Claude
Plan · reported
Organize your workflows
When your development workflow spans multiple agent harnesses, people, and CI jobs, Kiwi gives them a flexible shared structure and record.
What you need to organize work across harnesses.
Keep work, evidence, and authority visible across every step.
Ordered workflow steps
Publish a shared step graph for your runs. Each step declares only the requirements it needs, such as a plan, specific checks, review, or human sign-off.
Remote MCP for existing agents
Connect agents from the harnesses you already use through a hosted MCP endpoint using Clerk OAuth or a project MCP key, with no local Kiwi runtime.
People and review in the loop
Use the web app to inspect reported work, resolve blockers, accept plans, and record final human sign-off on the current candidate before handoff.
Scoped CI check reporting
Include automated checks in the same workflow record. CI jobs post results over HTTP using credentials scoped to the project and its permitted check names.
Attempt history and clean rework
Earlier attempts stay in history. Replacement work needs fresh checks, review, and sign-off wherever the workflow requires them.
Role separation and safe retries
Keep agent updates, independent reviews, human sign-offs, and CI results distinct in the record, with safe handling for retried requests.
How scattered agent work becomes a shared workflow.
Define the workflow
Publish an ordered workflow template, starting with software delivery or your own step sequence, and declare what each stage requires to move forward.
Connect harnesses and CI
Point existing agents to Kiwi's hosted remote MCP endpoint and create scoped HTTP checker credentials for any CI checks in your workflow.
Track progress in one place
Start a run and collect plans, implementation updates, blockers, and check results as agents, people, and CI report into the shared record.
Review, accept, and hand off
Complete required agent or human reviews, record final acceptance for the current candidate in the web app, and move to handoff with the full history intact.
Questions and boundaries
How does Kiwi connect to our agent harnesses and tools?
Existing agents connect explicitly to Kiwi's hosted remote MCP endpoint. People work in the web app, and CI jobs post named check results over HTTP. Kiwi does not passively watch harnesses, ingest external state, launch or stop agents, run code, or provide native Linear, Jira, or TestRail connectors.
Does every workflow step require CI checks and human approval?
No. Each step only checks the requirements declared for that stage in the workflow version pinned to the run. A step might only require structured output fields, while later steps might require an accepted plan, passing named checks, an independent review, or human acceptance.
How are reviews and human sign-off kept separate from implementation?
Kiwi records who submitted every update. An independent agent reviewer must use an eligible identity and session distinct from the implementer. Checker-only requirements accept results from scoped checker credentials. The designated person grants final candidate acceptance in the web app before handoff.
What happens when work needs rework or a run is canceled?
A failed check or review prevents advancement. Request rework, then submit fresh evidence for the replacement work. Earlier attempts stay in history. Replacing a candidate requires new checks, review, and approval for that candidate. Canceling prevents further workflow actions and preserves history without stopping external jobs.
Bring scattered agent work into one shared workflow.
Sign in to set up your workspace and organize your first workflow across agents, people, and CI.