex and runs Node's HTTP server and client on it; the Python SDK drives h11 over the stream. Both read the response body as it arrives, which an A2A message/stream (server-sent events) needs.
• URL parsers lowercase the host (WHATWG URL, httpx), and a base58 peer ID is case-sensitive. The URL an HTTP client uses for a peer's service therefore carries the peer ID in the path, http://mesh/mesh/<peer-id>/<type>/<name>/<path>, the shape of agentmesh-node's egress proxy and of an agent card rewritten for the mesh; the host is ignored.
• A member that publishes nothing has announced no address, so nothing in the directory names one. agentmesh-node dials /p2p/<router>/p2p-circuit for every router it authenticated with when it knows nothing else about a peer, as both SDKs do for a peer named by ID (internal/node/mcp.go, preparePeerAddrs).
• Every Go component runs GossipSub with StrictSign. @libp2p/gossipsub (the libp2p-maintained package; @chainsafe/libp2p-gossipsub stops at libp2p 2) and py-libp2p's GossipSub with strict_signing=True both exchange messages with agentmesh-router. py-libp2p's Pubsub learns of peers through a notifee it registers when constructed, so it is built before the first connection; it also opens a meshsub stream to every new peer and its host logs an error for peers that do not run pubsub, which is expected on a mesh with plain peers.
• py-libp2p's connection gate is address-based, not peer-based, so the Python SDK enforces a ban by disconnecting the peer and refusing its handshakes, requests and dials; js-libp2p's connectionGater also denies the connection itself, as agentmesh-node's gater does.

Wire contract

Everything below is what the Go implementation does today, with the source file that defines it. The SDKs must match it byte for byte; when a Go change alters one of these, the SDKs change with it.

Identity

• The identity is an ed25519 key pair. The public key travels in the libp2p protobuf encoding PublicKey{Type: Ed25519, Data: <32 bytes>}, which is the bytes 08 01 12 20 followed by the key. Private keys persist as PrivateKey{Type: Ed25519, Data: <seed || public>}, bytes 08 01 12 40 followed by 64 bytes; this is what agentmesh-node writes to its store.
• The peer ID is the identity multihash (00 24) of the public key encoding, written in base58btc. The control plane refuses an enrollment whose peer_id is not derived from public_key (internal/controlplane/server.go, pID.MatchesPublicKey).

State directory

What a member keeps between runs is api.MemberCredential plus its private key, and every implementation persists the same two things. The SDKs keep them in a directory: identity.key is the libp2p PrivateKey encoding above; credential.json is MemberCredential as protojson with proto field names (control_plane_url, biscuit, expire_time as RFC 3339, trusted_keys[].public_key, issued_under_keys, router_addresses, oidc_session), written with mode 0600 in a directory of mode 0700. An unknown field is an error. control_plane_url has no trailing slash. In a browser the JS SDK keeps the same two records in an IndexedDB database named after the state location, kept by the browser for the page's origin. agentmesh-node keeps the same message in agent.db, and agentmesh-node state export|import <dir> moves a member between the two (internal/node/statedir.go). TestNativeSDKExamples resumes each SDK's directory with the other SDK and with an imported agentmesh-node. Fields an SDK does not use (receive_time, oidc_session) are carried through on save, so a directory survives a round trip through any implementation.

enroll resumes the saved credential when it is for this control plane, has more than five minutes left, and the control plane still serves a key in trusted_keys: /keys is pulled and verified against the saved set before the member is returned, and the keys it names are adopted. A member off for longer than the key grace period holds a credential no current key signed and a trusted set no current key is in, so neither verifies; with a token given, enroll enrolls again as the saved identity, and without one it raises CredentialRetiredError. A control plane that cannot be reached leaves the saved credential in force; the pull before join tries again. TestKeyRotationWhileMembersOffline and TestKeyRotationPastGraceWhileMembersOffline in tests/integration run both cases against real routers and a real control plane.

Control plane (protobuf over HTTP)

All requests and responses are application/x-protobuf bodies of the messages in api/agentmesh.proto. Bodies are capped at 1 MiB on both sides.

