Ariadne brings agents, memory, tools and integrations into one continuous workflow. It holds long-term goals, notices what changes, acts on your systems — and an independent evaluator checks the work. Its mark is a glider: a pattern that moves under its own momentum.
Conventional AI reacts to prompts. Ariadne holds a live model of your problem space, remembers what it has learned, and pursues long-term intent — taking the lowest-effort actions that keep conditions moving in the right direction.
Watchers notice changes in feeds and web pages, goals are worked in the background, and a daily digest reports what happened while you were away. Ariadne doesn't wait to be prompted.
What Ariadne learns is embedded into vector memory automatically and eventually fine-tuned into the model itself. Skills are refined by self-assessment of success and supplied to the agent directly — not called via tools.
Run the whole platform on your own hardware with your own database, or in your cloud. Model-agnostic too — use our in-house Kaja models or any OpenAI-compatible endpoint.
Scheduled AI tasks re-run prompts. Ariadne pursues outcomes across days — and proves they're met.
Agents can form their own goals from what they learn about the people they work with — always budgeted, always reviewable by you.
A per-user contact policy — channels, urgency thresholds, quiet hours and rate limits — decides what interrupts you now and what waits for the morning digest.
Ariadne connects agents, your data, tools and communication in a single workspace shared by people and agents — switch modes without switching apps.
A conversational assistant, a coding mode that writes, reviews and runs code — with real file and terminal access via the Desktop, Console or Docker — and a query mode for structured data. Every conversation has its own settings: its own agent, tools and MCP servers, decided per chat rather than globally.
Goals decompose into a Kanban board shared between humans and agents — plans, journals and comments in one place — while automated briefings summarise your topics on the schedule you set, and can be published publicly when they're worth sharing. Boards belong to project workspaces, sitting alongside their charter, documents and release history.
Markdown notes with semantic search — and live co-editing: you and the agent edit the same note at the same time, every change kept with who made it. Dynamic knowledge is fetched from the web the moment it's needed and refreshed nightly; knowledge from conversations is captured as notes that you or your agents can add to. All of it is embedded automatically — no tool calls required.
A real-time collaborative canvas — shapes, text, connectors and freehand drawing, with others invited in live. Ask an agent to generate a Mermaid diagram, or to expand a rough sketch into a finished one.
Agents build in three dimensions — placing and arranging named objects in a shared scene that updates live as the work happens. System architectures, network topologies, data flows — no CAD tool required.
Configure agents with their own prompts, model, tool collections and knowledge. Browse a catalogue of 5,000+ MCP servers and install one onto an agent in a single click — or let external tools connect to Ariadne. And with A2A, agents on other platforms can work with yours, in either direction.
Ariadne sees and acts on the infrastructure your code runs on — Docker, Portainer, Azure or AWS — through a single audited gateway, with every mutating action scope-fenced and policy-gated. Plan and apply deployments with automatic rollback, or provision disposable remote coding sessions on machines you control.
Register a project's test suite and the agent detects the framework — NUnit, xUnit, pytest, Jest or Vitest — runs it, and reports structured results back in chat: passed, failed, skipped. No trip to a CI dashboard, and full runner output stays on hand for when you want the detail.
Incoming tickets become structured investigations: Ariadne correlates recent deployments, metric breaches, restarts and log errors into ranked hypotheses — read-only, and safe to run automatically. Remediation is classified, evidenced and gated; anything destructive is never automatic, and an approval is single-use.
One unit of work — a product, a client engagement, a research programme — gathered into a single place: charter, conversations, kanban work, documents, files, repositories, whiteboards, data connections, releases and support tickets. The charter is a standing brief Ariadne reads before every message in every conversation bound to the project — context written once, never re-explained.
A side panel docks beside every conversation. Artifacts collects what the chat produced — documents, rendered HTML, images, code and files — and Canvas is a document you and the agent co-edit live, with tracked revisions and attribution. One click promotes any artifact into a versioned canvas document; nothing is lost in the transcript.
An evolving model of the person you are — beliefs with the evidence behind them, your traits, mannerisms and preferences, refined against itself nightly. Read it, correct it, and when a belief is genuinely contested it asks you to settle it — never a silent guess. Off by default, and per agent.
Ariadne reaches out to real systems — and anything that speaks MCP or the OpenAI API can reach it back.
One question in, a cited report out. Ariadne plans sub-questions, searches, reads a dozen or more sources, extracts evidence and goes back out for what it still doesn't know — then writes the report with numbered citations you can check. Name a source in your question and it becomes a grounding anchor, read first and cited as yours. Reports stay editable afterwards, not just re-runnable.
Web pages, YouTube videos, audio recordings and documents condensed to what matters — brief, bullets, detailed analysis or full transcript.
Crawl and index the content of whole websites, then let agents search and retrieve exactly what they need.
Record and replay browser flows that repair themselves when a page changes — with credential isolation, so the model never sees your stored passwords.
