Skip to content

Ask the agent: self-configuration, magic links and admin tools

Every agent can reconfigure itself when asked to in a conversation — on any surface, the simple chat view included — and can hand a person a magic link whenever a browser is needed. AgentFather can do the same for the whole deployment. This is the default way to connect, configure and operate Olano: an agent never asks you to open a terminal, and there is no command line to learn.

Ask a bot to:

  • "Connect yourself to Notion" — every connection starts with the finder: the bot lists the supported ways to reach the service (a channel, an MCP server from the directory, a built-in toolkit — see MCP servers), adds the one you pick to itself, and hands you the sign-in: a Connect button card in the dashboard chat, a tappable link on messaging channels. You sign in from your own browser — Olano's app, your own OAuth app, or a code you type on your phone; the connection completes on its own and the bot's new tools appear after that. API keys are collected through a secure page and stored in the vault — never pasted into the chat or a config file. Something the bot cannot find in any catalog is external: it will read the project's page and tell you what it is before anything is installed, and it will never sign it in for you.
  • "Get yourself on Telegram" — it sends you one link whose page walks you through the bot token (validated live) or, for WhatsApp and WeChat (Weixin), shows a QR code to scan — in the dashboard chat and on messaging channels alike. The channel and the token land in its config and vault, and it restarts to go live.
  • "Switch to a cheaper model" / "Rewrite your instructions to always answer in Spanish" / "Add a research assistant" — it lists real options, saves the change, and restarts itself when you say so. Prompt edits land in Config → Revisions under the agent's name, ready to roll back.
  • "What could you become?" — it browses the archetype library of ready-made experts and teams and suggests specialists worth adding.
  • "Which skills do you have?" / "Stop using the spreadsheet skill" — it lists its skills with their on/off state and switches individual ones off (or back on) for itself. Only a switch: it never installs, edits or deletes a skill, and turning off a common skill leaves every other agent's copy alone. See Switching a skill on or off.

A magic link is a short-lived, PIN-protected page the agent sends into the chat; whoever opens it finishes one task on their own phone or laptop, and the chat is told the outcome (magic_links: has the knobs). Every agent has these tools by default:

Ask Tool What the person sees
Sign in to a service, an MCP server or an official connector (GitHub, Google, OpenAI Codex, Microsoft 365, Slack, Notion, …) magic_link_connect_service the provider's own consent screen, or a one-time code to type on the provider's device page — whichever method the service offers and the bot picked (Olano's app, your own OAuth app, a code, or a secure form for a key); an MCP server that needs an OAuth app of your own (HubSpot) first asks for its Client ID and secret on the same page, or offers Composio instead (MCP servers); the connection serves the service's tools, MCP servers and the git tools at once, no restart
Hand over an API key, token or password magic_link_request_secret a secure form; the value goes straight into the vault, the bot only learns its name
Connect a messenger magic_link_connect_channels one page per platform: QR for WhatsApp/WeChat, a validated bot token for Telegram/Discord/Slack
Edit the bot's instructions magic_link_edit_instructions a one-page editor; saving applies live, no restart
Edit a file magic_link_request_file_edit a one-page editor showing the file's current contents
Open the dashboard on a phone magic_link_open_dashboard a one-time sign-in focused on a section (prompt, vault, connections, providers, config, files, overview)
Check a link you sent magic_link_status —

Only the bot's owners can ask for a link (on a messenger, the channel's allowed-users list), group chats are refused, and on the stricter security profiles every link raises an approval card first. Never paste a key into a chat — if a bot asks you to, it has been misconfigured; the link is the route.

The self_* tools — what a bot may change about itself

Capability Tools Restart?
Know thyself self_get_config (always on while any capability is) —
Browse archetype library self_list_archetypes, self_read_archetype —
Change its own model self_list_models, self_set_model yes
Edit its own instructions self_update_prompt yes
Restart itself self_restart —
Add subagents self_list_subagents, self_add_subagent yes
Connect services self_find_integration (every supported way to reach a service, ranked, with the exact next steps), self_add_toolkit, self_remove_toolkit, self_inspect_external (read an outside project's documentation before anything is installed); self_search_mcp_directory, self_list_mcp_servers (each server's live state, stored sign-in and what would bring it back), self_list_mcp_tools (every tool a server offers, switched-off ones marked), self_reload_mcp_servers (reconnect and re-bind tools without a restart), self_add_mcp_server, self_remove_mcp_server, self_authorize_mcp_server yes for a toolkit or server; a sign-in or a reload applies on its own
Be reachable from outside self_set_external_access (publish itself over MCP / A2A with a public description — administrators only), magic_link_mint_access_token (send an admin a page that mints a caller token, shown once there) yes; the token link applies on its own
Connect messaging platforms self_list_platforms, self_connect_platform, self_connect_whatsapp, self_connect_weixin yes
Manage its own skills self_list_skills, self_enable_skill, self_disable_skill yes
Send magic links the magic_link_* family above —

What a bot may change about itself is controlled per capability — everything on by default — in Config → Agent self-configuration (whole deployment) and per agent under Agents → agent → Chat self-configuration (three-state: Inherit / On / Off). See agent_autonomy:. A single tool can also be switched off for one agent on its Tools tab, by AgentFather (olano_update_agent with disabled_tools), or for every agent with default_internal_tools: false. A bot never gains power over other agents this way: each one can only touch itself. Agents in customer-service mode inherit everything off.