• <ts> is the request's challenge_unix_ms, unix milliseconds, and must be within 5 minutes of the control plane's clock (challengeMaxAge). When the control plane answers 401 with "stale or invalid challenge timestamp", a client reads the response's Date header (Access-Control-Expose-Headers: Date), recomputes <ts> from it, and retries once; other 401s are not retried.
• The endpoints above answer a CORS preflight and mark their responses for any origin (Access-Control-Allow-Origin: *), so a page on another origin can call them. They authenticate by what the request carries, a token in the body or a biscuit as a bearer plus the signed challenge, never by a cookie. The operator plane (/admin/*, /user/*) and /routers/lease do not. Challenges are defined in api/network.go. It is the one instant on the wire that is an int64: it is the number in the signed text. Every other instant (expire_time, sign_time, event_time) is a google.protobuf.Timestamp, and a receiver rejects a message whose required instant is unset.
• A bootstrap enrollment answers PENDING until an operator approves it, unless the control plane runs with auto-approval. The client polls /enroll/status at the returned poll_interval_seconds.
• /refresh redeems only the last biscuit the control plane issued for the peer. A client must persist the new biscuit before using it; losing it means re-enrolling.
• /keys signatures cover the deterministic protobuf encoding of KeysResponse{public_keys, sign_time} with signatures cleared, one signature per key by that key. The receiver accepts the set when a key it already trusts (the enrollment key) vouches for it and sign_time is within 5 minutes (api/trust.go). Rotation keeps several keys valid, so a member must trust the whole set to verify peers enrolled under a retiring key.
• A plaintext http:// control plane URL is accepted only for a loopback host (api.ValidateControlPlaneTransport). The SDKs expose the same explicit opt-in (allowInsecure / allow_insecure).
• Bootstrap tokens are secrets. The SDKs read them from a file or a value the caller already holds; the conformance runners take a path in the environment. Nothing takes a token as a command-line argument.

libp2p host (internal/node/node.go)

• Transports: libp2p.DefaultTransports (TCP, QUIC, WebSocket).
• Security: TLS first, Noise accepted (libp2p.Security twice, in that order, on agentmesh-node and agentmesh-router). Both bind the connection to the peer ID. The Node and Python SDKs offer the same two in the same order and land on TLS with a Go peer and with each other; in a browser the JS SDK offers Noise alone, since a page cannot run libp2p's TLS, and a Node or Python member reached through a relay meets it on Noise. js-libp2p has both; py-libp2p gained TLS in libp2p/py-libp2p#831 and has passed the libp2p transport interoperability suite against the other implementations since libp2p/test-plans#798.
• Muxer: yamux (libp2p default).
• Relay: circuit relay v2, with the routers as static relays; hole punching is on. A node behind NAT is reached through a router.
• Directory: /mesh/directory/1.0.0 to a router the session holds; a member keeps no DHT and speaks no Kademlia.

Streams

Four protocols. The first two and the last start with a length-prefixed protobuf frame; framing is the go-msgio varint style (unsigned varint byte length, then the bytes), with a 64 KiB cap on the first frame.

• /mesh/auth/1.0.0 (HandleAuthHandshake, and the router's equivalent). Client sends AuthFrame{biscuit}; server verifies the biscuit against the control plane keys, checks it is bound to the connection's peer ID and not revoked, and answers AuthResponse{success, biscuit} with its own credential. The client verifies that credential the same way and, for a router, requires the router role. This is how a member joins: connect to a router, run this handshake, and the router admits it to the relay and gossip. Both SDKs answer it too, so a agentmesh-node can verify them before calling.
• /mesh/mcp/1.0.0 (WithBiscuitAuth then HandleMCPStream). Client sends AuthFrame{biscuit, target_service: "mcp://<service>", agent}; server authorizes the request with the Biscuit authorizer described below and answers AuthResponse. The stream then carries MCP JSON-RPC messages, each one varint-length-prefixed (StreamTransport in internal/node/gate.go). An empty target_service selects the node's own catalog tools. The SDKs are clients of this protocol only; they serve no MCP.
• /libp2p-http (go-libp2p-http; StartIngressServer and the egress proxy in internal/node). Plain HTTP/1.1 on a stream, one request per stream, Host set to the peer ID, the caller's biscuit in X-Mesh-Biscuit and the agent it speaks for in X-Mesh-Agent. The path is /<type>/<name>/<upstream>; the server authorizes <type>://<name> with the same authorizer, then forwards <upstream> to the service with those two headers stripped and X-Peer-Id set to the verified caller. The SDKs are clients of this for inference://, a2a:// and egress:// services (an egress destination is one a agentmesh-node serves for a hostname outside the mesh, as the control plane assigned it; the SDK reaches it as any other service, /egress/<hostname>/<path> on that node), and servers of it for their own agent, a2a://<name>, only. Bodies are framed by Content-Length or chunked transfer coding; a response with neither runs to the end of the stream. The JS SDK frames these itself (http1.ts), so the same code runs in a browser.
• /mesh/goaway/1.0.0 (HandleGoAway). A router opens it to a member it will no longer hold and sends one RouterGoAway{reason, retry_after}: DRAINING when it is stopping, OVERLOADED when it holds more members than its high watermark. Both SDKs register the protocol, act only on a message from a router the session holds, close the stream, record the router as shunned for retry_after (default 5m, cap 1h), and immediately attach and reserve on another router from the catalog; the router closes the connection 5 seconds after the message either way. When a held router's connection drops without a go-away, the member redials it first (2s, 4s, 8s) and, if it does not come back or refuses auth, shuns it and tops up from the catalog.

