The all-in-one platform for AI agents
Don’t watch a replay.
Step into the loop.
Auteryn agents run in loops — plan, act, verify, then book their own next pass — each on a real machine with a terminal, a browser, and files. Watch every pass live, take the keyboard whenever you want, and hand it back.
Gemini · Claude · GPT · Kimi · GLM · DeepSeek — your model, per agent.
In private beta. Built by a small team — talk to us directly.
- a single task can span — runs checkpoint after every step, and the loop continues until the goal is done or your budget says stop
- Daysa single task can span — runs checkpoint after every step, and the loop continues until the goal is done or your budget says stop
- tools, integrations, and MCP servers — a focused agent carries only what its job needs
- 100+tools, integrations, and MCP servers — a focused agent carries only what its job needs
- a run can park on your question — and resume with full state the moment you answer
- 72 ha run can park on your question — and resume with full state the moment you answer
- channels your agents answer on — WhatsApp to a real phone line, in a designed or cloned voice
- 7channels your agents answer on — WhatsApp to a real phone line, in a designed or cloned voice
The prompt is the job description.
Four prompts, four finished pieces of work — each recreated from a recorded run, prompts verbatim.
The outline lands in seconds, slides fill in as the agent researches, a theme applies everywhere — then it narrates every slide in your voice profile and publishes to a link that presents itself. Export to PPTX, PDF, or MP4.
The agent’s terminal is your terminal.
Auteryn agents work on a real machine — shell, Chromium, filesystem. You can open it at any moment: type commands while the agent works, take the browser cursor, inspect and edit files, restore a snapshot, hand control back. Nothing is hidden, because nothing needs to be.
- Control handoff
- Take the keyboard or the browser cursor mid-run; an expiry timer hands control back to the agent.
- Snapshots and checkpoints
- Save any state, fork the sandbox, or roll back. The run continues from exactly where you left it.
- Sessions persist
- The machine survives across runs, so tasks span days without losing files or context.
When other agents finish, they send you a replay. Here, you were in the room.
Start free — you’re in the room from your first run.
Give it a goal. It runs the loop.
A run is one pass: plan, act on the machine, verify what actually changed, checkpoint. Real work needs more than one pass — so the loop continues: woken by a schedule, by an event like a build finishing, or by a follow-up the agent booked for itself. It ends when the goal is verifiably done — judged by a separate check, not the agent grading its own work — when the budget says stop, or when you do.
Runs last up to two hours each; tasks span days, because the machine, files, and context persist between passes. Agents remember preferences, learn from their own reflections, and read the knowledge your team already writes.
Every loop is bounded. Runs carry hard budgets of time and model calls; follow-up chains are serial and capped; every credit is visible before it’s spent. And a loop only gets the autonomy you gave it.
It gets better every pass.
The loop doesn’t just continue — it compounds. Agents write their own notes, consolidate them, and pull them back by meaning, not keywords. The next similar task starts with the lessons already in context.
- Lessons ride the loop
- A reflection pass after every qualifying run writes down what worked and what didn’t.
- The next pass starts smarter
- A similar task begins with the lessons already loaded — no re-teaching, no repeated mistakes.
- You hold the pen too
- Memory is a page in the console — read, edit, or delete any entry the agent wrote.
Focused agents. One serious harness.
Give an agent forty tools and it spends most of its mind reading tool manuals — bloated context is why long runs wander. Auteryn agents carry only what their job needs: the integrations you connected, the skills you attached, nothing else.
The platform has 90+ built-in tools. A focused agent might carry nine — and ten focused agents beat one that does everything badly.
Schematic, not a benchmark — the shape of the problem, not a measurement.
You choose what it carries
Capabilities, connections, knowledge, and skills are picked per agent — nothing else rides along in its context window.
Skills load on demand
An agent sees a one-line catalog; the full instructions load only when a skill is actually used.
The same harness underneath
Every agent gets the full machine — checkpoints, memory, approvals. Focus costs nothing.
When the job is wide, it swarms.
