kvelmo

kvelmo is a local orchestration system for AI-assisted development.

It manages the full lifecycle of development tasks — from loading requirements through planning, implementation, review, and pull request submission. The Web UI and desktop app are the main experience, with CLI and TUI available for those who prefer them.

kvelmo is not the agent. It is the system around the agent: workspace setup, checkpoints, approvals, governance, observability, and recovery.

Why Use It

• Web-first workflow for creating tasks, watching progress, reviewing changes, and managing project state visually
• Human oversight at every step so planning, implementation, review, and submission stay explicit and inspectable
• Runs locally with no hosted service required for the core workflow
• Works across interfaces so you can start in the browser and continue in CLI or TUI without losing context
• Recovery built in through checkpoints, undo/redo, reset paths, and resumable task state
• Operational depth beyond the happy path: policy, CI, security, activity logs, exports, recordings, notifications, and task grouping

How It Works

start -> plan -> implement -> simplify? -> optimize? -> review -> submit -> finish

Each phase updates task state, records checkpoints, and exposes progress to every interface.

At a high level:

1. Start loads work from a file or provider and prepares the workspace
2. Plan produces a reviewable specification before code changes begin
3. Implement executes the approved direction in the project workspace
4. Simplify and Optimize are optional refinement passes
5. Review adds the human checkpoint before code leaves the machine
6. Submit creates the PR and updates provider-facing state
7. Finish cleans up after merge

Choose Your Interface

Desktop App

The desktop app is a native application for managing your workflow visually — no terminal required.

Desktop App Guide

Web UI

The browser UI gives you the same visual experience, started from one command:

kvelmo serve --open

Then open http://localhost:6337 if it did not open automatically.

Use the Web UI when you want:

• project and task dashboards
• visual workflow controls
• live output, chat, logs, and review context
• access to panels for memory, screenshots, browser automation, jobs, code graph, recordings, policy, CI, and more

Web UI Guide

CLI

The CLI is available for scripting, shell-native workflows, automation, and provider operations.

kvelmo start --from file:task.md
kvelmo plan
kvelmo implement
kvelmo review
kvelmo submit

CLI Reference

TUI

The TUI gives you a full-screen terminal dashboard with workflow control and live output, without switching to a browser.

kvelmo tui

What kvelmo Is

• An orchestrator for development work — it coordinates AI agents, it is not one itself
• A local product that runs on your machine with no hosted service required
• A workflow system with clear steps, approvals, and recovery at every stage
• Available everywhere — desktop app, Web UI, CLI, TUI, and RPC
• A governance layer for policy checks, approvals, auditability, and submission discipline

What kvelmo Is Not

• Not a one-click autonomous coder that silently edits and ships code
• Not a hosted service — the core workflow runs entirely on your machine
• Not only a shortcut to PRs; it also handles the operational edges around agent work
• Not limited to a single agent; Claude, Codex, and custom agents can all participate

Licensing and Third-Party Tools

kvelmo itself is BSD 3-Clause. There are two separate licensing questions — the agent CLIs kvelmo subprocess-invokes, and the libraries kvelmo links against. Both are covered below.

Agent CLIs (subprocess-invoked, not bundled)

kvelmo does not modify, patch, bundle, or redistribute any agent CLI. It invokes locally installed binaries as subprocesses through their documented interfaces — the same way a shell script or CI pipeline would. Users install and authenticate agent CLIs independently.

Tool	License	Binary invocation status
Claude Code	Proprietary (Anthropic)	Designed for programmatic use. The --print and --output-format stream-json flags exist for exactly this purpose, and Anthropic's own Agent SDK uses the same subprocess pattern. Consumer ToS explicitly exempts the CLI from the prohibition on automated access.
Codex CLI	Apache 2.0 (OpenAI)	Open source. Subprocess invocation, wrapping, and modification are all permitted by the license.
Custom agents	Varies	Any CLI exposing a supported protocol can participate.

Linked dependencies (Go, frontend, Rust/Tauri)

Every linked dependency across all three ecosystems is permissively licensed — MIT, Apache 2.0, BSD-2/3-Clause, ISC, MPL 2.0, or OFL 1.1. No GPL, AGPL, SSPL, or BUSL code is pulled in. The only LGPL touch is webkit2gtk on Linux via Tauri, which is dynamically linked against the system library (the standard Tauri pattern, compatible with BSD distribution).

This is enforced automatically by make quality, which runs:

• go-licenses check ./... --disallowed_types=forbidden,restricted — blocks GPL/AGPL/LGPL Go dependencies. Install: go install github.com/google/go-licenses@latest
• cargo deny check licenses sources — blocks non-allowlisted Rust crates (allowlist in web/src-tauri/deny.toml). Install: cargo install cargo-deny --locked

Both checks are skipped with a notice if the tools aren't installed, so they don't block local development — but when present, a license regression fails the build loudly.

Key Capabilities

Capability	What It Covers
Task intake	Files plus providers such as GitHub, GitLab, Linear, Wrike, Jira, and Azure DevOps
Workflow orchestration	Start, plan, implement, simplify, optimize, review, submit, finish
Checkpointing	Undo, redo, reset, abandon, and resumable task history
Observability	Status, watch, logs, activity, jobs, stats, recordings, event log
Governance	Approvals, checklisting, policy, CI checks, security scanning, audit trail
Project intelligence	Memory, screenshots, browser automation, code graph, discovery, recap, explain
Operations	Backups, exports, notifications, access tokens, hooks, worker pools, task groups

Installation

kvelmo can be installed with the install script or built from source.

Install script:

curl -fsSL https://raw.githubusercontent.com/valksor/kvelmo/master/install.sh | bash

Build from source:

git clone https://github.com/valksor/kvelmo.git
cd kvelmo
make build
make install

After installation:

kvelmo version
kvelmo diagnose

If your preferred agent CLI is available locally, kvelmo can orchestrate it. Claude CLI is the default path, and Codex plus custom/API-backed agents are also supported.

Getting Started

The fastest way to start is the desktop app or Web UI. For the Web UI:

kvelmo serve --open

For CLI users, the core workflow commands are:

kvelmo start "Refactor the settings panel"
kvelmo plan
kvelmo implement
kvelmo review
kvelmo submit

Documentation

Guide	Description
Quickstart	Get started and learn the workflow
Desktop App	Native application guide
Web UI Guide	Browser experience
CLI Reference	Command reference
Workflow Concepts	Steps, approvals, and recovery
Providers	External task sources
Configuration	Settings, tokens, and behavior

Development

make build
make quality
make test

See CONTRIBUTING.md for the full contributor workflow.

Prototype

prototype/ contains the historical reference implementation that informed the current architecture. It is preserved as read-only reference material and is not the active development target.

Data and Legal

kvelmo runs locally and does not collect usage analytics or phone home. Tasks, recordings, activity logs, and configuration stay on your machine; the only network traffic kvelmo originates is to providers, agents, and webhooks you explicitly configure.

• Data Contract — what kvelmo stores, where it lives, what is and is not transmitted, and what survives an upgrade
• Legal Disclaimer — AI risks, third-party compliance, acceptable and unacceptable use, financial responsibility
• Security Policy — vulnerability disclosure
• Trademark Notice — brand usage and fork naming
• Code of Conduct — community conduct

License

BSD 3-Clause License