Describe the work. Get a team that does it.
Traven reads your brief and builds a team of AI agents around it. They split the work between them, run it at the same time, and hand back the files — and you watch the whole thing happen.
$20 per seat, per month. Usage at provider cost plus 10%.
The idea
Not one assistant. An org.
A chatbot does one thing at a time and you are the one holding the plan — deciding what to ask next, remembering what the last answer said, stitching it together yourself.
Traven does the deciding. It reads the brief and builds a structure for it: an executive that works out what the job actually is, managers that each own a slice, and workers that go and do it. Each agent is given its own mission, its own model and its own slice of the budget, and reports back up the chain when it is done.
A manager can delegate twice — once from the brief, and again from what its reports brought back. That second round is the difference between running a plan and running a team.
Every agent is also a folder. The org chart and the file tree are the same tree, so “who is working on this” and “what came out of it” are one view rather than two.
The brief was “find our five closest competitors and write a positioning brief”. Nobody drew this tree — the executive decided the shape, named the roles and wrote each one its own brief.
PositioningExecutive
CompetitorsMarket Research
- Five competitors, in detail
PricingPricing Analyst
- How each one charges
BriefPositioning Writer
✓ finished and reported back · ● still working
What it can do
Reach, and the receipts for it.
- Runs real code
- Every run gets a disposable Linux container. Agents clone a repository, install its dependencies, run the tests and open a pull request — actual execution, not a description of what the code would do if you ran it.
- 108 connectors
- 108 systems in the catalog, and 66 of them connect in one click with nothing to set up. GitHub, Slack, Notion, Linear, Stripe, Supabase, Google Drive, Docs, Sheets and Calendar, HubSpot, Attio, Cloudflare, Neon, Railway, Sentry, Datadog, Midjourney, Runway, ElevenLabs.
- Reads the live web
- Agents search and fetch pages while they work, and bring back the sources they used, so a claim in a deliverable can be traced to where it came from.
- Remembers
- What a run establishes is written down and found again by the next one. Later work starts from what earlier work already learned instead of from nothing.
- Routes across models
- Twelve models from seven vendors, picked per job by what the job needs — and failed over across vendors when one is down. You set the cost tier, economy to premium, and it takes the best model under that ceiling.
- Waits when it should
- Anything that changes something outside Traven can be set to queue for your approval instead of executing. How that works.
How it works
Four steps, and you can see all of them.
Say what you want.
One sentence or one page. “Find our five closest competitors and write a positioning brief.” You are describing an outcome, not a workflow — there is nothing to wire up first.
It builds the team.
The executive decides the shape, names the roles and writes each one its brief. The tree fills in on screen as it goes, with the model each agent was routed to and why.
They work, at the same time.
Agents run in parallel, calling only the tools their connections allow, and report back to whoever created them. You can stop the run at any point.
You get files.
Deliverables land in the folder tree as markdown you can read and edit — with the brief, the tool calls and the report that produced each one still attached to it.
Trust
You are handing agents your credentials.
That is the first objection everyone has and it is the right one. So the controls are structural rather than promised, and they are written down in detail rather than summarised into a badge.
What is actually in place
- Every tool is classified by how far its effect reaches, and anything unrecognisable is treated as the dangerous kind.
- Each connection has one of three autonomy levels. On Ask me first, a change is parked in a queue for you instead of being sent.
- Agent code runs in a container created for the run and destroyed when it ends. Never on your machine.
- Every run carries a hard dollar ceiling, and admins can add monthly ones per person and for the whole team.
- Credentials are encrypted at rest, and stripped out of anything a model is shown.
- We hold no security certification, and we say so on the page rather than leaving you to notice.
Start
Give it something you would otherwise do yourself.
A competitor scan, a migration nobody has picked up, the report you rewrite every month. Watch what it builds, and stop it if you do not like the shape.
See pricing — $20 a seat, usage at cost plus 10%.