Discovery (api/directory_keys.go, internal/node/directory_client.go)

A provider announces a service to a router it holds as a record under the key sha256("mesh:service:<type>[:<name>]"), once for the type and once for the name (<type> is mcp, inference, a2a or egress), with the addresses it is reached at; the router keeps the record and places it on the routers closest to the key. A consumer asks a router it holds for the providers of the same key, dials one and opens /mesh/mcp/1.0.0. The SDKs look records up and announce none.

A peer that announced nothing is still reachable by peer ID: every router relays for the peers it admitted, so a caller dials <router>/p2p-circuit/p2p/<peer> through each router that admitted it. Both SDKs do this for a bare peer ID, and agentmesh-node does it when it knows no other address for a peer (preparePeerAddrs in internal/node/mcp.go). This is how an SDK agent is reached.

Authorization on the provider side (internal/node/middleware.go)

A node that serves a tool evaluates the caller's biscuit with an authorizer that has, in this order: the biscuit's own authority facts (node, role, label), injected facts service(<type>, <name>), connection_peer_id(<peer>) and time(<now>), the baseline checks, rules and policies from api/datalog.go, the target_fact facts derived from the provider's own biscuit, and the mesh policy rules. Deny by default.

All of that Datalog is text every Biscuit implementation parses, and none of it is derived in the SDK:

• The baseline (checks, rules, policies, fact names) is generated from api/datalog.go by hack/gen-sdk-datalog into sdk/js/src/gen/datalog.ts and sdk/python/src/agent_mesh/_gen/datalog.json; hack/verify-sdk-generated.sh fails when they are stale.
• The mesh policy arrives rendered: GET /policies answers PolicyConfigGetResponse{datalog_rules}, one rule per entry, rendered by the control plane with api.BuildPolicyRules. agentmesh-node adds the same text; no member compiles roles and bindings itself. An unparseable entry rejects the whole response.
• Presence-only facts carry the single term true (target_unrestricted(true), granted_service_all_types(true), agent_authorized(true), …). biscuit-go accepts a predicate with no terms; biscuit-rust, and so biscuit-python and biscuit-wasm, do not.
• An unconditional rule is written head <- true, which every parser accepts as a rule (role("dev") <- true for a mesh:system:authenticated binding).

Agents, not services

An SDK member calls services and accepts A2A requests for its own agent. It does not publish services: no MCP server, no named inference or A2A service, no directory record, no catalog entry. The reasons:

• The agent is the principal; a service is infrastructure. A harness (LangChain, Claude Code, a browser agent) is an MCP client. The tools it should see through the mesh are the ones agentmesh-nodes publish. Something that must be reachable by name, a tool, a model, a named agent, is a deployment concern, and the sidecar model exists for it: the agentmesh-node beside it publishes the service and enforces the policy once, in Go.
• Publishing is the expensive part. Being a provider needs an announcement loop, a catalog report and records that go stale on every rollout. Accepting requests for the agent itself needs the relay reservation, the auth handshake and the authorizer, all of which a member needs anyway to be verified by a agentmesh-node before it is called.
• One inbound protocol. A2A is the protocol for talking to an agent, and the A2A SDKs serve one agent card each. One ingress means one authorizer path to audit in three languages instead of two ingress protocols times a registry of services.
• The browser. A browser cannot listen. An agent in a browser is reachable through a router's relay and nothing else, which is what accept_a2a is. The JS SDK runs in a page: it reaches the router over WebSocket (wss when the router sits behind a TLS-terminating edge, as agentmesh-one --tunnel puts it), secures the connection with Noise, keeps its state in IndexedDB and answers its agent with a fetch handler.

