Flows
Templates and AI drafts
Start a LoCrew flow from one of five built-in templates, map its roles to your agents, import and export templates as JSON, or describe a flow and let one of your agents draft it.
You do not have to start from a blank canvas. New flow opens a gallery with a blank flow, five built-in templates, Describe it, and Import from file….
#Built-in templates
| Template | Starts | What it does | Roles |
|---|---|---|---|
| Research brief | Started by you | One agent researches a topic, another turns the findings into a short brief, and you approve it. | Researcher, Writer |
| Review and fix | /run review | A reviewer decides; if changes are needed, a second agent makes them. Type /run review followed by a path, branch or PR. | Reviewer, Fixer |
| Two reviews, one verdict | Started by you | Security and quality reviews run side by side; a lead combines them into one verdict. | Security reviewer, Quality reviewer, Lead |
| Daily digest | On a schedule | Every morning, one agent gathers what changed and another writes it up. The schedule starts off; switch it on when ready. | Collector, Writer |
| Hacker News on a topic | /run hn | Fetches the top Hacker News stories on a topic and has an agent summarise them. | Writer |
How some of them are built:
- Review and fix: the Review step replies in JSON with
verdict(approve or changes) andsummary. A condition, Changes needed?, checks whetherverdictis changes; Yes leads to Fix. - Two reviews, one verdict: two agent steps run in parallel, a Join branches step waits for both, then the lead writes the verdict.
- Hacker News on a topic: an HTTP step calls
hn.algolia.comwith your topic and keeps thehitslist.
Warning
Start Hacker News on a topic with /run hn <topic>, not the editor's Run button. It reads {{start.json.prompt}}, which only exists for runs started from a conversation. This public API was not contacted in the project's tests.
#Set up a template
- Choose a template. The Set up form opens.
- Give the flow a Name.
- Under Who does what, choose one of your agents for each role. One agent can fill several roles. An agent whose name matches a role is preselected.
- Read Before you run it: for example, that the schedule starts switched off, which
/runcommand it uses (a number is added if that command is taken, such asreview-2), and which hosts it calls. - Press Create flow.
#Import and export
Export as template… (in a flow card's ⋯ menu) saves the flow as a .locrew-flow.json file. Each agent becomes a role named after it. Anything specific to your workspace is removed, and a message lists what was left out:
- Send message steps lose their destination.
/runis allowed in every conversation.- HTTP credentials are left out.
- Local network and
http://access are switched off. - Text that looks like a key or token is replaced with
[redacted].
The flow-wide instructions are not exported; the run time limit is.
Import from file… reads a template up to 512 KB. LoCrew checks it strictly: a template may not name agents (only roles), conversations or credentials, may not allow private network or http:// access, and may not carry anything that looks like a secret. A valid file opens the set-up form.
There is no shared template library in the app: templates you export are files.
Note
A flow that uses a sends back connection cannot be exported in this version: the template format has no way to describe it.
#Describe a new flow
Describe new flow (or Describe it in the gallery) lets one of your agents draft a flow from a description.
- Write what the flow should do, up to 4,000 characters. Example prompts are offered, such as Review a pull request for security and quality in parallel, then give one verdict.
- Pick the agent that drafts it. Model agents are listed first.
- Press Draft the flow. You can leave the page; drafting carries on, and a toast says when the draft is ready.
- The draft opens in the set-up form with Drafted by …. Read every step below: a draft can be wrong, and nothing in it has run. Review every step, map the roles, and press Create flow.
Things to know:
- The agent only plans the flow, under its own permissions, and it may use a turn of its usage. Drafting costs one or two turns.
- Nothing runs, and nothing is saved until you press Create flow.
- The draft is checked like an imported template. If the agent's reply still has problems after two tries, LoCrew shows them and you can try again.
- No real model has drafted a flow in the project's tests, so draft quality is unknown.