The lead agent shards the work and spawns up to four focused agents, each on its own isolated machine. When a lane lands, the next shard takes its place — wave after wave, four live at a time — every lane streaming into the same glass box, until the lead merges everything into one answer.
- Four you can watch, not hundreds you can’t
- Swarms cap at four parallel agents on purpose — every lane stays legible and inspectable.
- Isolated by machine
- Each worker runs on its own sandbox — separate files, separate shell, nothing stepping on anything.
- One merged answer
- The lead collects every report and hands you a single result — not four tabs to reconcile.
The same agent answers the phone.
Deploy the agent you built to WhatsApp, Telegram, Slack, Gmail, SMS, a web widget — and real phone calls, in a designed or cloned voice. The same voice profile narrates decks and records podcasts.
the same voice narrates decks and records podcasts
over your threshold — escalated to a human
Phone calls are available on Business plans.
And the call works the other way.
Talk to your agent while it works. Hand it a task by voice and hang up — the work keeps running. Step away and come back, and it opens with what happened while you were gone.
- Delegate by voice
- A spoken request becomes a background run — with a time estimate up front and a heads-up when it finishes.
- Progress on demand
- Ask “how’s that task going?” — the agent answers from the live run: plan, steps, artifacts, elapsed time. Milestones are narrated as they land, including a swarm reporting in.
- Artifacts mid-call
- Charts and documents render in your workspace while you’re still talking.
Watch everything. Approve anything.
The live plan, the model’s reasoning, and every tool call stream into the run timeline as they happen — and the agent stops for a human whenever your rules say so.
Autonomy is a per-agent dial with four levels — Autopilot · Guarded · Supervised · Co-pilot — and approvals ride the loop: a scheduled pass obeys the same rules as one you started by hand.
the pause waits for you — up to 72 hours, state intact
run r-9381 · recreated from a run timeline
The anatomy of a run — the live timeline, the pause that waits for you, and the receipt every tool call leaves behind.
Mid-run, the agent stops and asks.
Six ways it hands you the decision: free text, a choice, a dropdown, a file upload, an approval, or sign-off on a risk-tagged plan. The run resumes with full state.
Permissions are scoped per agent and per API method — allow slack:post, deny slack:delete — with temporary tokens issued per task.
Approvals ride the loop — a scheduled pass obeys the same rules as one you started by hand.
The autonomy dial — set per agent, from Autopilot to Co-pilot.
Org-isolated by design; credentials encrypted and short-lived; every agent action lands in the audit log. The full control list.
Under the hood.
The machinery power users ask about, by name. Everything on this list is live in production today — nothing here is a roadmap item.
And every claim on this page traces to a source-of-truth register we keep in the repo — when a number changes, the register changes first.
- Wakes on events, not timers
- a background job finishes → the agent hears about it on its next turn; no polling loops burning credits
- Books its own follow-ups
- “check CI in 20 minutes,” “resume at 09:00” — serial chains, never trees, three pending at most
- Learns between passes
- a reflection pass after every qualifying run stores what worked and what didn't — the next similar task starts with the lessons
- Lands the plane at budget
- near the budget, a run checkpoints its todos and notes instead of dying mid-thought
- Context compaction + scratchpad
- working memory that keeps week-long tasks coherent between passes
- Repo-wide search + repo maps
- ripgrep in the sandbox; a symbol map of the repository injected into the agent's prompt
- Git worktrees
- parallel work on isolated branches of the same repository
- Swarms + sub-agents
- spawn up to four parallel agents, each on an isolated machine — plus research, review, and computer-use delegations in-run
- Background shells that wake the agent
- start a build, keep working — the agent is told when it exits, no polling
- Durable pauses, up to 72 hours
- a run parks on a question and resumes with full state when you answer
- Fork any conversation
- branch a task at any message; the full context is rebuilt on the new branch
- A shared terminal
- the agent's tmux session is your tmux session — type alongside it mid-run
- Snapshots + restore
- the machine's state, saved and restorable at any point — checkpointed after every step
- Self-evolving memory
- agents write and consolidate their own notes, searchable in natural language — and you can read and edit every entry
- Skills drafted by AI
- describe a workflow — or hand it a chat thread, a screen recording, or a YouTube link — and Studio writes the Skill for your review
- Any MCP server
- point an agent at yours — plus a featured catalog of 28
- Your model, per agent
- Gemini, Claude, GPT, or open-source — a dropdown, not a migration
And it runs on your own machine, too — enroll a Linux box with one command, outbound-only, zero sandbox-compute credits. Your machine.