What an SDK agent is on the wire: a peer with a relay reservation on a router, answering /mesh/auth/1.0.0 and /libp2p-http for a2a://<name>, with no directory record and no catalog entry. To a agentmesh-node it is a peer whose address nobody announced, reached through the router that admitted both, and whose credential it verifies before forwarding /mesh/<peer>/a2a/<name>/.... Routers, the control plane and the Datalog need no special case.

Two agents that both wrote nothing down still meet: A learns B's peer ID (an invite, an agent card, a coordinator), dials it through a router, both present their credentials, and A opens the A2A conversation on that connection; libp2p's secure channel is end to end, so the router carries ciphertext. An agent that must be found by name runs behind a agentmesh-node. Two agents that both can only call out (two browsers) meet at a third agent behind a agentmesh-node that both call.

Plan

Each milestone lands in both languages with its tests before the next one starts. Where a milestone needs a Go change, the Go change lands first.

Milestone 1 — identity and enrollment (done)

Described above. Tested by unit tests in each SDK against a fake control plane and by tests/integration/sdk_enroll_test.go, which runs each SDK's conformance runner against a real control plane in manual-approval mode, approves the request through /admin, and checks the biscuits the SDK holds against the control plane's records.

Milestone 2 — join the mesh (done)

• libp2p host per SDK, configured as agentmesh-node's: TCP and WebSocket, TLS, yamux, identify, circuit relay v2 client. The testnets' routers listen on TCP; agentmesh-one's router is a WebSocket listener on the port that serves its API (/ws, or /wss behind a tunnel), so a member reaches it over WebSocket or not at all; tests/integration/standalone_sdk_test.go holds both SDKs to it. JS: libp2p, @libp2p/tcp, @libp2p/websockets, @libp2p/tls, @chainsafe/libp2p-yamux, @libp2p/circuit-relay-v2, @libp2p/identify. Python: libp2p>=0.8 (security.tls, stream_muxer.yamux) plus the SDK's own relay client; py-libp2p is trio-based, so join() is an async context manager.
• Biscuit verification of a peer's credential: signature under any trusted control plane key, authority block only, expiry, binding to the connection peer ID, roles and labels. JS: @biscuit-auth/biscuit-wasm, instantiated by hand so no Node flag is needed; Python: biscuit-python. sdk/testdata/biscuit_vectors.json holds tokens minted by internal/identity (valid, other peer, expired, untrusted key, appended block, garbage) and both verifiers agree with the Go one on all of them.
• /mesh/auth/1.0.0 on both sides. Join connects to the routers in the credential, runs the handshake, requires the router role under the key that verified the token, and reserves a relay slot; the router grants one only to an authenticated peer. Inbound handshakes are answered the way HandleAuthHandshake does: a credential that does not verify gets a closed stream and nothing else.
• session.authenticate(addr) connects to any peer, directly or through a router (/p2p-circuit), and runs the handshake; session.authenticatedPeers is the admitted set. Background refresh is driven by the biscuit's expiration, as StartRenewalLoop does.
• Tests. Unit: each SDK joins an in-process fake router (its own libp2p with the handshake and a relay), and refuses a relay without the router role, a router that trusts another control plane, and a refused reservation. Integration: tests/integration/sdk_mesh_test.go runs one mesh with a control plane, a agentmesh-router, a agentmesh-node and one member per SDK, all real, then walks the connectivity matrix: every member on the router's lease; each SDK member verifies the agentmesh-node and the other SDK member on both the direct and the relayed path, and is verified by them; the agentmesh-node reaches each member through the router; an admitted Go peer does the same and sends a forged frame; an address behind the router for a peer that is not on the mesh fails. About 7 seconds.

