ProtoboxProtobox
Now in public beta

Deploy your secure MCP server for your team and every AI agent

Let your team and customers use your tools, prompts, and knowledge across Claude, Cursor, and ChatGPT — with controlled access, all managed from one place.

Free tier — 1 server, 1,000 requests/month. No credit card.

Speaks every MCP client

ClaudeOpenAIMicrosoft

Wraps every API you ship

HubSpotSalesforceSlackGmailNotionShopifyCalendarAnalytics
The problem

MCP is the spec. The server is still the part nobody warned you about.

010%

Of production primitives ship in the spec. Transport, auth, rate limits, observability, secrets, billing — all yours to build before a single tool is callable.

025+

MCP clients, each negotiating differently. SSE, Streamable HTTP, JSON-RPC quirks — every one is its own late night with the dev tools open.

03Weeks

Of plumbing before your actual idea ships. That's a quarter spent on infrastructure that was never your product.

MCP turns every backend into an AI tool — that's the promise. You shouldn't have to rebuild transport, auth, and metering to collect on it. That part should already be solved.

From the Protobox team

What is MCP

Model Context Protocol, in plain English

The standard plug between LLMs and your tools. Open spec, every client, no glue code.

The spec

JSON-RPC. Open. Transport-agnostic. Anthropic shipped it; current revision 2025-11-25.

The clients

Claude, ChatGPT, Cursor, Windsurf, Zed — plus every serious agent framework.

What we add

Hosting, OAuth, rate limits, audit logs. Stable URL. Every MCP client connects.

Meet your guide

We've shipped this server in production. Now you can skip straight to it.
Define, deploy, observe — without the infrastructure tax.

Define

Tools in TypeScript, Python, or YAML. Or import OpenAPI. Schemas, handlers, secrets — one config.

Learn more
Deploy

One command. SSL, subdomain, encrypted secrets, metering. URL ready in seconds.

Learn more
Observe

Live tool-call logs. Latency per handler. Error breakdown. Per-client usage.

Learn more
Proven in production

The same runtime already powers our own products.
Now it runs anything an AI client should call.

This isn't a hosting wrapper we built for a launch. It's the exact runtime behind Chanl, Audial, and Sixfactors — battle-tested in production, handed to you as building blocks.

Knowledge Bases

RAG-as-MCP. Point at docs, ship a search tool.

S3, Notion, GDrive, Postgres, files. Embeddings handled. Clients get search and get.

  • S3, Notion, GDrive, Postgres, files
  • Auto-chunked, auto-embedded
  • Citations + relevance scores
Agents as subagents

Multi-step LLM loops as one MCP tool.

Plan, retry, branch, call sub-tools. Workflow on the server, not in the prompt.

  • ReAct loops, branching, retries
  • Calls other Protobox tools
  • Streaming progress + final result
Memory

Persistent state across MCP calls.

Per-user, per-thread. Short-term promotes to long-term. The layer we run in production.

  • Per-user, per-thread scopes
  • Short-term → long-term promotion
  • Read + write tools, no glue code
Define

Tools, resources, prompts — typed and tested

Everything the MCP spec exposes, in idiomatic code

Define tools as typed functions. Mount resources from your database, S3, or anywhere. Publish prompts your client can call by name. Bring your existing OpenAPI spec and we generate the surface for you. Bring your own secrets — they live in a vault, never in your repo.

Everything an MCP server actually needs

Code, HTTP, workflows, toolsets, OpenAPI imports, real security primitives, and audit-grade observability. Skip the months of platform work.

Code tools

Define tools as typed TypeScript or Python handlers. Input schemas, async handlers, environment-aware secrets. The same code path you would have written by hand — minus the protocol plumbing.

HTTP tools

Wrap any REST endpoint as an MCP tool. Pass headers, transform requests, cache responses. Point Protobox at your existing API and ship in minutes.

Workflow tools

Compose multi-step tools with branching, retries, and downstream tool calls. Useful for orchestrations that would otherwise live in your agent prompt.

Toolsets

Group related tools, version them together, and share across servers. Promote a toolset from staging to prod with a single command.

OpenAPI → MCP

Drop in an OpenAPI 3 spec and Protobox generates the MCP tool surface for every endpoint. Customize, hide, rename, or extend any generated tool.

