公司即代码。面向智能体的后端框架和服务器。
项目就是一个 git 仓库。kortix.yaml 声明其代理、连接器、触发器和规则。Kortix 在独立的云计算机上运行每个会话，你可使用 @kortix/sdk 和 @kortix/ui 构建自己的体验。
One block per agent. It gets the tools, connectors and skills the block lists, and nothing else.
- # Engineering
- eng-triage:
- description: Triages every new GitHub issue.
- prompt_file: agents/engineering/triage.md
- tools: [read, grep, glob, web_search]
- connectors: [github, kortix_slack]
- skills: [triage-issue]
- workspace: { project_repository: true }
eng-triage
对每个新的 GitHub issue 进行分诊。
- read
- grep
- glob
- web_search
- github
- kortix_slack
- triage-issue
/workspace 中的仓库，位于其独立分支
一个仓库中的一个文件。 Kortix runs what it declares.
四个想法。 The whole framework, in four parts.
A project is a git repoYour company, versioned.
- Kortix hosts it, or link GitHub
- Agents, skills and memory are files
- Every change is a commit
kortix.yaml declares itOne file, checked before it runs.
- Agents, connectors, triggers, rules
- Access is deny-by-default
- kortix validate on ship and merge
每个会话一台计算机Where the agent works and the rules hold.
- pi runs every new session
- Tool rules checked before each call
- Its own branch in /workspace
Your own experienceThe same backend, from code.
- @kortix/sdk: one typed client
- @kortix/ui: the web app’s components
- A CLI and a REST API
工作原理。One part at a time.
Agents
An agent is a block in kortix.yaml and a Markdown prompt. The block names its model, tools, connectors and skills.
Access is deny-by-default. An agent gets no connector, secret or skill the block does not list.
kortix.yaml
# Marketing marketing: description: Drafts LinkedIn posts from shipped work. prompt_file: agents/marketing/content.md tools: [read, web_search] connectors: [linkedin, kortix_slack] skills: [write-post]agents/marketing/content.md
You write Arcadia LinkedIn posts from work the team shipped.Follow the write-post skill and the tone in the company memory.Every post waits for a person to approve it.架构。 One API. A computer per session.
Clients
People and code reach Kortix the same way.
- Web app
- Desktop
- Mobile
- CLI
- Slack
- Microsoft Teams
- Your app
Kortix API
Decides who may run which agent, starts sessions, holds connector logins and change requests.
Project repository
kortix.yaml, agents, skills and memory. Kortix hosts it, or you link GitHub.
会话计算机
One per session. pi works here, on the session’s own branch.
LLM gateway
Every model call. Kortix-managed models or your own keys.
Kortix Drive
Shared folders and company memory, open at /drives.
框架问题。Answered from the docs.
Anything else, write to the team.
Do I have to write YAML?
No. People manage agents, connectors and triggers under Customize in the web app, and Create in chat writes a new agent with you. Those edits land in the same kortix.yaml.
Can I keep the repository on GitHub?
Yes. With a GitHub origin, kortix ship links that repository, and plain git push and git pull work. Without one, Kortix creates and hosts the repository.
What does kortix validate check?
Every key against the schema, then the files around it: prompt files that do not exist, skill folders without a SKILL.md, and connector or secret names the file does not declare. It exits 1 on an error. kortix ship runs the same checks, and a merge with an invalid kortix.yaml is refused.
When does a change to kortix.yaml take effect?
When it lands on the base branch, through a merge or a push. New sessions start with it. A running session picks it up at the start of its next turn.
What runs inside a session?
pi, an open-source agent harness. Every new session runs it on its own cloud computer, and it calls models only through the Kortix LLM gateway.
Where do secret values live?
In Kortix, encrypted. kortix.yaml lists only the names a project needs. No page, command or API shows a stored value again.
Can I run it myself?
Yes. Kortix is open source. Run it on Kortix Cloud, in your VPC, or on your own on-prem network.