Documentation
What Keiki is, and what you can build with it.
Keiki helps you put AI agents into the conversations your customers and teammates already use: SMS, iMessage, WhatsApp, Slack, Telegram and email. You describe one agent once, and Keiki runs it everywhere.
You write down what the agent knows and what it can do. Keiki handles the rest: delivering messages on every channel, remembering each conversation, running tools, browsing the web, running code, taking payments, and handing a conversation to a person when it should. Adding a channel later does not mean rebuilding the agent.
These guides show you how to set each part up. Every page is also published as markdown at
/docs/<name>.md if you would rather read raw text.
Build an agent
- Cloud-served agents: write your agent's config, including its instructions, tools, plugins and channels.
- SDK reference: everything
@onkeiki/agentsexports, from agents, tools and alerts to plugins, capability manifests and service adapters. - Turn harnesses: choose where the agent's turn loop runs.
- External harness: run the turn loop in your app with signed async requests from Keiki.
- Agent swarms: put many agents on one goal, and watch and stop them.
- Rich cards: render swipeable card carousels into one media message, enable them on an agent, and wire the agent into them.
Connect your app
- Agents as models: call any agent from the OpenAI or Anthropic SDKs.
- Partner deployments: deploy an agent for each of your users from your own backend, over REST or MCP.
- Identity and permissions: tell the agent who is talking in your terms, and have every tool call checked against their scopes.
- Governed actions: preview, approve, run and audit the tool calls that matter, with no gap between the card and the call.
Slack
- Slack agents: when the agent answers, who it thinks is talking, bursts of messages, and how to stop it.
- Migrate a bespoke Slack agent: replace the runtime you built around your Slack bot with one Keiki agent per customer organization.
- Example: a billing agent per organization: the migration guide made concrete — one deployment request, two gated tools, an identity webhook.
Capabilities
- Capabilities: package a new ability and publish it for others.
- Publishing a capability: have your coding agent build one, then submit it to the marketplace.
Evaluate
- Eval Studio: replay saved cases to check how the agent behaves and pick the best model.
- Replay: test a prompt change against real turns from production.
- External-channel evals: grade cases on what a real contact actually received.
Operate and secure
- Agent wallets: give an agent a spending policy in front of your own payment account.
- Enterprise SAML SSO: let your team sign in through your own identity provider.