Architecture Comparison · ACK Protocol
ACK Protocol vs. A2A vs. MCP: where each layer fits
MCP connects an agent runtime to tools. A2A exposes an organization’s live task endpoint. ACK Protocol sits in the missing layer between independent owners: offline-durable, cryptographically receipted work delivery.
Short answer: ACK Protocol does not replace MCP or A2A. An agent uses MCP (pull-only) to talk to its own provider, independent providers use ACK Protocol for durable signed delivery and work-evidence records, and organizations can optionally bridge A2A at their public doorway.
1. Side-by-side comparison
The ACK Protocol column describes the Wire v2 design, which is not yet implemented. Today’s prototype already signs work events and storage receipts as ES256 JWS with zero LLM calls on intake.
| Dimension | MCP | A2A (v1.0) | ACK Protocol (Wire v2 draft) |
|---|---|---|---|
| Primary boundary | Agent runtime ↔ local/remote tools | Client ↔ live organization agent endpoint | Independent provider ↔ independent provider |
| Offline delivery | No (stateless session) | Assumes reachable server; not reliable store-and-forward | Yes — verify, persist durably, then issue signed receipt |
| Idempotency | Caller-defined per tool | Optional (MAY be idempotent) |
Mandatory (planned) — immutable JWS bytes on retry |
| Signed storage receipt | No | No (signs Agent Cards only) | Yes — ES256 receipt returned with zero LLM calls |
| Typed work evidence | No | Two-party task state only | Planned — separate signed facts: delivered, accepted, completed, verified |
| Cross-owner permissions | Single-owner OAuth/token | Endpoint-level auth | Scoped, revocable per-agent grants distinct from owner identity |
2. Why MCP alone is not cross-owner messaging
MCP is a tool-invocation protocol, not a federated delivery network. When a personal coding agent (such as Claude Code, Codex, or OpenClaw) closes its terminal, there is no persistent address or redelivery queue inside MCP itself. ACK Protocol uses MCP strictly at the agent-to-provider boundary: a small, fixed pull-only tool set so any fresh runtime session can fetch signed work and publish responses under a scoped grant.
3. Why ACK Protocol bridges A2A rather than tunneling inside it
A2A is designed for live two-party RPC, not offline-verifiable evidence. As verified in the A2A v1.0.1 specification, messages carry no sender field, idempotency is optional, and the spec explicitly states messages must not be treated as a reliable delivery mechanism for critical information. ACK Protocol keeps record UUIDs and work states mechanically mappable to A2A so an organization will be able to expose a planned A2A gateway at its edge without giving up signed storage receipts or offline audit trails.
4. How all three compose in one workflow
- Step 1 (Sender runtime → Sender provider): Alice’s agent calls an MCP tool to sign and submit a work handoff under her scoped grant.
- Step 2 (Sender provider → Receiver provider): Alice’s provider POSTs the compact ES256 JWS envelope to Bob’s provider over ACK Protocol. Bob’s provider verifies the signature, stores it durably with zero LLM calls, and returns a signed storage receipt.
- Step 3 (Receiver runtime ← Receiver provider): Hours later, Bob’s fresh agent session pulls verified items over MCP, completes the work, and signs a completion event that any third party can verify offline.
- Optional edge (Public doorway, planned): External A2A or email callers will be able to enter through an organization’s gateway extension without changing the core wire.