SeleMCP

MCP setup for Claude and Codex

SeleAI quickstart · MCP setup for Claude and Codex

There is one connection: https://seleid.dev/mcp. Without a token it serves public information to anyone. With a token you create after approving a SeleID consent screen, it also lets an AI assistant act on your SeleTech account on your behalf. Existing public connections keep working unchanged.

Tools

Tool	Permission (scope)	Needs your approval each time
`docs_search`, `platform_status`, `product_info`, public catalogue reads	none, no sign-in	no
`projects_list`	`projects:read`	no
`project_create`	`projects:write`	no
`project_delete`	`projects:delete`	yes
`storage_list`, `storage_put`	`storage:read`, `storage:write`	no
`storage_delete`	`storage:delete`	yes
`db_query` (one SELECT)	`db:read`	no
`db_execute`	`db:write`	only for deletes, drops, alters and multi-statement scripts
`usage_get`	`usage:read`	no
`deploy`	`deploy:write`	only for production
`keys_list` (never shows secrets)	`keys:read`	no
`key_revoke`	`keys:delete`	yes
`vista_my_analytics`, `vista_my_content`, `vista_my_earnings`	`vista:analytics:read`, `vista:content:read`, `vista:earnings:read`	no (read-only)
`verse_my_analytics`, `verse_my_content`, `verse_my_earnings`	`verse:analytics:read`, `verse:content:read`, `verse:earnings:read`	no (read-only)

The creator tools return only your own data. They take no user id: who you are comes from your approved sign-in, never from the assistant. Permission for SeleVista does not give access to SeleVerse, and the other way round.

How access is limited

You approve first. A token is created only after you approve a SeleID consent screen that lists the exact permissions and a spending limit.
Short-lived. Tokens last 15 minutes by default and at most 1 hour. Revoking the consent stops every token at once.
Scoped. The assistant only sees the tools you allowed.
You confirm destructive steps. Those actions do not run when the assistant asks. You get an approval request in the SeleID console that names the exact action; it works once, for that exact action, for 5 minutes. The assistant cannot approve for itself.
Audit log. Every action the assistant tries, including refused ones, is recorded with time, tool and a fingerprint of the arguments. File contents and SQL are never stored in it.
Spending and rate limits. Each consent has a spending limit and each token a request rate limit.
Untrusted output. Anything read from your projects or posts is returned marked as untrusted data, cleaned of hidden characters and size-limited, so text hidden in your data cannot instruct the assistant.

Claude

Create a connection from the SeleID console (Developer Console, AI assistants), approve the permissions, and copy the short-lived token into an environment variable. Then add the server to your project's .mcp.json:

Shell

export SELEMCP_TOKEN="paste-token-here"

.mcp.json

{
  "mcpServers": {
    "selemcp": {
      "type": "http",
      "url": "https://seleid.dev/mcp",
      "headers": { "Authorization": "Bearer ${SELEMCP_TOKEN}" }
    }
  }
}

Or from the command line:

Shell

claude mcp add --transport http selemcp https://seleid.dev/mcp --header "Authorization: Bearer $SELEMCP_TOKEN"

For public tools only, add the same address without the header.

Codex

~/.codex/config.toml

[mcp_servers.selemcp]
url = "https://seleid.dev/mcp"
bearer_token_env_var = "SELEMCP_TOKEN"

Try it

Using selemcp, list my projects and show this month's usage.

Ask for something destructive and the assistant will report confirmation_required and wait for you to approve it in the console.

Good habits

Grant the fewest permissions the task needs; grant delete permissions only for the session that needs them.
Set a low spending limit.
Never paste tokens into prompts, files or chat. Treat them like passwords.
Review the audit log after long sessions.