Security baked in

Per-server API keys, per-tool scopes, bearer-token passthrough for upstream calls, secrets encrypted at rest. OAuth for the AI client connection on the roadmap.

Audit logs + observability

Every tool call captured: request, response, latency, auth context, error trace. Replay any invocation. Drop-in OpenTelemetry export for your existing stack.

Rate limits + quotas

Per-key, per-tool, per-IP rate limits. Hard quotas that return graceful 429s with retry hints. Burst windows configurable per tier.

See your MCP server in action

Every tool call captured: request, response, latency, auth context. Replay any invocation. Drop into observability mode the moment something looks off.

mcp-stream
[09:41:12]connectionTransport: SSE | Status: Connected
[09:41:13]tool_list12 tools registered
[09:41:14]tool_callmemory.search({ query: 'billing...' })
[09:41:15]tool_result{ matches: 3, relevance: 0.94 }
[09:41:16]tool_callknowledge.query({ topic: 'refund...' })
[09:41:17]tool_result{ documents: 2, confidence: 0.91 }
[09:41:18]heartbeatlatency: 12ms
>
Observe

Every tool call is a signal

Logs, latency, errors, costs — per tool, per client

Live structured logs for every MCP request. P50/P95/P99 latency per handler. Error breakdown by class. Per-client usage and per-tool cost. The observability layer your homegrown MCP server was always going to need.

Test

Catch regressions before any AI client does

Synthetic clients, scorecards, and burst tests

Run a 100-call burst against staging before you flip prod. Synthetic MCP clients hit your server with realistic tool sequences. Scorecards check that every response matches your contract. CI fails if a regression slips in.

Platform capabilities

Everything you need to ship
production-grade MCP servers.

Define

Tool definitions in code

Typed handlers in TypeScript or Python. Input schemas inferred from your types. No JSON-by-hand.

Define

OpenAPI import

Point at a spec, get a working MCP server. Every operation becomes a tool with auth and pagination.

Define

Auth & rate limiting

Per-server API keys. Bearer-token passthrough. Configurable rate limits per tool, per client.

Define

Streamable HTTP & SSE

Both transports out of the box. Long-running tool calls stream tokens, results, and progress events.

Test

Synthetic load

Burst-test your server before AI clients do. 100-call sequences with realistic tool patterns.

Test

Response scorecards

Latency budgets, schema conformance, error rates. Auto-grade every response in CI.

Test

Version diffing

Compare two server versions side-by-side. Catch breaking changes before they reach production.

Test

Client probes

Claude-style, Cursor-style, GPT-loop probes. Each one stresses your server differently.

Test

Tool playground

Iterate on tool descriptions and input schemas with instant feedback. Test before deploying.

Observe

Tool-call traces

Every request, response, latency, and auth context. Replay any invocation.

Observe

Real-time analytics

Tool calls per minute, latency distributions, error rates. Per-client, per-tool, per-version.

Observe

SLO alerts

Define a P95 budget per tool. Get paged when a handler regresses. Auto-rollback on runaway errors.

Observe

Webhook fan-out

Trigger downstream actions on tool-call events. Forward to Slack, PagerDuty, your own pipeline.

Observe

Cost & usage tracking

Per-client request volume, per-tool compute cost. Set budgets and get alerted on spikes.

The plan

Three steps from config to connected client

No deploy pipelines to wire up. No transport code to maintain. Just tools and a URL — the same path you'd take if you'd already built the platform.

01

Define your tools

Write a typed tool in TypeScript or Python, or import an OpenAPI spec. Include input schemas, handlers, and any secrets you need.

02

Deploy with one command

Run protobox deploy. We provision SSL, allocate your subdomain, encrypt your secrets, and meter every request.

03

Hand the URL to any MCP client

Paste your your-name.protobox.app/mcp URL into Claude, ChatGPT, Cursor, or anything else that speaks MCP. Done.

MCP changelog and dev notes

Short, occasional updates on the MCP spec, new Protobox features, and patterns we've seen in production. No marketing fluff.

Be among the first

Frequently asked questions

Ship the server, or spend the quarter building one

Choose Protobox and your tools are live for Claude, ChatGPT, and Cursor this afternoon — with auth, logs, and metering already handled. Skip it and the plumbing is still waiting for you. Start on the free tier: one server, a thousand requests a month, no credit card.