agents

Agents

Agents are the AI personas at the heart of Ariadne. Each agent has its own system prompt, tool collection, reasoning settings, and optionally connected knowledge bases and MCP servers. You can create multiple agents for different roles.

Accessing Agent Settings

Go to Settings → Agents in the top navigation.

Creating an Agent

1. Click New Agent
2. Enter a name (displayed in the UI) and an optional description
3. Write the system prompt — this is the core instruction that shapes the agent's personality, behaviour, and capabilities
4. Select a tool collection (see below)
5. Choose a model and reasoning level
6. Click Save

System Prompt

The system prompt is the foundational instruction for the agent. It defines:

• The agent's persona and communication style
• Its primary role and areas of expertise
• Any constraints or guidelines it should follow
• Background context it should always keep in mind

Tips for effective system prompts:

• Be specific about the role: "You are a senior software engineer specialising in Python and distributed systems"
• Include communication preferences: "Always respond concisely; use bullet points for lists"
• Add relevant context: "You are working in a company that builds fintech products for SMEs"
• Use Knowledge → Agent Skills for longer, reusable instruction sets

Tool Collections

Tool collections define which tools are available to an agent during a conversation. Tools extend what an agent can do beyond text generation.

Built-in Tool Collections

Ariadne includes several pre-built tool collections:

Assigning Tools

1. Open an agent in Settings → Agents
2. Go to the Tools tab
3. Enable or disable tool collections as needed

You can also enable or disable specific tools within a collection for finer-grained control.

Conversation-Specific Tools

Some tools are automatically added for specific conversation modes (Code, Query, etc.) or when data sources are attached to a conversation.

Reasoning Levels

Agents can operate at different levels of reasoning depth:

Select the reasoning level in the agent settings. You can also override it per conversation.

Model Selection

Choose which underlying language model the agent uses. Available models depend on your configuration. Go to Settings → Agents → Model tab to select from discovered models.

MCP Servers

MCP (Model Context Protocol) servers extend agents with additional external tools, resources, and prompt templates. An MCP server is a separately running process or remote service that exposes capabilities via a standard protocol.

The MCP Catalogue

Ariadne carries a built-in MCP Catalogue of over 5,000 remote integrations, kept in step with the official Model Context Protocol registry and browsable from Settings → MCP Servers. Rather than configuring endpoints by hand, you browse or search the catalogue, pick a server, and install it onto an agent in one click — the URL, transport and authentication method all come from the catalogue entry.

Entries are grouped into categories (development, data & storage, productivity, communication, finance, cloud & infrastructure, search & research, AI & models, security, media & design) and badged with how much is known about who published them. See MCP Servers for the detail.

Installing a server attaches it to the agent. It is not automatically switched on in your conversations — see per-conversation MCP selection.

Adding an MCP Server

1. Go to Settings → MCP Servers
2. Select the agent you want to configure
3. Click Add Server
4. Choose a connection type:
   • URL (Remote Server) — connect to a remote MCP server over HTTP/SSE. Available to all users.
   • Command (Local Process) — run a local process (e.g. npx, python). Administrators only, because command servers execute arbitrary processes on the Ariadne host.
5. Enter the server URL or command and any arguments
6. Set the visibility (who else can see and use this server — see below)
7. Configure authentication if the server requires it
8. Click Add MCP Server

The server's tools, resources, and prompt templates are discovered automatically and become available to the agent shortly.

Advanced MCP Routing and Security (RouterMcpAuth & RouterMcpBridge)

For enterprise and multi-tenant installations, Ariadne incorporates advanced security and routing layers to mediate MCP tool execution:

• RouterMcpAuth: Manages secure authentication handshakes (including OAuth 2.1 and static API keys) on behalf of your agents, ensuring that user and system credentials are never exposed directly to models or logged in conversation histories. Supports both shared keys and per-user credentials.
• RouterMcpBridge: Seamlessly bridges tools between remote SSE (Server-Sent Events) hosts, local command processes, and custom cloud-hosted environments. It automatically handles routing protocol translation, optimizes latency budgets, and applies per-tool execution policies at runtime.

Visibility & Ownership

Every MCP server configuration has an owner (the user who created it) and a visibility setting that controls who else can see and use it:

The owner (or an administrator) can change visibility at any time via the Manage Visibility & Auth button on the server card.

Authentication

MCP servers that require authentication support two modes:

When a server uses per-user authentication, each user sets their own token via the server card in the MCP Servers page. Tokens are stored securely and never displayed back.

MCP Resources & Templates

Beyond tools, MCP servers can expose:

• Resources — readable data items (files, documents, data feeds) that the agent can access alongside conversation files. Resources appear in the list_blob_files tool output alongside uploaded files.
• Prompt templates — pre-defined conversational templates that users can invoke with /prompt <name>. Templates are listed by the get_blob_templates tool.

Managing MCP Servers

From Settings → MCP Servers you can:

• View all MCP servers visible to you, with owner and visibility badges
• Manage visibility, authentication, and connection details
• Set your own credential for per-user authenticated servers
• Remove servers you own (or any server if you are an administrator)

MCP servers that are unresponsive are automatically excluded from tool calls to prevent delays.

Ariadne as an MCP Server

Ariadne can also act as an MCP server itself, exposing its capabilities to external MCP clients such as Claude Desktop, other agent frameworks, or custom integrations.

• The MCP endpoint is available at the URL shown in the banner on the Settings → MCP Servers page
• Connections are secured with OAuth 2.1 — each external client gets its own authorization with per-connection scopes you approve
• Exposed capabilities include searching notes, reading conversations, listing agents, reading briefings, and summarising URLs
• Manage connected external clients under Settings → Connections

This means you can use Ariadne's knowledge base, notes, and conversation history from any MCP-compatible tool.

Agent Goals

You can assign long-term goals to an agent — ongoing objectives that the agent works toward over time. Agents can also define goals of their own during an overnight review, worked on in budgeted idle time. See Goals & Schedules for details.

Self-Improvement

As agents work, they automatically capture reusable procedures as Agent Skills and tidy their skill and knowledge sets overnight. Every automatic change is version-snapshotted and reversible from the Knowledge page.

Schedules

Agents can be triggered on a schedule — for example, to run a morning briefing or check your inbox daily. See Goals & Schedules.

Working With Agents on Other Platforms

Your agents can also work alongside agents built on other platforms using the open A2A protocol — either answering their requests or handing work to them. See A2A — Working with Other AI Agents.

Multiple Agents

Using multiple agents allows you to:

• Keep different concerns separate (a research agent vs. a coding agent)
• Use different models for different cost/quality trade-offs
• Configure strict tool access per domain

When starting a new conversation, select the agent you want to work with from the agent picker.