Milestone 3 — call tools (done)

• Discovery: session.discover("mcp://calc"), or (type, name), or a type alone, asks a router the session holds, over /mesh/directory/1.0.0, for the providers of the service key api/directory_keys.go derives.
• Naming a peer: every call takes a Peer, which is a provider discover returned, a peer id, or a multiaddr. For the first two the SDK dials the advertised addresses and <router>/p2p-circuit/p2p/<peer> through every admitted router, as preparePeerAddrs in internal/node/mcp.go does; a multiaddr is dialed as given.
• /mesh/mcp/1.0.0 client: session.openMCP(peer, "mcp://<name>") sends the AuthFrame naming the service, verifies the provider's credential and the caller's required labels (checkPeerLabels: every pair must be attested, as api.LabelCheck joins them with ,), then runs the official MCP client over the varint-framed stream. JS: a Transport for @modelcontextprotocol/sdk; Python: a pair of memory streams pumped to and from the libp2p stream for mcp.ClientSession. "" as the target is the provider's own catalog (list_local_services, get_mesh_info).
• Egress floor: join({ egressRequireLabels }) (join(egress_require_labels=)) is agentmesh-node's egress.require_labels for an SDK member: every provider the session calls must attest all of them, on top of a call's required labels, on every outbound call however the peer was named, MCP and HTTP alike. It is the same rule as the caller's requirement (api.LabelCheck), checked separately so no call can reach it. Stated once at join and held for the session; a call cannot waive or widen it. The three implementations agree on it, as they do on the caller's requirement. The HTTP path (request, fetch, MeshTransport) verifies the provider with or without a floor, through the mutual /mesh/auth/1.0.0 handshake, as agentmesh-node's VerifyPeerLabels does before its egress proxy sends anything; a positive verdict is kept per peer for five minutes (labelGateTTL); a refusal is not kept. An unmet floor is a LabelsNotSatisfiedError naming the floor.
• session.listTools(peer, service) and session.callTool(peer, service, tool, args) on top of that.
• Tests. Unit: each SDK calls a tool on an in-process provider that serves /mesh/mcp/1.0.0 as agentmesh-node does with the official MCP server behind it, reads its catalog, and is refused for missing labels, an untrusted provider, a forged caller credential and an unknown service. Integration: the mesh of sdk_mesh_test.go has an MCP backend behind the agentmesh-node (calc, tool add); each SDK member discovers it through its router, lists and calls add through the router, reads the node's catalog and is refused a service the node does not have. The runners took discover, tools and call commands for it.

Milestone 4 — be an agent (done)

• Datalog as the contract, Go side: the baseline moved into a generated artifact both SDKs embed, the control plane renders the mesh policy as datalog_rules, and agentmesh-node consumes that text like any other member. See the authorization section above.
• Provider authorizer in each SDK (authorizer.ts, authorizer.py), built from the artifact and datalog_rules, evaluated on every inbound /libp2p-http request in the order internal/node/middleware.go uses.
• /libp2p-http server for the agent: the path is /<type>/<name>/<upstream>, the biscuit is X-Mesh-Biscuit; after authorization for <type>://<name>, a request for the agent's own a2a://<name> goes to an A2A server beside the process or a handler in it, with the biscuit and agent headers stripped and X-Peer-Id set to the verified caller; anything else is 404, after authorization, so an unauthorized caller learns nothing. JS also takes a Node request listener, which is what an Express app with the A2A SDK's handlers is, and runs it on the ingress's own http.Server.
• session.accept_a2a(target, name=) / session.acceptA2A(spec) fetches the mesh policy, re-reads it on agentmesh-node's sync interval and starts answering. It announces nothing. One agent per session.
• /libp2p-http client for callers built on HTTP libraries: open_http_request / fetchOverStream return once the headers are in and stream the body; MeshTransport (httpx) and session.fetch() (fetch) carry a client's requests to the peer a mesh URL names. The A2A SDK's client takes either without changes. The card of an agent behind a agentmesh-node names the agent's own address; a GET of the well-known card path or of the service root is answered the way the node's egress proxy answers it: the SDK fetches the card itself, with identity encoding, and serves it rewritten (rewriteAgentCard / rewrite_agent_card): HTTP interfaces point at the mesh URL, gRPC ones are dropped, signatures go. Streaming stays as the agent declares it, since the transport streams.
• agentmesh-node: a peer it knows no address for is dialed through every router it authenticated with, so its egress proxy reaches an agent by peer ID (preparePeerAddrs).
• Examples with the official A2A SDKs (a2a-agent.ts, a2a-call.ts; a2a_agent.py, a2a_call.py): the A2A SDK's server is the listener acceptA2A runs (JS) or a uvicorn process accept_a2a forwards to (Python), its card names the agent by its mesh URL, and the A2A SDK's client runs on session.fetch() or MeshTransport. The A2A SDKs are dependencies of the examples only: dev dependencies in sdk/js, the examples extra in sdk/python.
• Tests. Unit: each SDK's ingress behind an in-process host, called with its own clients: a forwarded A2A server that sees X-Peer-Id and never the biscuit, an SSE body read event by event, an in-process handler (and a Node listener), an httpx client on MeshTransport, and refusals for a role that grants nothing, an ungranted type, another agent name (404), a missing biscuit and a dotted path; plus the authorizer alone against the decisions internal/node/middleware_test.go pins, and agentmesh-node's router fallback. Integration: in the mesh of sdk_mesh_test.go each SDK member accepts a2a://agent; the agentmesh-node reaches it by peer ID through its egress proxy, the other SDK member does the same with no lookup, a lookup for a2a://agent finds nothing, the member refuses a /mesh/mcp/1.0.0 stream, and a Go peer whose role the policy grants nothing gets 403 and is answered once it presents a node-role token. About 9 seconds for the whole mesh. TestNativeSDKA2A runs the A2A SDK agents and calls each from the other language with the A2A SDK's client; the answer names the caller's peer, and a agentmesh-node fetches the card through its egress proxy.