Read, send and triage connected IMAP/SMTP mailboxes, view calendar events, transcribe recordings and have any content read aloud.
Any note or folder becomes a read-only public page with a share link — no login for visitors, nothing editable. Scope is live and per-document, and every publication doubles as a tiny JSON API, so you can build a documentation site straight out of your notes. Ariadne's own documentation is served this way — pull it from the public API at /api/public/v1/docs/ariadne.
Your organisation's knowledge — notes, knowledge bases, conversations — exposed as an MCP server with OAuth 2.1. Any MCP-compatible agent, in your IDE or elsewhere, works with the same memory Ariadne does — while a 5,000+ entry catalogue, synced from the official MCP registry, connects new tools in one click.
Your agents work alongside agents built on other platforms, over the open A2A protocol. An external agent can send yours a task and watch the progress as it works — or your agent delegates a job to theirs and picks up the answer when it arrives. MCP adds tools; A2A adds colleagues.
An OpenAI-compatible /v1/chat/completions gateway — anything that speaks the OpenAI API can use your Ariadne agents, memory and tools included.
Plus multi-step workflows, inbound webhooks, cron schedules and email triggers — automation in every direction.
Ariadne doesn't wait for releases to get sharper. It reviews the day's work overnight — and an Ariadne instance actively works on Ariadne's own codebase.
New knowledge is embedded into vector memory the moment it's learned, then eventually fine-tuned into the model on the nightly cycle — not frozen between releases.
Agent skills are refined against the agent's own assessment of what succeeded, then supplied to the agent automatically — approaches that didn't work are dropped.
Episodic memory is deduplicated, decayed and promoted into durable knowledge — the world model improves rather than accretes.
Missing a tool? Ariadne writes, compiles and loads a new C# tool mid-conversation — then uses it.
Web, desktop, terminal, telephone — one memory behind all of it.
The full browser UI, plus a native Windows app with streaming chat, notes and session management — QR-code login, auto-updating, no .NET install needed. The web client is an installable PWA, so the same agent runs on a phone or tablet from your home screen. With your consent it can even operate Windows itself — scoped to a window or the full desktop, revocable, and screenshotted before and after every action.
A console client for developers and headless environments: run as an ACP server inside JetBrains or VS Code, with real local file access and program execution on your machine or in a container.
Voice satellites, telephone calls via SIP, Discord servers and Microsoft Teams — including meeting audio — Matrix, and inbound email triage. Plus physical accessories: buttons and lights wired into the same agent, and Home Assistant Voice hardware speaking Ariadne's own wake word, with no cloud in the middle.
Publish a capability-scoped, rate-limited, origin-locked chat widget on any website — no login required for your visitors. The chat widget in the corner of this page is a live Ariadne persona — say hello.
An Ariadne instance works on Ariadne's own codebase — the autonomy features are exercised on real software engineering, every day.
Every conversation lives in your own PostgreSQL with pgvector search. Nothing leaves your boundary unless you point it somewhere.
From a hosted trial to a fully air-gapped installation on your own hardware — the platform meets your governance, not the other way around. Every deployment includes the MCP context lake and the OpenAI-compatible API.
Get going in minutes on our managed Grace Blackwell infrastructure. Ideal for evaluation and smaller workloads.
A dedicated Ariadne instance in your own Azure, AWS or Google Cloud tenancy, inside your network boundary.
Full on-site deployment — including air-gapped — with our engineers handling installation, tuning and handover.
A chatbot answers and stops. Ariadne maintains long-term goals, keeps a live model of your systems, and takes real actions toward them — continuously, with an evaluator verifying the result.
Scheduled tasks re-run prompts. Ariadne pursues outcomes across days: planning, acting, noticing changes via watchers, and confirming its own success criteria before it calls a goal done.
Wherever you decide. With private-cloud and on-premise deployments your data and models never leave your boundary — every conversation lives in your own PostgreSQL.
Our in-house Kaja and Chikara models, trained on our own Grace Blackwell cluster — or bring your own: Ariadne works with any OpenAI-compatible endpoint.
Your organisation's knowledge — notes, knowledge bases, conversations — exposed as an MCP server, so any MCP-compatible agent works with the same memory Ariadne does.
Yes. Begin free on the hosted tier, prove the value on one problem, then scale to a private-cloud or on-premise deployment when you're ready.
Yes — over the open A2A protocol. Other platforms' agents can send yours a task and stream its progress, or your agent can delegate work to theirs and collect the answer when it's ready. Ariadne also exposes itself as an MCP server, so any MCP-compatible client can read the same memory.
One unit of work gathered into a single place — charter, conversations, kanban work, documents, repositories, releases and support tickets. The charter is a standing brief Ariadne reads before every message in every conversation bound to the project, so context is written once instead of re-explained in each chat.
Start free on the hosted platform, or talk to us about a private-cloud or on-premise deployment tuned to your organisation.
