Five steps. Every change earns its place.
Section titled “Five steps. Every change earns its place.”Each agent runs in a mode — Research, Plan, Implement, Validate, Review — and the harness runs the matching skill behind it. The same curated playbook every time, no ad-hoc prompts.
Map the codebase first
Agents read the repo and write findings to a shared thoughts repo — context that outlives the session and travels across the team.
- Parallel sub-agents sweep the code
- Findings persist as reviewable docs
- Cited to exact files and lines
Plans before code
Every change starts as a phased plan with explicit success criteria — reviewed and refined before a line is generated. Engineering rigor, not vibes.
- Phased, with success criteria
- Reviewed before any code
- The spec the team signs off on
Parallel worktrees
Each agent runs in its own isolated git worktree and branch, so the team ships many streams at once without stepping on each other.
- Isolated branch per agent
- Many streams in parallel
- A real shell in every worktree
Checked against the plan
The plan’s success criteria run as gates — nothing is “done” until the checks it promised actually pass.
- Runs the plan’s own criteria
- Surfaces deviations explicitly
- No silent “looks good to me”
Review-ready diffs & PRs
Rich diffs, staging, and one-click PRs with CI tracking — so AI output goes through the same review gates as every other change.
- Rich diff + staging
- One-click PR with CI status
- Same gates as human code
Every step writes to a shared Library — research, plans, and PRs your whole team can read, with git-tracked frontmatter and a link straight to the source.
Built for teams managing agents at scale
Section titled “Built for teams managing agents at scale”Most AI tools make one person fast for an afternoon. Hyprlayer is the workspace a whole team runs on — orchestrate agents across every project from one window, each working from shared, reviewable specs, so the mental model behind the code stays with the team instead of walking out the door with a prompt.
Your team shares one version-controlled set of AI skills — research, plan, implement, validate, review — so every agent runs the same playbook, not ad-hoc prompts.
Every change starts as a phased plan with explicit success criteria, reviewed and refined before a line is generated. Engineering rigor, not vibes.
Research and plans persist in a shared thoughts repo, so context outlives the session and travels across the team. No more understanding that vanishes with the prompt.
Each agent runs in its own isolated git worktree and branch, so the team ships many streams at once without stepping on each other.
Rich diffs, staging, and one-click PRs with CI tracking — so AI output goes through the same review gates as every other change.
A real shell in the worktree (⌘`) and a ⌘P palette across workstreams, agents, docs, and files — the whole loop without leaving the app.
Works with your AI
Section titled “Works with your AI”Drive the coding agent you trust — Hyprlayer wraps it in curated skills, shared specs, and team collaboration.
Anthropic’s official CLI, driven natively inside the app with slash commands, sub-agents, and permission modes.
OpenAI’s Codex CLI is landing next, with the same curated skills, shared specs, and collaboration wrapped around it.
What’s new
Section titled “What’s new”Ollama harness, auto-named agent tabs, and a dedicated workstream GIT panel.
- Run your agents on local or remote Ollama models. Settings → Harnesses → Ollama probes any reachable server for its installed models and exposes them in the composer picker.
- New agent tabs now auto-title from a summary of your first prompt with a small model (configurable in Settings → General). The title types out in the tab strip and locks once set — a manual rename is never overwritten.
- A new GIT panel in the changes sidebar collects per-workstream git controls: an ahead/behind chip, base-branch retargeting that re-points an already-open PR, pull-base (fetch + merge), and a prune-fetch refresh. Archiving a merged workstream can now also delete its remote branch when you've opted in.
Prefer the terminal?
The same curated skills and spec-driven workflow ship as a CLI for Claude Code, Copilot, and OpenCode.
Explore the Hyprlayer CLI →