AI workers · your tools, your rules
Tell it. Watch it ship.
Crewshal gives your organization AI workers with a role, memory and a manager. Give one a task, keep your corrections, and continue through Claude Code or Codex—with the work and its evidence in one place.
Replay a past PR and compare the resultYou · WhatsApp
The payments list hides the drafts from the new invoice import. Can Ada take NW-142?Native engines and connected workspaces
Availability depends on your workspace setup.
Engines and models
- Claude Code
- Codex
- Gemini
Workspace connectors
- GitHub
- GitLab
- Slack
- Linear
- Notion
- Jira
- Confluence
How it works
From a message to a merged pull request.
No new tool to learn. You ask the way you would ask a colleague, and the work arrives the way a colleague would deliver it.
- 01
Give it the task and context
Start with the manager conversation, a connected ticket or a project brief. Keep the goal, your corrections and the next decision together.
- One conversation, shared context
- Connected routes depend on setup
- 02
It works like an engineer
It reads the code and the history, plans, then hands the implementation to Claude Code or Codex running natively in a sandbox, on your own subscriptions.
- One round, full context
- Your conventions, your rules
- 03
Inspect the result, then ship
Review the checks, changes and open questions. Your approval rules decide what can proceed; the work and its evidence stay together.
- Checks and changes to inspect
- Your approval rules
Your work, carried forward
Change the model. Keep the context.
Your brief, decisions and corrections belong to the work. Start with Claude Code, continue in Codex, and pick up with the context you already built.
- 01Ask
Tell it what needs doing.
The task, accepted constraints and open questions stay together from the first brief.
- 02Correct
Teach it your way.
A decision such as “exclude account IDs” becomes part of what the next round can inspect.
- 03Switch
Choose the engine for the next step.
Move from Claude Code to Codex when the work calls for another native tool or account.
- 04Inspect
Open the result. Check the work.
The handoff ends with evidence you can review before deciding what happens next.
Add CSV export to billing
- Correction
- Exclude account IDs
- Route
- Claude Code Codex
- Inspect
- Changes Checks Questions
Command Center
See the work. Follow the evidence.
Start at the fleet, open a worker, then step inside a delegation. Changes, checks and questions remain visible in the same trail.
- 01FleetWorkers, current work and operating signals
- 02WorkerRole, manager, memory and assigned items
- 03DelegationBrief, changes, checks and open questions Selected
Evidence stays attached to the work you are inspecting.
Safe by default
It asks when it matters.
Every action goes through one gateway that authorizes it, classifies its risk, budgets it and writes it down. Policies decide what a worker may do alone and what waits for you.
| Risk | For example | What the worker does |
|---|---|---|
| None | Read, search, analyze | Acts on its own |
| Low | Edit code, commit, run tests, push a feature branch, write documents and messages inside the platform | Acts on its own |
| Medium, inside the platform | Install packages, save business records | Acts, then tells you |
| Medium, outside the platform | Open a pull request, create or comment on a ticket in your tracker | Waits for your approval |
| High | Push to the main branch, delete data | Waits for your approval |
| Always-ask list | Merge, deploy, Slack or Teams messages, new integrations, hiring, reorganizing, spend above your limit | Always waits for your approval |
| Critical | Force-push, merge into a production repository's main branch | Blocked unless a policy asks for approval |
These are the defaults for a new organization in Auto mode. Your policies, your settings and a level that a worker earns can change them.
Credential access is scoped, and authorized actions leave an audit trail.
The experience behind Crewshal
Previous employers and consulting clients of SASID’s founder. See SASID.ai
Career experience; no Crewshal customer relationship or endorsement is implied.
Historical replay
A real pull request. Your own verdict.
Explore a recorded h3 middleware fix, then compare the candidate with the original result. Inspect the checks, elapsed time and cost estimates before you choose.
Replay the past PRRecorded open-source benchmark · h3js/h3.
Edited actual product screens from the fictional Cobalt demo.
- Reference
- h3 route middleware
- Shown
- Task, candidate and checks
- Separated
- Human effort and charges
- Decision
- Compare before you continue
Put a worker on it.
Give your first worker a clear task, your repository context and your rules. Review what comes back.