Claude and ChatGPT are excellent assistants, and we'll be honest about where they are better - but Ariadne is a different kind of system: it holds long-term goals, notices what changes, acts on your systems, verifies its own work — and runs on your hardware, with your models, under your rules. Ariadne can even use Opus and GPT models.
Ariadne against Anthropic Claude (claude.ai, Cowork, Claude Code) and OpenAI ChatGPT (agent mode, Projects, Tasks). Yes means shipped today; partial means some of it; a dash means not offered.
| Capability | Ariadne | Claude | ChatGPT |
|---|---|---|---|
| Long-term goals, worked continuouslySuccess criteria, plans and journals pursued across days — not prompts re-run on a timer | Yes | —Scheduled sessions re-run prompts | —Tasks re-run prompts |
| Evaluator-verified “done”An independent agent must agree the success criteria are met | Yes | — | — |
| Perception layer & proactive contactTen kinds of watcher; a contact policy decides what interrupts you | Yes | — | — |
| Self-hosted, private, air-gappedYour database, your hardware, your boundary | Yes | — | — |
| Model choiceWhose models you can run | YesAny OpenAI-compatible endpoint, plus our own Kaja & Chikara — mixed per agent | PartialTheir menu | PartialTheir menu |
| Inference routing across your own fleetCache affinity, priority, load; hosts started and stopped on demand | Yes | — | — |
| Project workspacesOne container for a unit of work | YesCharter, repos, releases, tickets, data connections, whiteboards | YesChats + project knowledge | YesChats + files |
| Deep research with citationsMulti-source, multi-round, structured report | YesGrounding anchors, verified answers, editable reports | Yes | Yes |
| Canvas co-editingSide-by-side document the agent edits with tracked changes | Yes | Yes | Yes |
| Live dashboards & BI over your dataCross-filtering, drill-down, forecasts, public snapshots | Yes | — | PartialAd-hoc analysis, not live dashboards |
| MCP — client and serverConnect tools, and expose your own memory as a context lake | Both5,000+ catalogue, OAuth 2.1 | Client | Client |
| A2A — work with other platforms' agentsDelegate to, and accept work from, non-Ariadne agents | Yes | — | — |
| Telephone, voice satellites & IoTSIP calls, ESPHome wake-word hardware, buttons and lights | Yes | — | — |
| Native mobile apps & pushPolished apps with voice, camera and notifications | PartialInstallable PWA; Android app in beta; push designed, not yet built | Yes | Yes |
| Real-time voice conversationLow-latency, interruptible | PartialStreaming voice, room-aware barge-in; full duplex remains | Yes | Yes |
| Runtime tool creationThe agent writes, compiles and loads new tools mid-conversation | Yes | — | — |
| Nightly self-improvement, validation-gatedSkills must beat the old version on held-out evidence | Yes | — | — |
| A model of you that can be wrongBeliefs with evidence, argued against themselves nightly | Yes | PartialMemory facts | PartialMemory facts |
| Run an AI service, not just consume oneMetered billing, spend caps, multi-tenant, public pricing API | Yes | — | — |
| Browser extensionA side panel on the page you're already signed into | —Designed, in progress | Yes | Yes |
| Image & video generationBuilt-in generative media | PartialImages via tools; no video | — | YesIncluding Sora video |
| Enterprise trust artefacts (hosted)SOC 2 / ISO 27001, DPAs, SCIM, audit export | PartialSSO, RBAC, scoped tokens; certification open — self-hosting sidesteps much of it | Yes | Yes |
Snapshot: November 2026. Both competitors ship fast — this page is reviewed against the published comparison in our documentation, which tracks every gap as work, not as an unknown.
Things Claude and ChatGPT users take for granted that Ariadne lacks or only partially covers. Every item here is tracked work with a design behind it — not an unknown.
The web UI is an installable PWA with mobile layout, location sharing and camera handling, and a native Android app is in beta. Push notifications are designed but not yet built — so proactive contact currently reaches a desk, not a pocket.
Our voice stack was rebuilt around streaming transcripts, neural voice-activity detection and interruption that adapts to room noise. Genuinely close now — the remaining gap is true full-duplex (listening while speaking) and the last of the latency.
