Features
The platform behind every shipped PR
cyql is more than a model call. It's the engine, the isolation layer, and the review loop that turn autonomous generations into pull requests you can trust.
The plan → ship loop
A durable workflow engine drives every task from draft to merged PR, resuming exactly where it left off after any failure.
Runtime isolation
Each agent runs in a locked-down, isolated pod with egress firewalled to your allow-list. Cloud metadata and private ranges are blocked by default.
Parallel agents
Independent subtasks run at the same time across an autoscaling worker pool, so big changes land in minutes, not hours.
Built-in review
Every diff is checked against your conventions and test suite before it reaches you — agents review agents, you review the result.
Short-lived credentials
Workers never hold long-lived secrets. The orchestrator mints per-task, scope-limited tokens that expire when the job ends.
Full execution traces
Replay every decision, prompt, tool call, and file change. Cluster recurring failures and turn them into guardrails.
Connects to your stack
GitHub, GitLab, Linear, Jira, and Slack out of the box. Trigger runs from a PR comment, an issue label, or the API.
Spend controls
Per-project budgets, model routing, and hard caps. See exactly what each task cost before and after it runs.
Workflow engine
A durable loop that never loses its place
Every task is a durable workflow. If an agent crashes, a test flakes, or a node disappears, the run resumes from the exact step it was on — not from the beginning. State lives in the engine, so long-running, multi-step changes are safe by construction.
- Exactly-once step execution with full resumability
- Automatic replan on failure instead of a dead end
- Human approval gates on the decisions that matter
# workflow engine
›exactly-once step execution with full resumability
›automatic replan on failure instead of a dead end
›human approval gates on the decisions that matter
✓ ready
Security model
Isolation, least privilege, and zero standing secrets
Each agent runs in its own pod on a dedicated, tainted node pool, with egress firewalled to an allow-list. The cloud metadata endpoint and private ranges are blocked outright. Workers carry no long-lived secrets — only a bootstrap token that can request scoped, short-lived credentials per task.
- Pod-per-task isolation, no shared state
- Egress allow-list; metadata + private ranges blocked
- Per-task credentials that expire on exit
# security model
›pod-per-task isolation, no shared state
›egress allow-list; metadata + private ranges blocked
›per-task credentials that expire on exit
✓ ready
Parallelism
Fan out across an autoscaling fleet
Independent subtasks run in parallel across a worker pool that scales on queue depth. A large refactor that touches dozens of files lands in minutes because the work is spread across many agents — then the pool scales back to zero.
- Queue-driven autoscaling worker pool
- Dependency-aware scheduling of subtasks
- Per-project budgets and hard spend caps
# parallelism
›queue-driven autoscaling worker pool
›dependency-aware scheduling of subtasks
›per-project budgets and hard spend caps
✓ ready
Ship your backlog while you sleep.
Start free in minutes. Connect a repo, describe a task, and watch a reviewed pull request show up.
No credit card · Free forever for individuals