NEWMilestone 4: WASM plugins, split-key TLS & telemetry

Memory-safe.Secure by default.

MiraServe is a memory-safe, secure web server and edge gateway written in Rust, with post-quantum HTTPS and AI-native protection built in. Every request is checked at the edge before it reaches your site.

Alpha Milestones 0–4 are complete. Milestone 5 (packaging, an external security audit and v1.0) is in progress, so configuration and APIs may still change. See the roadmap
Why MiraServe

Traditional web hosting vs MiraServe

Most sites still run on servers designed decades ago, with security bolted on afterwards. MiraServe starts from safe defaults.

Traditional hosting
MiraServe

Security

Code safety
Written in C or C++, where buffer overflows and use-after-free bugs remain possible.
Pure Rust with unsafe code forbidden, so whole classes of memory bugs are ruled out.
HTTPS
TLS tuned by hand; older protocol versions are often left switched on.
TLS 1.3 only by default, with X25519MLKEM768 post-quantum key exchange first.
Private keys
Loaded into the same processes that face the internet.
Can be held by a separate signing daemon, miraserve-keyd, outside worker memory.
Extensions
Native modules run with the full privileges of the server.
WebAssembly plugins with CPU, memory and time limits and no ambient access.
Isolation
Workers can usually read anything the server user can.
Landlock and seccomp-bpf sandboxing on Linux; files confined beneath the web root.

Operations

Certificates
Separate tools and scheduled jobs to renew certificates.
Built-in ACME v2 with renewal information (ARI) and short-lived certificates.
Configuration
Long config files where security mistakes pass a syntax check.
Typed KDL config with a strict security audit before it runs, and atomic hot reload.
Deploys
Files copied over the live site, with no clean way back.
Signed, atomic releases with a zero-downtime swap and instant rollback.
Migration
Moving servers means rewriting configs by hand.
miraserve import drafts a config from Nginx, Caddy, Apache or OpenLiteSpeed for you to review.
Monitoring
Plain log files, where secrets can end up in logs.
Prometheus and OTLP export, with access and audit logs that redact secrets.

AI era

AI crawlers
Identified by user-agent strings, which anyone can fake.
Verified cryptographically with RFC 9421 Web Bot Auth signatures.
Unknown bots
Handled with IP lists and robots.txt, if at all.
Unverified AI traffic is throttled, challenged or blocked by policy.
AI tools (MCP)
Exposed directly or behind a generic proxy.
MCP security gateway with OAuth metadata, per-tool access control and redacted audit logs.
AI-written sites
Generated code goes live as-is.
Pre-flight checks for leaked secrets, inline script and CSP before publishing.
LLM endpoints
Proxy buffering can stall streamed responses.
Zero-buffer SSE streaming, plus llms.txt and Markdown content negotiation.
Built for the AI edge

AI now reads, writes and calls your site. Your server should know who is who.

Crawlers, AI agents and MCP tools are a growing share of web traffic, and anyone can fake a user-agent. MiraServe checks identity cryptographically and applies policy at the edge.

Person in a browser

GET /pricing HTTP/3
Host: shop.example.com
  1. TLS 1.3 handshake with post-quantum key exchange
  2. Request normalised and checked for smuggling
  3. Within rate limits for this client
  4. Served from the confined web root with strict security headers
ResultServedmiraserve-http · miraserve-static

Verified AI agent

GET /docs/install HTTP/2
Accept: text/markdown
Signature-Agent: "https://agent.example"
Signature-Input: sig1=("@authority" "signature-agent")…
  1. RFC 9421 Ed25519 signature verified (Web Bot Auth)
  2. Agent allowed by this site’s AI policy
  3. Accept: text/markdown negotiated, clean Markdown returned
  4. Request logged with secrets redacted
ResultServed as Markdownmiraserve-ai

Bot faking an AI crawler

GET /wp-login.php HTTP/1.1
User-Agent: ExampleAIBot/1.0
  1. User-agent claims to be an AI crawler
  2. No valid Web Bot Auth signature
  3. Policy for unverified AI traffic: throttle and challenge
  4. Event recorded in the audit log
ResultThrottled & challengedmiraserve-ai · rate limiter

Request smuggling attempt

POST / HTTP/1.1
Content-Length: 44
Transfer-Encoding: chunked
  1. Conflicting Content-Length and Transfer-Encoding detected
  2. Rejected by the RFC 9110/9112 normaliser before routing
  3. Connection closed; upstream never sees the request
  4. Event recorded in the audit log
ResultBlocked · 400miraserve-security

AI tool call over MCP

POST /mcp HTTP/2
{"method":"tools/call","params":{"name":"delete_user"}}
  1. OAuth token validated (RFC 9728 metadata)
  2. JSON size and nesting within limits
  3. This tool is on the site’s deny list
  4. Denied call written to a secret-redacted audit log
ResultDenied by policyMCP security gateway

Web Bot Auth

RFC 9421 signatures prove an AI agent is who it says it is.

MCP security gateway

OAuth metadata, per-tool access control and redacted audit logs.

Safe AI publishing

AI-written sites are scanned for secrets, inline script and CSP issues, then swapped in atomically.

AI-friendly content

llms.txt, Markdown negotiation and zero-buffer streaming for LLM endpoints.

Protection

What MiraServe protects you from

Each threat below is handled in the server itself, with no extra add-ons to install.

Threat

Memory corruption bugs

Buffer overflows and use-after-free are ruled out: pure Rust with unsafe code forbidden.

Threat

Request smuggling

An RFC 9110/9112 normaliser rejects ambiguous message framing before routing.

Threat

Floods & Slowloris

HTTP/2 and HTTP/3 flood guards, slow-body limits and a bounded-memory GCRA rate limiter.

Threat

Path traversal

Files are opened beneath the web root only, enforced by the kernel (openat2 RESOLVE_BENEATH).

Threat

Stolen TLS keys

In split-daemon mode private keys live in the separate miraserve-keyd daemon, never in worker memory.

Threat

Harvest now, decrypt later

Post-quantum hybrid key exchange (X25519MLKEM768) protects traffic recorded today.

Threat

Risky configuration

miraserve check --strict audits a config for security mistakes before the server will run it.

Threat

Malicious plugins

WebAssembly plugins run with CPU, memory and time limits and no file, environment or network access.

Threat

Compromised PHP apps

PHP runs outside the server process over FastCGI, with HTTPoxy protection and a hardened WordPress preset.

Threat

Leaked secrets in deploys

Published sites are checked for secrets, inline script and CSP problems before they go live.

Threat

Fake AI crawlers

Unsigned bots claiming to be AI agents are throttled, challenged or blocked by policy.

Threat

AI tool abuse

MCP tool calls are checked against allow and deny lists, with size limits and audit logs that redact secrets.

  • TLS 1.3
  • HTTP/3 · QUIC
  • HTTP/2
  • ACME v2 + ARI
  • wasi:http
  • FastCGI · PHP-FPM
  • Prometheus · OTLP
Get started

From Nginx to MiraServe in three commands

Import your existing config as a draft, audit it, and run it with TLS 1.3, post-quantum key exchange and HTTP/3 switched on.

In private testing. Source release coming soon.

MiraServe is being tested privately before anyone else runs it. The source code will be published after testing on real sites and an independent security review.