Milestone 5 — parity and release (done)

• Control plane pull (session.sync(), mesh.syncControlPlane()), as SyncControlPlane in internal/node/controlplane_sync.go: /keys verified against the keys already trusted, so whoever answers the URL cannot become the trust root; a credential refresh when a key trusted now was unknown when the credential was issued (issuedUnderKeys in the credential file); /info for router addresses and the ban set, reconciled with the rule that a ban recorded after the request went out survives an answer that omits it; the mesh policy while accepting callers. Runs once shortly after join, then on agentmesh-node's interval with the same jitter, and whenever an event triggers it.
• Gossip events on /mesh/events/v1: the topic validator rejects anything not signed by a trusted control plane key and ignores stale events; BANNED evicts the peer at once, KEY_ROTATION adopts the key and triggers a pull, POLICY_UPDATE triggers a pull. Bans are not persisted; a restarted member reads them from /info.
• Ban enforcement: the auth handshake and the A2A ingress refuse a banned peer before looking at its token; connect() refuses to dial one; its connections are dropped when the ban lands.
• Publishing: .github/workflows/sdk.yml stamps the release tag's version on both packages (hack/sdk-version.sh) and publishes agentmesh-p2p to npm and agentmesh-p2p to PyPI through trusted publishing. A prerelease tag (v0.1.0-rc.4) publishes under the npm dist-tag next, a stable tag under latest; PyPI needs no tag, pip skips prereleases on its own. When a publish job (publish-sdk-js or publish-sdk-python) fails after the GitHub release exists, re-run the failed job, or run the workflow by hand from the Actions tab with the tag and target SDK as input; it checks out that tag and publishes the selected SDK. One-time setup by a package owner: on npmjs.com, publish agentmesh-p2p 0.1.0 once by hand (cd sdk/js && npm publish --access public; the trusted-publisher settings live on the package page, which exists only after that), then register google/agentmesh with workflow sdk.yml under the package's Settings, Trusted publishing; on pypi.org, add a pending publisher for project agentmesh-p2p with the same repository and workflow (environment release), which reserves the name and lets the first tag create the project. No publishing token is stored in the repository.
• Docs: site/content/docs/guides/native-sdks.md.
• Tests. Unit: a fake control plane rotates its key and bans a peer; each SDK learns both from a pull, refreshes under the new key and only under it, drops and refuses the banned peer, and lifts the ban when the control plane does; BanSet and verifyMeshEvent alone. Integration: the mesh of sdk_mesh_test.go gets the control plane's real event publisher; an enrolled Go peer admitted by every member is banned with POST /admin/revoke and each member learns it from the gossip event through the router without pulling, refuses the peer's next handshake, and a pull agrees; then the control plane rotates its key and announces it, and every member ends up trusting both keys, holding a credential that verifies only under the new one, and still authenticating with the agentmesh-node.

