From 498ce7a51a5785d614bea197fe78893474d7241a Mon Sep 17 00:00:00 2001
From: Shivakumar <kumar@appz.dev>
Date: Wed, 26 Aug 2026 20:52:09 +0530
Subject: [PATCH] mcp --http done: Arc-shared AgentflareMcp across sessions,
 loopback-only default, 244 tests green. Follow-ups: per-process agent
 identity conflates multi-agent handoff; add run_http concurrent test.

Agentflare-Agent: claude-code_2-1-251_agent
Agentflare-Branch: HEAD
Agentflare-Session: 42b8c2f2-a60f-4b46-9fab-fb3d3370d30f
---
 Cargo.toml        |  2 +-
 src/cli/mcp.rs    | 27 +++++++++++++++++++++++++--
 src/mcp_server.rs | 35 +++++++++++++++++++++++++++++++++++
 3 files changed, 61 insertions(+), 3 deletions(-)

diff --git a/Cargo.toml b/Cargo.toml
index a9afc4df..f23dc2a2 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -47,7 +47,7 @@ toml = { version = "0.8" }
 serde_json = "1"
 dirs = "6"
 chrono = "0.4"
-rmcp = { version = "1.8.0", features = ["server", "transport-io"] }
+rmcp = { version = "1.8.0", features = ["server", "transport-io", "transport-streamable-http-server"] }
 tokio = { version = "1", features = ["rt", "rt-multi-thread", "macros", "io-util", "io-std", "sync", "net", "time"] }
 ureq = { version = "2", features = ["json"] }
 sha2 = "0.10"
diff --git a/src/cli/mcp.rs b/src/cli/mcp.rs
index 77a06e97..0020627e 100644
--- a/src/cli/mcp.rs
+++ b/src/cli/mcp.rs
@@ -1,11 +1,34 @@
 use clap::Args;
 
-/// Run the agentflare MCP server over stdio (for editor/agent MCP clients).
+/// Run the agentflare MCP server (stdio by default; `--http` serves many
+/// concurrent agent sessions from one persistent process over Streamable HTTP).
 #[derive(Args)]
-pub struct McpArgs;
+pub struct McpArgs {
+    /// Serve over Streamable HTTP/SSE instead of stdio, so one agentflare
+    /// process backs many concurrent agent sessions instead of one per session.
+    #[arg(long)]
+    pub http: bool,
+    /// TCP port for `--http`. Default 35274 ("FLARE" shifted one up from the
+    /// dashboard's 35273); 0 = auto-assign.
+    #[arg(long, default_value = "35274")]
+    pub port: u16,
+    /// Interface to bind for `--http`. Loopback only for now: rmcp's default
+    /// Host allowlist rejects remote clients, and there is no auth yet.
+    #[arg(long, default_value = "127.0.0.1")]
+    pub host: String,
+}
 
 impl McpArgs {
     pub fn run(self) {
+        if self.http {
+            let runtime = tokio::runtime::Runtime::new()
+                .expect("failed to build tokio runtime for mcp server");
+            if let Err(e) = runtime.block_on(crate::mcp_server::run_http(&self.host, self.port)) {
+                crate::ui::error(&format!("agentflare mcp: {e}"));
+                std::process::exit(1);
+            }
+            return;
+        }
         let runtime = tokio::runtime::Builder::new_current_thread()
             .enable_all()
             .build()
diff --git a/src/mcp_server.rs b/src/mcp_server.rs
index 4de0fbb2..5e598f3b 100644
--- a/src/mcp_server.rs
+++ b/src/mcp_server.rs
@@ -1798,6 +1798,41 @@ pub async fn run() -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
     Ok(())
 }
 
+/// Serve the MCP server over Streamable HTTP/SSE so one persistent agentflare
+/// process backs many concurrent agent sessions, instead of one cold-started
+/// OS process per connecting agent (the stdio default).
+pub async fn run_http(
+    host: &str,
+    port: u16,
+) -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
+    use rmcp::transport::streamable_http_server::{
+        StreamableHttpServerConfig, StreamableHttpService, session::local::LocalSessionManager,
+    };
+
+    // One server instance shared across every session: the point of HTTP mode
+    // is that a single process (and its lazily-opened skills/gateway registries
+    // and backend/store DB connections) backs many concurrent agent sessions,
+    // instead of cold-starting one per connection. The factory clones the Arc,
+    // so each rmcp session routes into the same `AgentflareMcp`.
+    let shared = std::sync::Arc::new(AgentflareMcp::from_env());
+
+    let service: StreamableHttpService<std::sync::Arc<AgentflareMcp>, LocalSessionManager> =
+        StreamableHttpService::new(
+            {
+                let shared = shared.clone();
+                move || Ok(shared.clone())
+            },
+            std::sync::Arc::new(LocalSessionManager::default()),
+            StreamableHttpServerConfig::default(),
+        );
+    let app = axum::Router::new().nest_service("/mcp", service);
+    let listener = tokio::net::TcpListener::bind((host, port)).await?;
+    let addr = listener.local_addr()?;
+    eprintln!("agentflare MCP listening on http://{addr}/mcp (streamable HTTP)");
+    axum::serve(listener, app).await?;
+    Ok(())
+}
+
 #[cfg(test)]
 mod optimize_instructions_tests {
     use super::build_optimize_instructions;