Connect your AI to SubReply

The shortest path, with no key to copy anywhere. The connector sends you to SubReply, you sign in with your usual account and it is plugged in.

The approval lasts 30 days. After that the client reopens the same screen, and one click is enough to carry on.

The key travels in the Authorization header. Every request is re-authenticated, so revoking a key cuts it off that very moment rather than when some session expires. It is the same sr_live_… key as the REST v1 API, generated on your API keys page.

Claude Code and Claude Desktop

claude mcp add --transport http subreply https://subreply.io/api/mcp \
  --header "Authorization: Bearer sr_live_YOUR_API_KEY"

Or straight into the MCP configuration file, if you would rather keep it in version control:

{
  "mcpServers": {
    "subreply": {
      "type": "http",
      "url": "https://subreply.io/api/mcp",
      "headers": {
        "Authorization": "Bearer sr_live_YOUR_API_KEY"
      }
    }
  }
}

n8n

Add an MCP Client node and fill in four fields:

• Endpoint — https://subreply.io/api/mcp
• Server Transport — HTTP Streamable
• Authentication — Header Auth
• Header Auth credential — name Authorization, value Bearer sr_live_…

The node then lists all six tools. Mind the timeout: publish_reddit_comment can take two minutes, well beyond the n8n default.

Checking that it answers

When a client refuses to connect, this call settles the question: it depends on no MCP client and spends no credits.

curl -s -X POST https://subreply.io/api/mcp \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -H "Authorization: Bearer sr_live_YOUR_API_KEY" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list","params":{}}'

Six tools listed means the server is fine and the problem sits on the client side. A 401 means the key. A 406 means the Accept header: the transport requires both types, even though it answers with plain JSON. The ping tool does the same job once your client is connected: it answers pong and proves your key works at the same time.

Same economics as the REST API: searching is free, only production is billed. A scan that finds nothing therefore never costs anything.

Charges are idempotent per post: if the network drops and your agent replays generate_reddit_comment on the same Reddit URL, the second call returns credits_used: 0.

Quotas are counted per account and per tool: exhausting generate_reddit_commentthrottles neither your other tools nor anyone else's account. ping is deliberately unlimited — it is the probe, and it has to stay reachable exactly when you are hammering it because you suspect an outage.

Which is why it is off by default. To turn it on, open your API keys page and enable publishing on the key in question.

The permission lives on the key, not on the account, and that is deliberate: a key dedicated to an n8n publishing workflow can have it while a key pasted into a conversational assistant does not. After a leak, turning publishing off on that one key breaks none of your other integrations.

• 5 calls per minute at most, far below what the infrastructure would take. A burst of publications does not just cost a balance: it costs a shadowban.
• The 50 credits are charged afterwards, once publication is confirmed. A failure bills nothing.
• This tool is not idempotent: two calls on the same URL leave two comments. The second one is free, but it is very much there.
• Allow up to two minutes per call — a real browser opens the post and types the comment. Set your client timeout accordingly.

A failing tool never raises a JSON-RPC exception: it returns a result flagged isError: true, with a sentence the agent can read. Four cases exist only on the MCP side.

The rate limit message carries the retry delay, so your agent knows when to pick back up instead of looping:

Rate limit exceeded for generate_reddit_comment:
20 calls per minute maximum. Retry in 34 seconds.
No credits were charged.

The refusal lands before any business logic: nothing is called, nothing is charged. An agent that burns through its 20 generations per minute has lost no credits.

HTTP codes and the REST error format are described on the error codes page.