On the testnets

Both testnets run the example programs, unchanged, as canaries beside the agentmesh-node ones (.github/k8s/agentmesh-sdk-canary-template.yaml): two pairs, an agent written with one SDK and the official A2A SDK (a2a-agent.ts, a2a_agent.py) and a caller written with the other in one pod, enrolled with the pod's projected service account token through AGENTMESH_JWT_PATH. The agent accepts a2a://agent; the caller reads its peer ID from the line it prints, through a volume the pod shares, sends it a message with the A2A SDK's client every five minutes and is Ready while the last answer named the caller, so the Deployment's availability says whether an agent is still reachable after hours on the mesh. Two CronJobs cross every implementation boundary every 15 minutes and once per rollout: the agentmesh-node cold-path probe (agentmesh-probe-cronjob-template.yaml) runs an agent per SDK as sidecars and fetches each one's agent card by peer ID through the egress proxy (node → SDK), and the SDK cold-path probe (agentmesh-sdk-probe-cronjob-template.yaml) runs each SDK's callers against the everything canary by name (SDK → node) and the other SDK's agent by peer ID, the card with the mesh SDK's client and a message with the A2A SDK's (SDK → SDK). The images (Dockerfile.agentmesh-sdk-js, Dockerfile.agentmesh-sdk-python) are built per commit by deploy.yaml, so bananas runs the SDKs at the same commit as the Go components they talk to.

Later

• The connector interface for platforms (issue #480): Attach, Detach, Refresh, Status and the agent bundle, so a scheduler places SDK members the way it places agentmesh-box sandboxes.
• Reading the control plane's events also for router address changes, and re-dialing a router that replaced another.

Non-goals

• Rewriting the control plane, router or sandbox components; they stay Go.
• Publishing services from an SDK: an MCP server, a named inference or A2A service, a directory record or a catalog entry. That is agentmesh-node's job; see Agents, not services.
• Any SDK-only wire protocol. If an SDK needs something the Go node does not speak, the Go node learns it first.

Working on the SDKs

# Regenerate protobuf bindings and the Datalog artifact after editing
# api/agentmesh.proto or api/datalog.go
./hack/gen-sdk-proto.sh

# JavaScript
cd sdk/js && npm ci && npm test && npm run build && npm run examples

# Python
python3 -m venv sdk/python/.venv
sdk/python/.venv/bin/pip install -e 'sdk/python[test]'
sdk/python/.venv/bin/pytest sdk/python/tests

# Both against a real control plane, router and agentmesh-node, and the example
# programs the docs embed against the same
go test ./tests/integration -run TestNativeSDK -v

# The JS SDK in a browser page against agentmesh-one, in Chromium (Playwright);
# also run by `make ui-test`
cd sdk/js && node scripts/bundle-browser.mjs examples/browser/app.js build/browser-example
cd tests/ui && npm ci && npx playwright install chromium && npx playwright test browser-sdk

make sdk-test runs all of the above. The integration tests skip an SDK whose toolchain is missing, and CI (.github/workflows/sdk.yml) installs both so nothing skips there. The SDK members in the mesh test are the runners sdk/js/src/conformance-join.ts and sdk/python/src/agent_mesh/conformance_join.py: each joins, prints what it holds, then takes JSON commands on stdin (auth, discover, tools, call, accept, http, peers, sync, banned, quit) so the Go test can drive both languages through the same script.

The programs the package READMEs and the Native SDKs guide show are sdk/js/examples/*.ts and sdk/python/examples/*.py. The Markdown embeds them between <!-- embed: <path> --> and <!-- /embed --> markers; make sdk-docs (go run ./hack/gen-sdk-docs sdk site/content/docs) copies the files in, hack/verify-sdk-generated.sh fails when a copy is stale, and TestNativeSDKExamples runs the files themselves against a mesh. Edit the example, regenerate, commit both.

sdk/js/.npmrc pins the public registry so package-lock.json never resolves packages through a local mirror; the verify script checks the lockfile too.

Interoperability facts that the tests pin: sdk/testdata/identity_vectors.json holds key encodings, peer IDs and challenge signatures produced with go-libp2p, sdk/testdata/biscuit_vectors.json holds tokens minted by the control plane's code, and every SDK reproduces or agrees with them.

The package READMEs (sdk/js/README.md, sdk/python/README.md) are the registry pages on npm and PyPI: they hold what a user of the package needs and nothing about this repository's internals. Everything below is for contributors.

Layout

Both SDKs mirror the same Go code, one module per concern; the Go file each one follows is named so a change on one side can be carried to the others.

Unit tests sit beside the code (*.test.ts, tests/test_*.py) and build a fake control plane and router in the process; the real ones are exercised only by tests/integration.