Works with your stack.
20+ native integrations and customer channels. A featured catalog of 28 MCP servers, vendor-hosted and self-hosted. Every Google Workspace surface from the sandbox CLI — and any MCP server you point it at. All integrations.
Native integrations, channels, and sandbox CLIs
The MCP catalog — vendor-hosted and self-hosted
If your stack isn’t here, point the agent at any MCP server you run.
One subscription. A whole team of agents.
Pricing is live; capacity is gated while we’re in beta.
Founding members get 30% off for 12 months — ask us for your code.
A coding agent before standup, a research agent at noon, a support agent all night, a voice that answers the phone and narrates the deck. On other stacks that’s four subscriptions and a prayer of integrations. Here it’s one platform and one balance of credits — AI, sandbox compute, storage, voice, and generation all draw from it. Start free; scale when the work does.
One balance. Hard budgets.
Budgets are per run and per goal — a loop can never outspend the caps you can see.
1 credit = $0.001 · 1,000 credits = $1 — forever.
Free
Run a full agent and see the sandbox for yourself.
- Credits / mo
- 1,000
- Compute hours / mo
- 1
- Storage
- 100 MB
- 1 user
- 3 agents
- 1 knowledge store
- 3 loops
- 5 connections
- Workspace chat only
- Community
Pro
Most chosenFor solo builders running real work daily.
- Credits / mo
- 15,000
- Compute hours / mo
- 25
- Storage
- 5 GB
- 3 users
- 10 agents
- 10 knowledge stores
- 10 loops
- 20 connections
- Bring your own machine — no sandbox-compute credits
- Workspace voice
- WhatsApp · Telegram · Slack
- Video generation (10 clips/day)
- Email · 24h response
Business
Customer-facing agents on real channels.
- Credits / mo
- 60,000
- Compute hours / mo
- 100
- Storage
- 50 GB
- 10 users
- Unlimited agents
- 100 knowledge stores
- Unlimited loops
- 100 connections
- Bring your own machine — no sandbox-compute credits
- Everything in Pro
- Inbound phone numbers
- Video generation (50 clips/day)
- Priority · 4h response
Enterprise
Isolation, SSO, audit, and a human on call.
- Credits
- Custom
- Compute
- Unlimited
- Storage
- Unlimited
- Unlimited users · agents · knowledge
- SSO / SAML
- BYOK (customer-managed keys)
- Audit logs · Custom roles
- Bring your own machine — no sandbox-compute credits
- Dedicated cloud account
- VPC peering · private endpoints
- Custom telephony
- 24/7 · Dedicated CSM
- 99.9%+ custom SLA
Sandbox compute beyond your plan’s hours: 1.5 credits/min. Storage is included with your plan. All prices in USD. Taxes may apply. Need a tailored plan? Talk to us.
I built Auteryn because I wanted an agent I could trust with real work — which meant I had to be able to see everything it does and interrupt it at any moment. So the agent got a real machine, and you got a seat at it.
Agents everywhere run in loops now — plan, act, verify, again. We built the loop with a seat in it.
We’re a small team in beta, and I want that to work in your favor. Three promises: nothing an agent does here is ever hidden from you. When something breaks, you tell me — and watch the fix land in the changelog. And every email is answered by me, not a queue.
Founder, Auteryn
See a run for yourself.
The free plan is a real agent on a real sandbox — 1,000 credits a month, no sales call.