Vibe-code with discipline

koryph turns AI coding agents into a software factory. Point it at a git repo and it plans the work, runs a fleet of agents in parallel without merge conflicts, keeps them inside budget and policy, merges only what passes your gate — and ships signed, attested releases at the end. One static binary. No SaaS. Nothing to unsubscribe from.

Install · Quickstart · Features · Zero to shipped · How koryph compares · Download

Using an AI agent? Point it at llms.txt →

Why koryph exists

Tools like Claude Code changed the constraint. An AI agent writes a feature in minutes — so the bottleneck is no longer typing code, it is everything around the code: keeping five agents from trampling each other, keeping spend inside your subscription, keeping unreviewed changes off main, keeping commits signed and settings hardened, and turning the result into a release someone else can actually trust and install.

Vibe coding without that process is how you get a repo full of merge conflicts, a surprise bill, and an unshippable pile of code. Doing that process by hand is how you lose the speed you just gained.

koryph is the missing assembly line. It carries the process — planning, parallelism, review, budgets, hygiene, releases — so you and your agents can carry the ideas. It is opinionated about process, never about your application: no frameworks chosen for you, no folder layouts, no lock-in.

The three pillars

Build — the agent factory. koryph reads your project's task graph (beads), batches conflict-free work by each task's declared footprint, and dispatches headless agents into isolated git worktrees — in parallel, continuously, under the correct account. Finished work goes through review → rebase → your green gate → fast-forward merge. Nothing lands that doesn't pass.

Protect — hygiene as code. Branch-protection rulesets, repo settings, and security posture live as committed JSON you can check and apply (koryph repo check|apply), with named profiles like the built-in oss-solo-maintainer. Commit signing is enforced from vault-served keys. Protected paths keep agents away from your CI, hooks, and policy files. koryph doctor catches drift before it bites.

Ship — the release train. Conventional commits drive versioning for any language. Releases are draft-until-complete and immutable: binaries, checksums, SBOMs, cosign signatures, and SLSA build provenance all attach before anything publishes. A vault-backed release bot keeps PR checks flowing — with graceful fallbacks when you can't install one.

New in this release

• koryph adopt — one command takes any existing repo to a fully-onboarded, validated project: detect → plan → consent → execute → verify. Adoption is now the ten-minute front door.
• Intent routing — describe what you want built in a normal Claude session; koryph routes the ask through /koryph-design into a reviewed design doc and an implementable, conflict-aware bead graph.
• Typed recovery, hardened — stall detection, a dead-agent patrol, evidence-bound bounded continuation and repair, and operator overrides that a running loop respects instead of clobbers (koryph merge --close-bead, koryph inject, koryph status --frontier).
• Terminal cockpit, overhauled — a live activity tail of each agent's thinking and tool calls, structured stall flags, accurate fleet/quota readouts, and a queue that explains why every bead is or isn't dispatching.

The full history is in the CHANGELOG.

Datasheet

What koryph is not

• Not a SaaS. There is no account, no telemetry, no server. Everything runs on your machine against your git remotes and your AI subscriptions.
• Not a framework. koryph never chooses your language, layout, or dependencies. Delete koryph and your repo is still a perfectly ordinary repo — you lose the factory, not the product.
• Not a replacement for judgment. The gate, the review stage, and protected paths exist precisely so that speed never outruns your standards.

Get started

Adoption is two commands and about ten minutes:

# 1. Install — single static binary, no runtime needed
brew install koryph/tap/koryph

# 2. Adopt any existing repo: the wizard detects what's missing, shows a
#    plan, asks consent, and finishes with a green `koryph validate`
koryph adopt ~/src/myproject

# 3. Describe what you want built (in an agent session), then run the loop
koryph run --project myproject --once --auto-merge --review

Continue with the quickstart, or read Zero to shipped for the full journey — plan, build, protect, release. Prefer to drive every step by hand? The wizard's lower-level verbs are all still there.

This book is layered so you can stop reading whenever you have enough: Features for the map, Concepts for the ideas, the user guide for operators and collaborators, and the developer guide for contributors to koryph itself.

Project

• How koryph compares — the 2026 agent-orchestration landscape, honestly mapped.
• Security — how koryph secures itself and your machines, and how to report a vulnerability.
• Community & contributing — filing issues, contributing code, and the Apache-2.0 licensing model.

The name

koryph comes from the Ancient Greek κορυφαῖος (koryphaios) — the leader of the chorus in classical Greek drama. The koryphaios stood at the head of the chorus and spoke on its behalf whenever it took part in the action: one voice fronting many performers moving in step. The root κορυφή (koryphē) means "crest" or "summit", and the word lives on in several modern languages as a term for the leading figure in a field.

That is exactly this tool's job: one process fronting a fleet of autonomous coding agents — queueing their work, dispatching them in parallel, and speaking for them at the merge. It is pronounced KOR-iff.

For AI agents and tools: a machine-readable index of the canonical docs (llmstxt.org format) is published at /llms.txt — ingest it to map the project and its operating contract.