Kestrel
Build and orchestrate AI agents that work while your team sleeps.
Business · Sand & Petrol · Fraunces / Instrument Sans · motion warm
Build and orchestrate AI agents while you sleep.
Compose agents from a library of parts, wire them into a workflow and let them run. Kestrel handles the scheduling, the retries and the bill.
Start building- 50M+tasks run
- 99.9%uptime
- 180+countries
One platform. Every part of the job.
Build, watch, remember and connect, in one place.
Visual agent builder
Drag a goal, a model and some tools onto a canvas, and it runs
Real-time monitoring
Every agent's current step, cost and output, live
Memory and context
A shared store agents read and write, that you can inspect
Tool marketplace
Three hundred connectors, each with its own permission scope
Plug-and-play agents
Eight agents that work on day one. Copy one, change what you like, and it is yours.
Research analyst
Reads, summarises and cites, with a source for every claim
Sales prospector
Finds accounts that match a profile and drafts the first email
Support triage
Reads every ticket, tags it and answers the ones it is sure about
Code reviewer
Runs the tests, reads the diff and leaves comments a senior would
Data cleaner
Normalises spreadsheets and flags the rows it cannot fix
Content editor
Rewrites to a style guide and explains every change
Ops monitor
Watches dashboards and pages a human when something drifts
Recruiting screener
Reads applications against a rubric and ranks them, with reasons
From idea to running agent
Four steps, usually an afternoon.
Define
Write the goal in plain language and pick the tools it may use
Compose
Chain agents into a workflow with branches, loops and approvals
Test
Run against real data in a sandbox and read every step
Deploy
Schedule it, trigger it from an event, or expose it as an API
Build custom tools
A tool is a function with a description. Write one, register it, and every agent can ask for it.
Any language
Tools run as HTTP endpoints, so use whatever you already have
Typed inputs
Describe the parameters once and the agent fills them correctly
Scoped access
Each tool gets its own credentials and its own rate limit
$ npm install @kestrel/tools
Safe by construction
An agent cannot do what its manifest does not allow. That is not a setting, it is how the runtime is built.
Permission manifests
Every tool, every scope, declared before the first run
Approval gates
Mark a step and a person has to say yes before it continues
Spend limits
A budget per agent, per run and per month, enforced by the scheduler
Developer experience
Everything the canvas does, the CLI does too.
Scaffold
kestrel init writes an agent with a goal, a model and two tools
Iterate
kestrel run streams every step to your terminal while you edit
Ship
kestrel deploy puts it on a schedule and hands you a URL
Live right now
Numbers from the scheduler, as of this second. Nobody rounds them.
- agents running
- 23,118
- tasks today
- 1.9M
- median step
- 312ms
- average task cost
- ₹0.04
Plans
Start free. Pay when the agents are earning their keep.
Sandbox
₹0/month
Billed monthly
- 5 agents
- 10,000 tasks
- Community support
Team
₹7,900/month
Billed monthly
- 50 agents
- 500,000 tasks
- Shared memory
- Priority support
Enterprise
Billed monthly
- Contact sales
- Unlimited agents
- Dedicated runtime
- SSO and audit log
- Named engineer
- Sand & Petrol
- grounded · warm · established
- Fraunces / Instrument Sans
- pill corners · lifted elevation
- Warm
- rise · 0.6s · business, clinical, professional