KernelAgentOpen the app
Open-source agent runtime

Run AI agents like operating system processes.

KernelAgent splits a task into agent processes, schedules them, and lets them act only through permission-checked syscalls in a sandbox. Every step is logged, so you can see exactly what your agents did, what it cost, and why.

Open the app →View the code

Bring your own Claude or OpenAI API key. Nothing to install.

kernelagent · job research-pipeline
PIDROLESTATETOKENS
p-1plannerTERMINATED4.1k
p-2researcherRUNNING12.8k
p-3coderBLOCKED9.3k
p-4reviewerREADY0
#482 SYSCALL p-2 READ_FILE notes.md · allowed
#483 LLM_CALL p-2 claude-sonnet-5 · 88% cached
#484 SYSCALL p-3 EXEC pytest · waiting for approval
How it works

An operating system for agents

The same ideas that keep programs from stepping on each other keep agents in check: processes, a scheduler, permissions, isolation, and an audit log.

  1. 01
    Job

    You describe a task. It becomes a job: one or more agent processes, with dependencies between them.

  2. 02
    Scheduler

    A priority queue with aging runs ready processes under a concurrency cap and a token rate limit.

  3. 03
    Syscalls

    Agents act only through syscalls: read, write, exec, spawn, send. Each one is checked against the process's capabilities.

  4. 04
    Sandbox

    Code and files run in an isolated sandbox per process. Risky commands can wait for your approval.

  5. 05
    Event log

    Every state change, model call, and syscall is appended to a log. Replay it to rebuild any moment.

What you get

See how your agents behave

Chat

Ask for something. Watch each step the agent takes, the files it writes, and what it cost.

Process console

A live table like top, a task graph, IPC messages, sandboxes, and traces you can rewind.

Metrics

Tokens, cost, cache savings, and errors over time, split by model. Compare Claude and GPT on the same work.

Your account

Chats and agent memory are saved to your account, so your agents remember context on any device. Your API key stays in your browser.

Before you start

Bring your own API key

Agents run on your key, so you pay your provider directly and nobody else's usage lands on your bill. Add one key or both.

ClaudeAnthropicGet a key
  1. Open the Anthropic Console and add billing.
  2. Create a key under Settings → API keys.
  3. Paste it on the Connect page after you sign in.
ChatGPT modelsOpenAIGet a key
  1. Open the OpenAI platform and add billing.
  2. Create a secret key under API keys.
  3. Paste it on the Connect page after you sign in.
Why not sign in with Claude.ai or ChatGPT?
Consumer subscriptions do not include API access, and neither company lets other apps run models on your chat account. An API key is the supported way.
Where your key goes
It stays in your browser. Each job sends it to the API, which holds it in memory until the job ends. It never reaches the database or the event log.
Coming next

Connectors

Connect GitHub, Google Drive, or Slack through MCP and let agents use them, under the same capability checks, approvals, and audit log as every other syscall. Watch how Claude and GPT agents handle real tools side by side.

TypeScriptNext.jsFastifySQLiteAnthropic SDKOpenAI SDKE2B sandboxesOpenTelemetryAuth.js
Try it with your own agents
Sign in, add a key, and give your agents a task.
Open the app →