Run agents

Interact with agents through Slack, the CLI, and your own code.

Agents can be used in a variety of ways:

Surface	Status	Best for
Slack	Available	The primary way Keystroke users interact with agents today, directly in Slack
CLI	Available	Testing and debugging agents during development
HTTP API	Available	Use agents via API and integrate them into your own services
Workflow steps	Available	Use agents as steps in deterministic workflows
Triggers	Available	Run an agent automatically on a schedule, webhook, or poll
In-app chat	Available	Chat with any deployed agent from its page in the Keystroke web app
External channels	Coming soon	Use agents in tools like Microsoft Teams, Telegram, Linear, and more
MCP server	Coming soon	Expose agents as tools to other MCP clients

Use agents in Slack

Slack is how most teams use Keystroke agents today. Connect the Keystroke Slack App (keystroke connect slack --kind keystroke; catalog slug slackbot — not a Personal Slack Account user token), bind Slack channels to your agent, and teammates can message it like any coworker. The gateway turns each message into an agent session and posts the reply back as the bot.

See external channels for the three Slack connection kinds, connecting, binding channels, listen modes, and channel sessions.

Use agents in the web app

Every deployed agent has a chat surface on its page in the web app: open Agents, select the agent, and message it directly. Each conversation is an agent session — start a new one with New chat, and revisit past sessions from the history rail on the right. Sessions started in the app appear in History alongside every other surface.

This is where teammates who never touch the CLI use what you build. Chatting with a deployed agent is different from the platform agent, the built-in builder chat that creates and edits agents rather than running them.

Use agents from the CLI

The CLI is the most practical way to run an agent while you build and test it. Deploy your project, then prompt the agent against the deployed project:

keystroke deploy --project <slug>   # or --filter agents/support
keystroke agents prompt support \
  --message "What is Keystroke?" \
  --wait \
  --timeout 120

The support value is the agent slug from defineAgent(). With --wait, the response includes the completed session and its messages. Without it, the command returns the queued sessionId and runId immediately.

--wait is bounded and defaults to 120 seconds for agents. Choose another duration with --timeout <seconds>. A timeout means the session is still active, not failed. Continue waiting with:

keystroke agents sessions get support <session-id> --wait --timeout 120

When you call an agent from code and need a typed object back, pass an outputSchema and read result.output — see structured output.

Continue the same session by passing its ID:

keystroke agents prompt support \
  --session-id <session-id> \
  --message "Now turn that into a Slack reply" \
  --wait

Agent runtime commands run against your deployed cloud project:

keystroke agents prompt support --message "Prompt the deployed agent" --wait

keystroke agents list is platform/org-level and lists agents in the active organization or project. Prompt and session list/get commands run against the resolved project target. Session cancellation is platform-backed today.

List recent sessions for one agent:

keystroke agents sessions list support

Filter while debugging:

keystroke agents sessions list support --status failed
keystroke agents sessions list support --source gateway

Get details — start lean, then expand:

keystroke agents sessions get support <session-id> --wait --timeout 120
keystroke agents sessions get support <session-id> --summary
keystroke agents sessions get support <session-id> --include messages,trace

Use --include gateway,messages,events,trace when you need all available session detail. See the CLI reference for the full command list.

Cancel a running cloud session by ID:

keystroke agents sessions cancel <session-id>

Cancellation is useful when an agent is stuck in a long-running tool call or no longer needs to finish. This command uses the active platform context today, so use it for deployed/cloud sessions.

Use agents in workflow steps

Agents can be called from workflows and actions with .prompt():

const result = await supportAgent.prompt({
  message: "Draft a reply for this customer thread.",
});

When called inside a workflow, the agent prompt is part of the workflow execution and also creates an agent session. Inspect both the workflow run and the agent run when debugging handoff behavior.

Run agents from triggers

A trigger attaches an event source (a schedule, a webhook, or a poll) to an agent, so the agent runs on its own when the source fires instead of waiting for someone to prompt it. Triggers live in src/triggers/; a source exposes .attach(), where you pass the agent and the prompt to run.

src/triggers/morning-check.ts

import { defineCronSource } from "@keystrokehq/keystroke/trigger";
import support from "../agents/support";

export default defineCronSource({
  slug: "morning-check",
  name: "Morning check",
  description: "Asks support to summarize overnight issues each morning.",
  schedule: "0 9 * * *",
}).attach({
  agent: support,
  prompt: () => "Summarize anything that needs attention this morning.",
});

Webhook and poll sources pass their payload to prompt, so the agent can act on the event that fired it:

src/triggers/new-inbox.ts

}).attach({
  agent: support,
  prompt: (payload) => `Summarize this new inbox item: ${JSON.stringify(payload)}`,
});

Each fire starts a new agent session, just like any other surface. Inspect them in History (filter Type to Agent) or from the CLI:

keystroke triggers runs list morning-check --agent support

See triggers for the full source-and-attach model, payload filters, and transforms.

Agents can also schedule their own runs at runtime instead of waiting for a code trigger. See ephemeral triggers.

Use agents via API

You can also prompt a deployed agent over HTTP. The route is asynchronous: it enqueues the prompt and returns 202 with a sessionId and runId to inspect later, rather than waiting for the agent to finish.

curl -X POST https://<your-project>/api/projects/<project-id>/agents/support \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer <org-api-key>" \
  -d '{ "message": "Summarize our refund policy" }'

Continue an existing session by including its sessionId in the body alongside the new message; send only the new message, not the prior turns. Inspect the result later in History or with keystroke agents sessions get.

This is mainly useful for wiring agents into your own services. For most internal tooling, Slack and the CLI cover what you need.

Review agent runs

Every surface that runs an agent creates an agent session you can review later. Open History in the web app and filter Type to Agent. The detail panel shows messages, tool calls, metadata, trace data, timing, and errors.

From the CLI, use session commands when you want to inspect runs while debugging:

keystroke agents sessions list support
keystroke agents sessions get support <session-id> --wait --timeout 120
keystroke agents sessions get support <session-id> --summary
keystroke agents sessions get support <session-id> --include messages,events,trace

See agent runs for the full run history view.

Next steps

Triggers

Run agents on a schedule, webhook, or poll.

External channels

Connect Slack and route channel messages to agents.

Agent runs

Inspect messages, tool calls, errors, and traces.

Deploy a project

Ship agent changes to the platform.