Raycast expanded its desktop launcher into a more capable AI workspace with its September 2026 v2.3 release. The update adds Raycast Automations for scheduled prompts, Projects for persistent context, and an AI Chat that can plan multistep work, use extensions, run code, check results, and retry when a step fails.
These features can save time, but they also introduce practical constraints. Automations run on your device rather than in a continuously available cloud service, tool approvals can pause unattended work, and complex runs consume more AI credits than a short answer. This guide explains what is actually available, how to set it up, and where human review still matters.
What are Raycast Automations?
Raycast Automations run an AI prompt on a schedule and place the result in AI Chat. Create one in Settings > AI > Automations, choose a model and destination chat, write the prompt, then select a one-time, hourly, or daily schedule. Your computer must be awake with Raycast open, and tool approvals can pause a run.
An automation is essentially a scheduled AI Chat message. It can use the same AI Extensions, Model Context Protocol (MCP) servers, and Skills that you can mention in an ordinary conversation. That makes it suitable for recurring summaries, reviews, and preparation work, but it does not turn Raycast into an always-on cloud agent.
What changed in Raycast v2.3?
Raycast's official v2.3 notes describe three connected additions:
- Agentic AI Chat: Raycast can break a task into steps, use installed extensions, write and execute code, inspect the output, and try another approach after a failure. It pauses when it needs a decision.
- Projects: Related conversations can share their own memory. A Project can also point to a working directory, such as a notes vault, codebase, or document folder.
- Automations: Prompts can run once or on a recurring schedule, with each result delivered to a new or existing chat.
Raycast also lets users on a paid plan connect supported ChatGPT or Claude subscriptions. Bring-your-own API keys, custom providers, and local models remain alternatives, but the Raycast documentation says those options also require a paid Raycast plan.
How to create a Raycast Automation
- Open the automation settings. Go to Settings > AI > Automations, or search for Create Automation in Raycast's Root Search.
- Choose manual or AI-assisted setup. Manual setup exposes each field directly. The AI-assisted option can draft the schedule, destination, and prompt, but asks for confirmation before saving it.
- Name the automation and choose a model. Match the model to the work. A daily summary rarely needs the most expensive frontier model.
- Select the destination. A new chat keeps each result separate. An existing chat preserves earlier runs as context, which is useful for cumulative tracking but can increase context and credit use.
- Write a bounded prompt. Specify what to inspect, what output to produce, what not to change, and which actions require approval.
- Set the schedule. The interface supports one-time, hourly, and daily patterns. Daily schedules can be limited to selected weekdays; custom schedules may appear when AI creates a more complex cron expression.
- Save, then test immediately. Use Run Automation Now, inspect the result, and correct the prompt before relying on the schedule.
A safe copy-ready automation prompt
This example prepares a weekday project brief without authorizing changes:
Every weekday at 8:30 AM, review the project notes and open tasks
in the sources I explicitly provide. Create a concise briefing with:
1. deadlines in the next seven days,
2. blocked items and their owners,
3. decisions recorded since the previous run, and
4. three recommended priorities for today.
Do not send messages, edit files, close tasks, run destructive commands,
or access another folder. Link each statement to its source when possible.
If a source is unavailable or the evidence conflicts, label that item
for manual review instead of guessing.The prompt works best in a Project connected only to the relevant working directory. Replace vague phrases such as “check everything” with named sources and a defined output. If an extension can change external data, keep its approval requirement enabled until repeated test runs prove the workflow is predictable.
Schedules and local-device limits
| Behavior | What to expect |
|---|---|
| One-time schedule | Runs at the selected date and time, then removes itself. |
| Hourly schedule | Runs every 1 to 24 hours and can be limited to selected days. |
| Daily schedule | Runs at a chosen local time on selected days. |
| Device availability | The computer must be awake and Raycast must be running for an on-time run. |
| Missed runs | Raycast runs once at the next launch; multiple missed occurrences collapse into one run. |
| Busy destination chat | The automation waits for the chat to become idle, with due occurrences combined rather than queued repeatedly. |
| Approval or question | The run pauses until the user responds. |
| Storage and sync | Automations stay on the device where they were created, do not use Cloud Sync, and are limited to 20 per device. |
These rules make Raycast Automations best for work that can tolerate a delayed run. Do not use them as the only alerting system for an incident, a legal deadline, a payment, or another time-critical process.
