MCP · AI access

Let any AI model work inside your Conflux — safely.

Conflux Desktop embeds a local MCP server. Connect Claude Code or any MCP client with one command and let the agent read, build, run and test your APIs — while every change stays a local draft until you publish.

Connect Claude Code Download Conflux

What it is

The Model Context Protocol (MCP) is an open standard that lets AI agents like Claude Code call tools exposed by an application. Conflux Desktop runs a local MCP server on 127.0.0.1:4630, speaking Streamable HTTP with a Bearer token and an Origin allowlist. It's off by default, and it stops the moment you close the app window. Every tool call is answered by the same code paths the Conflux UI itself uses, so an agent's edits show up in your sidebar immediately, with the same dirty-tab badges as a change you made by hand.

How it works

Three steps from a fresh install to an agent reading your first collection.

Enable AI access in the app

Open the app-bar menu in Conflux Desktop, choose "AI access (MCP)…", and toggle "Enable AI access". The dialog shows the running URL, a masked Bearer token, and an "Allow AI to delete" toggle that is off by default.

Connect Claude Code

Copy the ready-to-run command from the dialog, or run it yourself with your own token, then confirm the connection.

claude mcp add --transport http --scope user conflux http://127.0.0.1:4630/mcp \
  --header "Authorization: Bearer <token>"

$ claude mcp list      # should show conflux: Connected

Ask the agent to work

Inside a Claude Code session, ask it to read, build, run or test your APIs in plain language — it calls the Conflux MCP tools directly.

"List every request under my Auth folder and add a status-code test to each one."

Tools, resources & prompts

31 tools grouped by what they're allowed to do, plus 2 resources and 2 prompts that give an agent extra context without a tool call.

Read (11)

list_workspaces — Lists every workspace visible to the signed-in account, and which one is active.
get_workspace_tree — Returns the full collection/folder/request tree for a workspace.
get_request — Reads a saved request — headers, params, body, auth, scripts. Secret auth values are masked.
get_folder — Reads a folder or collection node — metadata, default auth, variables, and children.
list_environments — Lists environments for a workspace. Secret variable values are always masked.
get_environment — Reads one environment's variables. Secret values are always masked.
get_history — Lists recent sent requests for a workspace. No response bodies included.
list_snapshots — Lists frozen response snapshots saved under a request.
get_snapshot — Reads one frozen snapshot's request + response, truncating a large body.
generate_code — Generates a curl command or fetch() snippet for a saved request, variables substituted.
get_script_api_reference — Returns the reference docs for Conflux's pre-request/post-response/global scripting API.

Write (11)

create_collection — Creates a new root collection in a workspace.
create_folder — Creates a new folder inside an existing folder or collection.
update_folder — Renames a folder/collection and/or updates its description.
create_request — Creates a new HTTP request inside a folder/collection.
update_request — Updates an existing request; omitted fields are left as-is.
move_node — Moves/reorders a folder or request under a new parent.
create_environment — Creates a new environment in a workspace.
update_environment_variables — Merges given variables into an environment (adds/overwrites by key).
set_active_environment — Sets the active environment for the current workspace.
import_collection — Imports a Postman v2.1 export or a Conflux .conflux.json export into a workspace.
export_collection — Exports a subtree or the whole active workspace as Postman v2.1 or Conflux JSON.

Delete — gated (4)

delete_request — Deletes a single request.
delete_folder — Deletes a folder or collection and everything under it.
delete_environment — Deletes an environment.
delete_snapshot — Deletes a frozen response snapshot.

Execute (2)

execute_request — Sends a saved request (scripts + tests run) and returns the response and test results.
execute_adhoc_request — Sends a one-off HTTP request without creating a saved request.

Run & snapshot (3)

run_collection — Runs every request under a folder/collection headlessly, in tree order.
create_snapshot — Runs a saved request headlessly and freezes the request + response as an immutable snapshot.
compare_with_snapshot — Re-runs a snapshot's exact request and diffs the fresh response against the saved one.

31 tools total.

Resources (2)

conflux://script-api — Markdown reference for the current pre-request/post-response/global scripting API.
conflux://workspace/{id}/tree — The full collection/folder/request tree for a workspace, as JSON.

Prompts (2)

write_tests_for_request — Drafts conflux.test()/expect() assertions for a saved request's post-response script.
debug_failed_request — Scaffolds a debugging pass over a request's recent failures using its history.

Security & control

Loopback bind, a Bearer token, and an Origin allowlist mean only a process on this machine, holding the token, can reach the server — and even then, you decide what it can touch.

Loopback-only, token-gated

Bound to 127.0.0.1, guarded by a Bearer token and an Origin allowlist. No interactive OAuth or credential prompt an agent could trigger.

Drafts, not publishes

Every write goes through the same local-draft path a human edit uses. Nothing reaches Conflux.Server until you press Ctrl+S.

Deletes are opt-in

Delete tools refuse to run until you flip "Allow AI to delete" — off by default, and even then it only affects local drafts.

Secrets always masked

Vault values, environment secrets, and Authorization/Cookie headers come back as *** in every tool result, with no exceptions.

Connect other clients

Claude Code is first-class, but the server speaks plain Streamable HTTP with a Bearer header — it works with any MCP client that supports that transport.

Project-scoped setup: drop a .mcp.json next to your code, expanding the token from your shell environment rather than committing it.

{
  "mcpServers": {
    "conflux": {
      "type": "http",
      "url": "http://127.0.0.1:4630/mcp",
      "headers": {
        "Authorization": "Bearer ${CONFLUX_MCP_TOKEN}"
      }
    }
  }
}

Claude Desktop is stdio-only, so it needs mcp-remote to bridge to the local HTTP server:

{
  "mcpServers": {
    "conflux": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "http://127.0.0.1:4630/mcp",
        "--header",
        "Authorization: Bearer ${CONFLUX_MCP_TOKEN}"
      ],
      "env": {
        "CONFLUX_MCP_TOKEN": "<paste the token from the AI access dialog>"
      }
    }
  }
}

Any other MCP client that speaks Streamable HTTP with a Bearer header can connect the same way — point it at http://127.0.0.1:4630/mcp with an Authorization: Bearer <token> header.

Frequently asked questions

Is the MCP server on by default?

No. It's off until you toggle "Enable AI access" in the app-bar menu, and it stops the moment you close the app window.

Can an agent publish changes to Conflux.Server on its own?

No. Every write an agent makes is a local draft, exactly like typing in the UI. Nothing reaches Conflux.Server until you press Ctrl+S yourself.

Can an agent delete things?

Only if you turn on "Allow AI to delete" — off by default. Delete tools refuse to run until then, and even after that a delete only affects local drafts until you publish.

Will an agent ever see my real API keys or tokens?

No. Vault values, environment secrets, and Authorization/Cookie headers are always masked as *** in tool output, even in a resolved-request preview.

Does this work with clients other than Claude Code?

Yes. The server speaks plain Streamable HTTP with a Bearer header, so any MCP client that supports that transport can connect — Claude Desktop needs the mcp-remote bridge since it's stdio-only.

What happens if I regenerate the token?

The old token stops working immediately. Re-run claude mcp add (or click "Add to Claude Code" again) with the new token to reconnect.

Does the agent need the app to be running?

Yes. The MCP server lives inside the Conflux Desktop process — if the app is closed, claude mcp list shows conflux as disconnected.

Try it with your own APIs

Download Conflux, enable AI access, and connect Claude Code in under a minute.

Download Conflux