Skip to content

Agent Mesh

decouple a2a and mcp

Antonio Ojeaa02161ab864a
diff --git a/README.md b/README.md
index cbcebaa0..a8a03573 100644
--- a/README.md
+++ b/README.md
@@ -73,7 +73,7 @@ Trust flow summary:
- Discovery: agents find peers by capability in an isolated federation namespace.
- Authentication: callee verifies caller identity with a locally trusted vouch.
- Authorization: Biscuit caveats constrain which capability can be executed.
-- Execution: skill runs locally (for example via MCP) and response returns over A2A.
+- Execution: skill runs locally through the agent backend (JSON-RPC) and the response returns over A2A.
## Development
diff --git a/docs/README.md b/docs/README.md
index 4bfd632f..5211f96a 100644
--- a/docs/README.md
+++ b/docs/README.md
@@ -139,7 +139,7 @@ Agent A DHT (Discovery) Agent B
│ ├─ Check vouch
├─ Send A2A request ──────────────────────────────────────────→│ BiscuitGate
│ ├─ Verify skill
- │ ← ── MCP Response ────────────┤ Execute
+ │ ← ── Task Response ────────────┤ Execute
│ │
└─ Return result to CLI │
```
diff --git a/docs/concepts/a2a-protocol.md b/docs/concepts/a2a-protocol.md
index 27058b91..cfc0c7ab 100644
--- a/docs/concepts/a2a-protocol.md
+++ b/docs/concepts/a2a-protocol.md
@@ -6,7 +6,7 @@ The **A2A (Agent-to-Agent)** protocol is how agents call skills offered by other
- **Transport**: libp2p streams (TLS-encrypted)
- **Port/Namespace**: `/sam/a2a/1.0`
-- **Message Format**: JSON-RPC 2.0 + MCP
+- **Message Format**: JSON-RPC 2.0 task envelope
- **Authentication**: Vouch (federation membership)
- **Authorization**: Biscuit token (skill access)
@@ -79,8 +79,8 @@ Client Server
{
"jsonrpc": "2.0", // Protocol version
"id": "string", // Request ID (echo in response)
- "method": "message", // Always "message" (MCP calls skill)
- "params": { // MCP parameters
+ "method": "message", // Always "message" (A2A task request)
+ "params": { // task parameters
"message": "..." // Natural-language prompt
}
}
@@ -158,7 +158,7 @@ If no → Return ErrSkillNotAllowed
### Step 3: Execution
-Server calls local MCP endpoint with the skill request.
+Server calls the local agent backend with the skill request.
```
POST http://127.0.0.1:8080
@@ -196,7 +196,7 @@ Server returns response to client
| `ErrNotInFederation` | Vouch not in federation | Authenticate: `sam-agent identity login --hub <url>` |
| `ErrVouchExpired` | Vouch timestamp > expiry | Re-authenticate: `sam-agent identity login --hub <url>` |
| `ErrSkillNotAllowed` | Skill not in Biscuit caveats | Get new Biscuit with skill, or ask for permission |
-| `ErrMCPFailed` | MCP server error | Check MCP server logs |
+| `ErrBackendFailed` | Local backend error | Check local backend logs |
---
@@ -210,7 +210,7 @@ Server returns response to client
3. Server validates vouch (FederationGate)
4. Client sends biscuit header + skill request
5. Server validates biscuit (BiscuitSkillGate)
-6. Server executes skill (calls local MCP)
+6. Server executes skill (calls local backend)
7. Server sends response
8. Stream closes
```
@@ -266,8 +266,8 @@ func (s *A2AService) handleStream(stream libp2pnet.Stream) error {
return fmt.Errorf("decoding request: %w", err)
}
- // 7. Execute skill (call local MCP)
- resp, err := s.executeMCP(ctx, capability, req.Params)
+ // 7. Execute skill (call local backend)
+ resp, err := s.executeBackend(ctx, capability, req.Params)
if err != nil {
return fmt.Errorf("executing skill: %w", err)
}
@@ -370,7 +370,7 @@ Typical A2A call latency:
- DHT discovery: 500ms - 5s (cached after first call)
- Stream setup: 50-200ms
- RPC roundtrip: 100-500ms
-- MCP execution: 1-10s (depends on implementation)
+- Backend execution: 1-10s (depends on implementation)
- **Total**: 2-15 seconds
### Firewalls