How Projects change ongoing work
A Raycast Project groups conversations around one subject and gives them shared memory. You can attach a working directory so the agent knows where it may work with files. This can reduce repetitive setup for activities such as maintaining a knowledge base or preparing recurring reports.
Persistent context also deserves review. Old instructions and past decisions can become stale. Keep Projects narrow, remove irrelevant context, and periodically start a clean conversation when the accumulated history begins to distort the current task. A directory boundary is useful, but it should not be treated as a substitute for backups, file permissions, and tool confirmations.
Raycast AI pricing and credits
Raycast switched its hosted AI from request-rate limits to usage-based credits in September 2026. The official billing documentation lists these individual plans as of September 13, 2026:
| Plan | Monthly billing | Annual billing | Monthly AI credits |
|---|---|---|---|
| Free | $0 | $0 | No Raycast AI credits |
| Pro | $10 | $96 per year | 500 |
| Pro+ | $20 | $192 per year | 3,000 |
| Max | $50 | $480 per year | 7,500 |
Pro+ replaces the former Advanced AI add-on. Pro+ and Max include top-tier models, while Pro has a smaller model selection. Monthly credits roll over for one additional billing cycle and expire 60 days after they are granted; purchased top-ups have a longer stated expiry. Prices and allowances can change, so verify the live billing page before subscribing.
Credits measure usage rather than a fixed number of messages. Long context, attachments, tool calls, regenerated answers, and a more expensive model can all raise the cost of one run. Raycast's own documentation warns that an automation using several tools costs more than a short prompt.
Five checks before letting an automation run
- Test with non-sensitive data. Confirm the prompt handles missing files, conflicting dates, empty results, and tool errors.
- Limit the sources. Give a Project only the folder and services needed for the task. Avoid broad home-directory or account access.
- Preserve approval boundaries. Require confirmation for sending, deleting, publishing, purchasing, merging, or changing records.
- Check the destination chat. Results sent to an archived chat are skipped until it is unarchived, and a deleted destination causes the next run to start a replacement chat.
- Monitor both results and credits. Review early runs for accuracy and inspect Settings > Account or the
/usagecommand to understand cost.
If you are designing a more complex agent workflow, ToolNexa's n8n Assistant guide includes additional checks for credentials, idempotency, failure paths, and production rollout.
Frequently asked questions
Do Raycast Automations run when the computer is asleep?
No. Raycast checks schedules while it is running, and the computer must be awake for an on-time run. A missed schedule runs once after Raycast next launches rather than replaying every missed occurrence.
Can a Raycast Automation use extensions and MCP servers?
Yes. Raycast says each automation run is a normal AI Chat message, so the prompt can mention AI Extensions, MCP servers, and Skills. Their tools follow the destination chat's normal permission settings.
Are Raycast Automations available on the free plan?
Raycast documents Automations as an AI feature, while Raycast Free does not include AI. A paid plan is therefore required for the documented automation workflow.
How much does one automation run cost?
There is no fixed per-run credit amount. Cost depends on the selected model, the input and output length, attachments, context, and tool calls. You can inspect usage for individual responses and the account's remaining balance.
Can automations create or edit other automations?
No. Raycast's current documentation says an automation cannot create or change another automation while it runs.
Conclusion
Raycast Automations are most useful for repeatable preparation and review tasks: daily briefs, inbox triage suggestions, project summaries, and scheduled checks whose results can wait in chat. Projects give that work reusable context, while the expanded AI Chat can handle more of the execution loop.
The safe starting point is narrow: use a low-risk source set, request a draft or report rather than an irreversible action, test with Run Automation Now, keep approvals on, and watch both output quality and credit use. The feature can remove repetitive prompting, but responsibility for scope, accuracy, and consequential actions still belongs to the user.