Platform
Every agent gets a real computer.
Helios runs each agent in its own isolated cloud sandbox: frontier models, code execution, a file system, and live web access — wired to the full APIs of your tools, your databases, and any MCP server you attach.
Our Q3 usage looks off. Pull the numbers from the warehouse, work out which accounts drove the change, and write me a memo.
- clickhouse·query events · Q2 vs Q3 by account
- code·run cohort-delta.ts · group by account
- web·search funding news for top 5 accounts
- workspace·write q3-usage-memo.md
- google-docs·create doc · share with #revops
Q3 usage memo delivered
- Change traced to 3 accounts that migrated off the legacy endpoint
- Two raised a round this quarter — expansion candidates flagged
- Chart and raw CSV saved to the agent's workspace
The sandbox
A computer of its own.
Each run starts in an isolated cloud sandbox, walled off from your infrastructure and from every other run.
Code execution
Agents run shell commands and write scripts inside the sandbox. When a task needs a join, a chart, or a one-off parser, the agent writes the code and runs it.
A file system
Read and write real files. Download an export, transform it, keep the intermediate artifacts, and hand back the finished document.
Web search and scraping
Search the live web and pull the contents of a page into the run, so an agent can research a company or check a status page before it acts.
A persistent workspace
The sandbox is disposable. The workspace endures — files written on Monday are still there on Friday, backed by durable object storage.
The model
A frontier model, tuned for the task.
Every agent runs on Claude Sonnet, tuned for long-running tool use — reading APIs, writing code, and checking its own work across a run.
One model, deliberately chosen, means every task gets the same judgment — the scheduled report and the one-off research question alike. The rest of the agent — its computer, its tools, its memory — is built around it.
Research the three competitors we lost deals to last month and summarize how they price.
- attio·list closed-lost · reason=competitor
- web·search pricing pages · 3 vendors
- apify·scrape pricing tables
- notion·update Competitive Landscape
Competitive brief updated
- 3 vendors · pricing model, entry price, and enterprise gate
- Sources cited inline for the sales team
The tools
Reach that matches the task.
An agent is only as useful as what it can touch. Helios hands it everything your team already works in.
The full API of every integration
Helios indexes each service's complete OpenAPI spec and lets the agent search it at run time, so it can call whichever endpoint the task needs. Credentials are attached server-side and stay out of the model's view.
Your databases
Connect Postgres, MySQL, or ClickHouse — over an SSH tunnel where your network requires one — and let agents query production data directly.
Any MCP server
Attach an MCP server and its tools join the agent's toolbox. Helios speaks MCP outward too, so your own clients can drive Helios agents.
Memory
The work compounds.
Learnings, files, and pinned context carry from one run into the next.
Memory across runs
Corrections and preferences persist. Teach an agent your escalation rules once and every future run applies them.
Pinned repositories
Pin a GitHub repository as a scoped resource and the agent carries it as standing context on every run.
Variables and secrets
Store the values a task needs — an API key, a threshold, a channel name — encrypted and reusable across every agent you run.
Two modes
The same agent, on demand or on duty.
Talk to an agent in Chat when you need something now. Save that agent as a Workflow to put it on the task permanently. Same runtime, same tools, same memory.
Which enterprise accounts opened a support thread this week and have a renewal inside 60 days?
- plain·list threads · 7 days
- postgres·join renewals on account_id
4 accounts match
- Answered in chat · 12 seconds
Do that every Monday and post it to #customer-success.
- plain·list threads · 7 days
- postgres·join renewals on account_id
- slack·post #customer-success
Workflow saved and scheduled
- Runs forever · full run history kept
Keep going
Go deeper
The runtime is one story. Here is how teams put it to work.
Chat
Talk to your agents
Delegate in conversation, upload files, pin the repositories an agent should know about, and share the session with your team.
Workflows
Put an agent on the task
A plain-English prompt and a trigger. Cron, signed webhooks, integration events, or on demand.
Use case
Database reporting
Watch an agent query your warehouse, analyze the result in code, and deliver the report where your team reads it.
Give an agent a computer.
Start free, describe the first task, and watch the run unfold step by step.