Testnizer
AI, MCP & streaming

Test the servers your agents talk to — then mock them.

An MCP client that shows every frame, a security scan for MCP servers, an AI chat that can use their tools, and an SSE client — next to mock SSE, LLM and MCP servers that run on your machine.

MCP client

Everything a Model Context Protocol server exposes, in one tab.

Connect over Streamable HTTP, legacy SSE or stdio (Testnizer starts the local server for you). Testnizer negotiates protocol 2024-10-07 through 2026-07-28 — leave it on Auto, force Legacy, or pin a version per tab — and when a server asks for more input mid-call, it shows a form for it.

  • Tools with a form built from each tool’s JSON Schema (and a JSON toggle), Resources with templates, Prompts and Notifications
  • A raw JSON-RPC frame view of everything on the wire, the handshake included, plus the negotiated version and server instructions
  • No Auth, Basic, Bearer, API Key or OAuth 2.1 — with a step-by-step OAuth 2.1 debugger: metadata, dynamic client registration, PKCE, sign-in, token exchange
  • Custom HTTP headers and stdio environment variables, with {{variables}}; paste or export the config for VS Code, Cursor and Claude Desktop
  • Save with Ctrl+S, add Tests and Scripts (pm.test, pm.expect, pm.mcp), and run MCP requests in the Runner, a Test Suite or a schedule

This connects to the MCP server you point it at. OAuth tokens, the PKCE verifier and the client secret stay in the app’s main process and are never saved. A run only starts a stdio server whose exact command you have connected to once on this computer.

Read the docs

MCP security scan

A graded A–F report on an MCP server, with the evidence for every finding.

Point the scan at an HTTP or SSE MCP server and it checks how the server is reached, how it authenticates, how it behaves at the protocol level and what its tool descriptions say. Each category gets its own score, each finding carries its evidence, and the whole report exports as one self-contained HTML file.

  • Transport: HTTPS and TLS, downgrade, cross-origin redirects
  • Authentication: unauthenticated initialize, WWW-Authenticate, protected-resource and authorization-server metadata, tokens in the query string
  • Prompt-injection heuristics in tool descriptions — instruction overrides, hidden Unicode, embedded URLs, tool shadowing, secret-looking schema fields
  • Protocol behaviour, information disclosure, CORS and security headers, an opt-in rate-limit probe, and header and cache-scope checks for 2026-07-28 servers
  • Credentials are scrubbed from every finding before it is shown or exported

The scan sends real requests to the server. Run it only against servers you are authorized to test — the scan screen says the same.

Read the docs

AI Chat

Talk to a model, and let it use your MCP servers — with you approving each call.

Fourteen providers — OpenAI, Anthropic, Google, xAI, DeepSeek, Mistral, Groq, Perplexity, Cerebras, Cohere, Fireworks, DeepInfra, Together and OpenRouter — plus a Custom URL for your own endpoint. Replies stream in, with system prompt, temperature and max tokens under your control.

  • MCP servers as tools: a saved MCP request or an ad-hoc URL or stdio command, with per-server and per-tool switches
  • Every tool call asks first: Allow once, Allow this tool for this conversation, or Deny. "Run tools without asking" lasts for the open tab only and is never saved
  • Each reply shows its status, time to first token, duration and token usage — "not reported" when the provider does not say, never a made-up zero
  • A provider error part-way through a reply keeps the partial text and shows the error under it
  • Conversations are stored locally, masked when written, and never exported or committed; API keys are encrypted per provider with the OS key store

AI Chat is not offline: your messages go directly from your machine to the provider you configure, and tool calls go to the MCP servers you enable. Nothing passes through Testnizer or Apinizer.

Read the docs

Server-Sent Events client

Open a long-lived event stream and watch it arrive.

Connect to an SSE endpoint with your own headers, resume from a Last-Event-ID, filter the log by event type and read JSON payloads pretty-printed. When the server ends the stream and refuses the reconnect, the tab says so instead of logging an error.

  • Custom headers and a Last-Event-ID to resume from
  • Event-type filter, with JSON data detected and formatted
  • Cmd/Ctrl+Enter connects, and a custom proxy applies as it does to HTTP
  • Received events live in memory for the session — they are not written to disk

This connects to the endpoint you enter, and nowhere else.

Read the docs

Mock all three

A fake model, fake tools and a fake stream — on your machine.

The clients above need a real server. These don't. Create a mock from the Mocks panel, pick a preset, and point Testnizer's own client at it with Start & open — a chat tab opened from a mock never sends your saved provider key. This is the part of the page that works with the cable pulled.

No provider account, no token bill

Build and test an agent, an SDK integration or a chat UI against a model that answers instantly and costs nothing per call.

Failures on demand

A 429 on the third request, an error after the fifth chunk, a tool that times out every other call — the cases real providers never produce when you need them.

Deterministic and offline

Mocks bind to 127.0.0.1 by default and need no network, so the same run gives the same answer on a laptop, in a lab or on an air-gapped machine.

Mock SSE server

An event stream that behaves exactly as you script it.

  • Events you define, with {{stream.seq}}, {{stream.iteration}} and {{stream.now}} per frame
  • Interval, heartbeat, loop, keep-open and Last-Event-ID resume
  • Fail on purpose: an event: error after N events
  • Presets: Ticker, Finite replay, Notifications feed, Slow / heartbeat
Read the docs

Mock AI (LLM) server

The OpenAI and Anthropic APIs, answered locally.

  • OpenAI chat completions (and /v1/models) and the Anthropic messages API — streaming when the client asks, JSON otherwise, token usage on the last chunk
  • Chunk by a fixed size or send whole; end with the stop reason you pick
  • Error injection: 401, 429 and 500 replies, an Anthropic 529 overloaded_error, or an error after N chunks
  • A tool-call round-trip preset that pairs with the Mock MCP Echo server; the editor shows the base URL an SDK should use
Read the docs

Mock MCP server

A local MCP server with tools, resources and prompts you define.

  • Streamable HTTP and legacy SSE; one endpoint serves both 2025 and 2026-07-28 clients
  • Tools with JSON Schema, resources, prompts, and bearer authentication that behaves like a real protected server
  • Latency and error scenarios: JSON-RPC error, isError, timeout, HTTP status, every Nth call — and input requested mid-call
  • Presets: Echo, Auth required, Error server, Complex schemas, Slow; live logs and Open in MCP tab
Read the docs

Put them together and an agent's whole loop runs on one machine: AI Chat talks to the mock LLM, the mock LLM asks for a tool call, and the Mock MCP Echo server answers it. HTTP mocks live in the same panel — see the mock server docs.

What connects, and to where

Offline, except where you point it.

Testnizer has no server of its own to talk to. AI Chat sends your messages to the provider you configure, the MCP client and the security scan connect to the MCP server you enter, and the SSE client opens the stream you ask for. Those connections go straight from your machine to that address — the same rule as every HTTP request you send.

What stays local: conversations, the OAuth state of an MCP session, provider API keys (encrypted with the OS key store), received stream events, and every mock server on this page.