Admins decide what agents can do
Every agent role runs under permissions your admins control — what tools it can call, what files it can touch, what credentials it can ask for. Project owners can narrow further; they can never widen.
Crewstack turns a Work Item into a pull request. An Orchestrator agent refines the scope with you, delegates execution to specialist agents in real sandboxes with real GitHub access, and brings the result back for your review.
Dashboard feels slow on load. Auth seems to be the bottleneck — needs to be faster without breaking the existing client.
Scope
Trace the four sequential auth calls fired on /dashboard mount — session, profile, prefs, and feature flags. Coalesce them into a single batched endpoint behind /session/bootstrap that preserves the public client API surface.
To be defined with the Planner.
Acceptance
Planner
Refining scope with you
Files changed
How it works
The same loop a senior engineer runs — scope, delegate, review, ship — running on a contracted runtime that your admins control.
Open a Work Item, paste a one-line ask. The Orchestrator agent reads your repo, asks the questions a teammate would, and writes the scope back into the Work Item Document — with the right files, the right references, the right acceptance criteria.
Scope
Trace the four sequential auth calls on /dashboard mount. Coalesce into a single batched endpoint behind /session/bootstrap. Keep the existing client API.
Acceptance
Specialist roles — Planner, Code Writer, Code Reviewer, Docs Writer — pick up the work in real sandboxes with real GitHub access. Each role runs under a Policy Bundle your admins control. The delegation tree is visible while it runs.
Crewstack opens the pull request on your repo and links it to the Work Item. You review on GitHub the way you always have — comments, requested changes, suggested edits — and the agent responds.
crewstack/webPR-482 opened byCode Writer
Merge the PR. Crewstack flips the Work Item to done automatically and the Orchestrator closes anything outstanding. The whole loop is recorded — tool calls, credential releases, and every state transition.
Recorded: 47 tool calls · 3 credential releases · 9 state transitions
The substrate
The competitors orchestrate sessions in vendor-owned workspaces with opaque runtimes. Crewstack is built contract-first so a CTO can actually deploy it.
Every agent role runs under permissions your admins control — what tools it can call, what files it can touch, what credentials it can ask for. Project owners can narrow further; they can never widen.
Agents run inside real sandboxed workspaces with your code, your dependencies, and your tooling installed. Same environment your engineers would use. Not a chat surface pretending to have a terminal.
Secrets are encrypted at rest, fetched just-in-time inside the sandbox, and never appear in any prompt sent to a model. Every release is logged with the agent, the tool, and the reason.
When a Work Item produces a pull request, the Work Item and the PR stay linked. Merging the PR flips the Work Item to done automatically. Your tracking surface and your engineering surface stop drifting apart.
External context
Agents are only as good as the context they can pull. 17 providers — GitHub, Linear, Slack, Jira, Sentry, GitLab, and more — wire them into your real stack as a unified set of tools. Events route back into the Work Item automatically; agents react to what's happening across your stack.
Your code, your docs, your tickets, your PDFs — searchable by the agents through a permission-aware retrieval layer. Agents only see what they're allowed to see, scoped per query, per role, per project.
Skills are reusable instruction fragments scoped to your org, workspace, project, or a single agent. Use them to teach agents your patterns, your style, your guardrails — without forking templates or hardcoding policies.
Agents produce research notes, design memos, and summaries as they work. Those land as searchable artifacts attached to the Work Item. Tomorrow's planning question can answer itself from yesterday's exploration.
Who this is for
Teams of 5–50 shipping product weekly who want a sharper answer to “what should we delegate?” than a chat tab.
Solo or small founding teams who want a senior engineer on tap for work that isn't worth a senior hire yet.
Leaders who need an audit trail, an admin-controlled policy surface, and a runtime that doesn't change under them.
Not built for: non-technical buyers, chat-assistant shoppers, vibe-coders looking for an AI pair, or anyone shopping for a Jira replacement.
Start
We'll set up a sandbox against your repo and walk through a Work Item end to end. About 30 minutes. Goes directly to the founder.