Connect an AI tool
Connect Claude Code · Claude Desktop to AxHub and inspect apps, deploys, and data by talking
Connect an AI tool like Claude Code to AxHub and you can say things like "show me my apps' recent deploy status" to inspect apps and data — and even trigger deploys. You connect once, sign in with your own account, and everything runs within your own permissions.
What connecting gets you
A connected AI tool handles these things for you, in plain conversation.
- Inspect apps and deploys — check deployment history and status, and pinpoint the cause when one fails.
- Read data — read tables and rows from the app-dedicated database, plus connected company data (connectors).
- Search AxHub knowledge — look up SDK usage and company documents (NAS) and answer from them.
You connect to one thing: the AxHub MCP server. Whatever AI tool you use talks to AxHub through this address.
https://mcp.axhub.ai/mcpConnect
One line in the terminal is all it takes.
claude mcp add --transport http --scope user axhub https://mcp.axhub.ai/mcpThe first time a tool is called, a browser opens; sign in, consent, and you're connected. There's no client_id or secret to create — the client handles it all.
If you're using the axhub plugin, onboarding ("set up axhub") walks you through this MCP connection as part of the same flow.
Other MCP clients
It doesn't have to be Claude — any MCP client that supports Streamable HTTP (POST/JSON) and OAuth dynamic client registration (DCR) + PKCE connects the same way. The client configuration needs exactly one value: the endpoint.
| Item | Value |
|---|---|
| Endpoint | https://mcp.axhub.ai/mcp |
| Transport | Streamable HTTP (POST/JSON) |
The rest (authorization server, client registration) is handled automatically via standard discovery — metadata is advertised at https://mcp.axhub.ai/.well-known/oauth-protected-resource/mcp.
What success looks like
Once connected, the tool list shows names like sdk_search · connector_list · deployment_list. No need to dig through the list — you can confirm in natural language.
show me my apps' recent deploy statusThat one sentence calls a deployment-read tool (deployment_list or tenant_recent_deployments), confirming both the connection and your permissions in one go.
What you can do once connected — 20 tools
A tool is a button the AI presses on your behalf. The default surface is 20 read-oriented tools — they look things up and search, but change nothing. The AI picks the right tool for the conversation on its own.
| Area | Tools | What they do |
|---|---|---|
| SDK knowledge | sdk_search · get_recipe · validate_usage | Search SDK usage, get recipes tailored to your app's schema, and statically check SDK/DB code you wrote. Call sdk_search first before writing AxHub SDK code. |
| App data | raw_table_list · raw_row_list · table_list · table_get · row_list | Read the app-dedicated DB's table structure and rows (row reads capped at 200). raw_* serve the current database model; table_* and row_list serve the legacy table model. |
| External connectors | connector_list · connector_resources · connector_query · connector_document_query | Browse granted external DB/API connectors and run read-only queries — SQL allows only a single read statement starting with SELECT or WITH; document stores (NoSQL) allow find/findOne/count/distinct only. |
| Deployment reads | deployment_list · deployment_status · deployment_diagnosis · deployment_logs · tenant_recent_deployments | Deployment history, status, failure diagnosis, runtime logs (secrets masked), and a tenant-wide recent-deployment briefing. |
| Company documents (NAS) | nas_search · nas_get_document · nas_list_sources | Semantic search over company NAS documents, parsed-text retrieval, and the list of indexed sources. Original binaries are never returned — parsed text only. |
Operational tools — 21 behind the gate
When the server's deploy-write gate (AXHUB_MCP_DEPLOY_WRITE_ENABLED) is on, 21 operational tools are added for a total of 41. These require the deploy:write scope.
| Area | Tools |
|---|---|
| App management | app_list · app_get · app_create · app_update · app_suspend · app_resume · app_archive · app_reactivate · app_delete |
| Deploy actions | deployment_trigger · deployment_cancel · deployment_rollback |
| Environment variables | env_var_list · env_var_set · env_var_delete |
| GitHub integration | github_link_start · github_link_poll · github_accounts_list · github_repos_list · git_connection_status · git_connection_set |
Operational tools execute immediately on call — there is no web approval or extra confirmation step. If an agent uses them, have it confirm with the user before calling. Permanent deletion is not available over MCP; it exists only in the web console.
Safety boundaries
- Data reads are read-only — connector queries run read queries only, and table-row reads return at most 200 rows per call. Writes to an app's DB happen from the deployed app itself (
DATABASE_URL), not through MCP. - Scoped grants block free-form SQL — if a grant has a table scope attached, free-form SQL through
connector_queryis rejected; only reads within the designated scope go through. - Secrets stay hidden — environment-variable values are masked as
***, and secret-like patterns in logs are redacted.