model choice

pick a model. Ivy plans the run and picks the tools.

anthropic direct or through openrouter, across the model families your plan includes — or bring your own key once you sign in. one request reaches the biomedical domain, where Ivy finds the tools that can answer it, runs them in order, and keeps what they returned.

try a guest run see the runtime

four ways in

one run, four ways in.

the web app, curl to /api/v1/orchestrate, the ivybloom terminal app and any MCP client all reach Ivy; a single tool job can also be dispatched directly with curl to /api/v1/jobs, ivybloom run, or a studio project. the tabs say which doors need an account.

one runtime, four ways in — start free here or bring it to your agent.chat + curl need no account; cli + mcp add saved sessions, higher quota and the full catalog.

ask in the browser. guest chat starts a bounded run without an account.

open a guest chat

no account needed · start with up to $1 in guest computesame request, same .bloom* registry — an agent pastes what a person would.just the compute? submit one tool job directly

what happens after you ask

from model choice to artifact.

between the model you choose and the biomedical work: Ivy plans the run, chains the tools, writes it up and, for signed-in runs, keeps the record.

model
Choose among the model families your plan includes; that is what you get. Going beyond them with your own provider key is planned. Anthropic is called directly when configured; the other families go through OpenRouter. A self-hosted deploy can run on your own hardware.
question
The request is interpreted for capability and intent; you do not need to select a tool by hand.
plan
The tools that answer the question are composed in dependency order, with the underlying outputs retained.
result
The run returns a citable answer and whatever its tools made, with a record behind the work.

computed, not described

the language layer routes. the scientific models compute.

the catalogue spans 500+ tools and 50+ scientific databases. what comes back is grounded in the run, not only in what a model has read. every result names the tool, parameters and run that produced it. signed-in runs leave a tamper-evident record you can check.

ready when you are

you leave with what the run made — a structure, a ranked table, an export.

start with a guest run, or compare the access and plan boundaries before choosing a longer-term setup.

try a guest run compare access