A hosted MCP client for every model

1. On a paid plan, open Settings → MCP and add a server: a name, its https:// address and its transport, Streamable HTTP or SSE.

2. If the server needs a token, add it as a header, such as Authorization with a bearer token, or whatever header the server asks for: up to 8 of them.

3. Press Test. llmwise connects, lists the server's tools, and shows any it would leave out, and why.

4. Turn the server on. In any chat, the model can now call its tools. By default, each call shows an approval card first.

5. Keep up to 20 servers, with up to 5 turned on at once.

• Header values are encrypted before they're stored (AES-256-GCM), tied to your account and that one server, so a copied value can't be opened anywhere else.

• Settings never returns saved header values: it only shows that a server has headers.

• llmwise tries to redact saved header values from tool text and errors. Redaction can miss content, so connect only servers you trust with your data.

• By default, every MCP tool call waits for your approval, showing the tool and what the model wants to send it. For a tool you trust, choose Always allow on its approval card or in Settings → MCP, one tool at a time; from then on that tool runs without asking, until you turn it off.

• At most 10 MCP calls in one reply, each with 60 seconds to finish. A slow or failing server becomes an error the model reads, never a broken chat.

• Up to 64 tools are offered in a reply, and each result is cut at 100,000 characters.

• llmwise connects only to public https:// addresses: no private networks, no local addresses, and no redirects. Connecting and listing tools get 10 seconds.

A server's tools work with every model in llmwise: Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5, Claude Haiku 4.5, GPT-6 Astra, GPT-6 Sol, GPT-6 Luna, Gemini 3.1 Pro (preview), Gemini 3.8 Flash, DeepSeek V4.1 Flash, DeepSeek V4 Pro, Grok 4.7, Kimi K3, GLM 5.3, and GLM 5.3 Flash. Start a task on a fast model and switch to a bigger one mid-chat; it sees the whole conversation, tool results included, and can call the same tools. Tool calls inside a reply don't change the message's price.

Connectors bring Gmail, Google Drive, Notion, Slack, GitHub and more into the chat: 100+ apps connect in one click · 1,500+ in Composio’s catalog. Reading needs no approval. Writes ask first by default; you can allow individual actions you trust. Sending, deleting, payments and sharing always ask.

Sign-in runs through Composio (SOC 2 Type II, ISO 27001), which holds the connections and their sign-in tokens, encrypted; llmwise never sees your app passwords. Some sign-in screens show Composio's name rather than llmwise's. Like MCP, connectors are on paid plans.

• Run a local server on your computer: llmwise connects to remote servers only. For local servers, a desktop MCP client is the right tool.

• Sign in to a server with OAuth. A server that only accepts OAuth won't connect; llmwise tells you “The server asked to authorize”.

• Read a server's resources or prompts: llmwise uses its tools.

• Run on the free trial: MCP needs a paid plan.

What a server you add receives, and how llmwise keeps its secrets, is in Privacy, security and your data.