Only the bot's owners can ask for a change. Whoever creates a bot owns it; deployment admins count as owners of every bot, and more owners are granted in Admin → Access. Someone who was merely given access to chat with the bot can still ask what it could do — list models, browse services and templates — but a request to actually change something comes back as a refusal telling them to ask an owner. On a messaging channel the bot's allowed users list plays the same role, so the person who set the bot up can still configure it from their phone.

AgentFather — the olano_* tools for the whole deployment

One agent, AgentFather, holds the tools that reach beyond itself. It is human-only (other agents cannot message it), every write raises an approval card, and it narrates what you are approving. Ask it to:

Job Tools
See and change any agent olano_list_agents, olano_get_agent, olano_create_agent, olano_update_agent, olano_delete_agent; olano_start_agent / olano_stop_agent / olano_restart_agent
Build from the catalog olano_list_builtin_agents, olano_list_archetypes, olano_read_archetype, olano_scaffold_archetype; olano_list_models
Tools and connections for any agent olano_find_integration (every supported way to reach a service for an agent), olano_add_toolkit, olano_remove_toolkit, olano_inspect_external; olano_list_tools, olano_list_bundles, olano_list_integrations, olano_find_tools, olano_set_tool_enabled; olano_search_mcp_directory, olano_get_mcp_directory_entry, olano_test_mcp_server, olano_add_mcp_server, olano_remove_mcp_server, olano_authorize_mcp_server, olano_mcp_oauth_status; olano_list_platforms; the magic_link_* tools aimed at another agent
External access for any agent olano_set_agent_external_access (publish or unpublish an agent over MCP / A2A and set its public description — administrators only); tokens are minted on the dashboard or with magic_link_mint_access_token
Skills and subagents olano_list_skills, olano_list_builtin_skills, olano_create_skill, olano_install_skill, olano_repair_skill_names; olano_list_subagents, olano_add_subagent, olano_toggle_subagent
Teammates and access olano_list_users, olano_get_user, olano_create_user (an invite link, never a password), olano_reset_user_password, olano_set_user_role, olano_update_user, olano_delete_user, olano_grant_agent_access, olano_revoke_agent_access, olano_access_overview
Diagnose olano_diagnose_agent, olano_runtime_status, olano_agent_status, olano_agent_activity, olano_agent_errors, olano_recent_errors, olano_tail_log, olano_list_sessions, olano_read_session, olano_mcp_status, olano_doctor
Deployment settings, disk, memory, updates olano_get_home_config, olano_update_home_config (every config.yaml block); olano_disk_report, olano_disk_monitor, olano_memory_report, olano_memory_monitor; olano_update
Another agent's files olano_list_agent_workspace, olano_read_agent_file, olano_write_agent_file, olano_run_agent_command
Try an agent out olano_test_agent (send it a test message and see its reply and every tool it used; see below)

Testing an agent before you rely on it

Ask AgentFather to "test the support agent with a refund question" and it sends that agent a message on a separate test thread, then reads back the reply and every tool the agent called, with what each one returned. It uses this after building or changing an agent: it tests, adjusts the agent, tests again, and tells you what it tried and what happened rather than what it expects to happen. A test can be one message or a whole conversation, since AgentFather can keep sending follow-ups on the same test thread.

It can also test instructions that are not saved yet, to compare a prompt rewrite before applying it, and test against the agent's saved configuration so a tool or MCP server it just added can be tried before the agent restarts. An agent that is stopped can be tested too. Nothing a test does changes the agent's instructions or configuration.

  • Administrators only. It runs from a conversation with AgentFather, never from a scheduled run, another agent or an outside caller, and you approve each test individually: the card names the agent and the message.
  • The agent's tools run for real, and a test costs a normal turn of that agent's model. Anything the agent needs approval for raises its own card.
  • The agent under test cannot change itself (its instructions, model or connections) during the test.
  • Test threads stay separate from real conversations. They appear in that agent's conversation list with the Autonomous badge (the Agents filter shows them), so you can read the whole exchange, and they are never added to the agent's long-term memory.
  • A test turn stops after 10 minutes unless AgentFather asks for a different limit (at most 30).

The playbooks AgentFather follows ship as two skills only it can see: olano-agent-builder (designing and building an agent from a conversation, Building agents from a conversation) and olano-admin (operating the deployment). Every other agent gets olano-agent (how to do the things above for itself — one finder, one ladder), olano-platform (the map of the platform) and olano-cloud (plans, credits and usage).