Computer use and recorded browser automation cover part of it, but neither reaches the page you are already signed into and looking at. A Chrome/Firefox extension is designed and in progress.
Image generation exists through tools; there is no video generation. ChatGPT's built-in image models and Sora are ahead here, and we don't pretend otherwise.
SCIM provisioning, audit-log export and formal SOC 2 / ISO 27001 certification remain open for the hosted offering. Self-hosting sidesteps much of this — your data never leaves your database — but the artefacts matter to procurement teams.
Slide decks are first-class — co-edited live, presented, exported to PowerPoint — and the artifact panel collects everything a conversation produced. Spreadsheets and web apps as first-class rendered deliverables are the remaining slice.
The full numbered list lives in our documentation. These are the themes that matter when you're choosing a platform rather than a chatbot.
Goals with success criteria worked in idle and background lanes; an independent evaluator verifies completion; ten kinds of watcher notice changes; per-goal token budgets and a full activity feed meter every autonomous action. “Keep improving this until 6am” is a first-class instruction — and a watchdog catches runs that go quiet mid-task, the failure that otherwise looks exactly like finishing.
The agent writes, compiles and loads new C# tools mid-conversation. Skills and knowledge are refined nightly from real usage, and a proposed skill edit must beat the old version on held-out evidence before it is accepted. Episodic memory is consolidated, not accreted — and repeated work is mined into inspectable learned reflexes.
Claude and ChatGPT remember facts about you. Ariadne holds beliefs with the evidence behind them, argues against itself nightly, and invalidates beliefs that stop being true — with a date, not a silent overwrite. It treats your different selves on different channels as a first-class distinction, and asks you to settle genuinely contested beliefs instead of guessing. You can read and correct the whole thing.
A model fleet you route, not a menu you pick from: the router resolves each request to a host by cache affinity, priority and load, keeps warm hosts warm, and fails over. Add a machine and it joins the fleet. Hand someone an API key scoped to named models with a spend cap — a key that cannot overspend.
Web, Windows desktop, CLI, IDE, Discord, Teams (including meeting audio), Matrix, email — and telephone calls via SIP, voice satellites, and IoT buttons and lights. Your own ESPHome voice hardware with a custom wake word and no cloud in the middle. It knows who is speaking, and interruption survives a crowded room.
Named agents with distinct skills delegate under evaluator oversight. Goals decompose into a Kanban board humans and agents share; whiteboard, 3D scene, notes and briefings are all agent-writable. And over the open A2A protocol, Ariadne works with other companies' agents — with an explicit allowlist controlling what it may reach.
Support tickets are a first-class object an agent triages and works. Infrastructure thresholds are themselves watchers. Deployments are authored, applied and rolled back through an audited gateway; remediation is evidence-gated. A headless one-shot command lets Ariadne run as a step inside your CI pipeline, not only in a chat window.
Prepay credit ledger, per-model rating including cached-input tokens, plan allowances, spend caps and Stripe top-ups — metered, billable and multi-tenant out of the box. Public endpoints serve your documentation, live prices and embeddable personas. Neither competitor lets you run an AI business on their platform; Ariadne is the platform.
Self-hosted, private, model-agnostic — no lock-in on the model or the platform. Every token, including background work, is attributed to a user, agent, model and purpose. Guard rails are a classifier engine you configure; browser automation never sees your stored credentials. And an Ariadne instance works on Ariadne's own codebase daily — the autonomy is dogfooded, not demoed.
A quick list of things people assume a self-hosted platform must lack. Ariadne ships all of them: scheduled and recurring tasks, code execution in a safety-reviewed sandbox, computer use, persistent cross-session memory, file knowledge and RAG, custom public bots, MCP on both client and server side, a 5,000+ connector catalogue with trust tiers, project workspaces, deep research, inline canvas co-editing, prompt caching, model choice and routing, and cross-session personalisation.
Start free on the hosted platform, or talk to us about running Ariadne inside your own boundary — where the gaps that matter to procurement disappear entirely.