Blog
No. 03·Product preview · KumiChem

Your research.
Your way to run it.

Tools, models and research workflows—available from the platform or through your own agent.

Choose the mode that fits the task: a single tool, your own agent loop, a delegated workflow or platform chat.

Research without the environment setup

We are extending KumiChem so researchers can log in, use tools and models, and take a research task through to a report—without first assembling the underlying software environment.

The platform will connect sequence search, RNA structure tools, prediction models and research workflows. Models and tools that need specialized dependencies or reference databases will be provided through our cloud services. Users can focus on their inputs and research questions rather than maintaining those environments.

If you already work with Codex, Claude Code or another agent, you will also be able to connect through APIs and Skills. Skills are instructions your agent can read to understand a tool's inputs, supported settings and outputs, or to follow a research template.

Find the mode for this task

KumiChem will support four modes, A–D. The decision tree gives you a starting point; the examples below explain what each mode looks like in practice.

Q1Can one tool or model handle this task?
YesAUse an individual toolUse it from a web form, a Skill or the tool API.
No→ Go to Q2
Q2Do you want your own agent to carry the workflow forward?
YesBRun the workflow in your own agentYour agent manages the whole process and uses KumiChem's tools and templates.
No→ Go to Q3
Q3Will your software or agent call a KumiChem hosted workflow?
YesCUse the hosted workflow APIYour agent delegates the work and KumiChem runs it.
NoDUse platform chatWork directly in the browser.

One tool can process a batch of inputs. A hosted workflow is a complete task delegated to KumiChem, rather than a series of tool calls managed by your agent.

What each mode looks like

A · Use an individual tool or model

Choose A for a defined task, such as RNA folding, sequence search or a set of activity predictions. One tool call can accept a batch of inputs; it does not need to start an entire research workflow.

On the website, select the tool, follow its input example, adjust the supported settings and submit. The result can include structured predictions, tables, figures or downloadable files. In your own agent, a Skill explains how to call the same module and interpret its output. Your agent can then combine that result with other resources.

“Calculate the secondary structures of these RNA sequences and help me compare the results.”

B · Run the workflow in your own agent loop

Choose B when you want your host agent to organize the research steps. This can suit users who already have a Codex or Claude Code subscription and prefer to keep planning, context, writing and review in that environment.

To start, give your agent the setup instructions from API & Skills and let it install the whole Skills suite, then use the example request below. The bundled research template supplies the method's steps, default settings and evidence requirements. Your agent reads it, discusses the target, chemistry and study parameters with you, and confirms the choices before calling tools. These defaults provide a starting point; they do not make every study identical.

You can inspect intermediate results, change the exploration path, combine other resources, and decide how to organize the report and its checks. Your host runs the loop and manages revisions, while KumiChem supplies the requested tools and models.

“Use the KumiChem ASO template to explore LNA gapmers for my target. Confirm the transcript, chemistry and study parameters with me before using the tools.”

C · Delegate a complete workflow through the API

Choose C when a study fits a prepared KumiChem workflow and you want its template, established agent loop, validator and review process to run together. This reduces the work of rebuilding and coordinating that complete execution process inside your own host.

Your agent can still manage the overall project. It treats KumiChem's research agent as a submodule: submit the agreed task through the hosted API or Skill, keep its task reference, monitor progress, and retrieve the report, review findings and supporting files.

01 · Your agentPlan the projectCompare early results and define the ASO study to delegate.
02 · KumiChemRun that workflowExecute the study, validate its evidence and review the report.
03 · Your agentContinue the projectUse the returned findings to decide the next task with you.
“Delegate this agreed ASO study to KumiChem. Bring back the report, review findings and supporting files so we can decide the next step.”

D · Work directly in platform chat

Choose D for a guided browser experience, without setting up an external agent or writing an API integration. Log in, describe the study, and discuss the inputs and method before execution.

The workspace will keep the request, progress, report and follow-up discussion together. A complete study can produce a readable report, including a PDF and supporting tables, figures and evidence files. You can download them or continue the conversation to request a revision.

01Ask & configureConfirm the goal, inputs and study choices.
02Run & reviewFollow tool execution and evidence checks.
03Read & reviseDownload the report or request a follow-up change.
“Revise the explanation of these recommendations without changing the sequences or repeating the tool runs. Keep the previous report and provide a revised PDF.”

B and C can both sit inside a project managed by your own agent. B executes the workflow steps in your host; C delegates that complete workflow to KumiChem. C and D use the same hosted execution and review capability, accessed through different interfaces.

What the hosted workflow checks

In C and D, validation is part of the execution process. Programmatic checks and multiple large-model agents in author and reviewer roles compare the report with the tool outputs and supporting evidence.

  • Evidence consistency: Do the sequences, settings, tables and claims match the recorded outputs?
  • Model interpretation: Are prediction direction, applicability and experimental conditions described correctly?
  • Completeness: Are required steps and report elements present, with missing evidence identified?
  • Revision: Have the review findings been addressed and checked again?

Review findings are returned to the author for revision, followed by another check within defined limits. The handover retains the review trail and any unresolved issues; reaching the end of a run does not by itself mean every scientific requirement has passed.

In B, the template still describes the evidence expected from the study. Your host decides how to carry out the checks and revisions. Reading a template alone does not invoke KumiChem's hosted validator.

These checks support traceability and consistency of the analysis. Biological performance still requires experimental validation.

Context, subscriptions and handover

With your own agent, conversation and project memory follow the host you choose. Tool inputs still go to the relevant cloud service. If you delegate a workflow, you select the inputs, evidence and scope to hand over; the platform does not need your entire local agent history.

A host subscription and KumiChem service usage are separate. Using an existing subscription can be a reason to choose B, while tool calls or hosted execution remain subject to the corresponding service usage. You can combine modes within a project rather than committing every task to one arrangement.

Tools and models, where you work.

Explore the platform we are developing for guided research and agent integration.

Explore KumiChem ↗