Skip to main content

processkit

processkit is a provider-neutral process layer for AI-assisted software projects.

It gives agents structured project memory, reusable domain skills, and validated MCP tools. The practical effect is simple: agents can read and write durable work items, decisions, notes, artifacts, migrations, and other project records through explicit contracts instead of loose files and provider-specific conventions.

processkit is designed to be used directly by MCP-capable harnesses or installed by an environment manager. aibox is one supported managed installer, not a runtime dependency.

What ships​

  • 140 skills across engineering, product, research, data, design, documents, devops, and processkit operations.
  • 25 MCP server entry points for entity management, search, routing, release checks, projections, and gateway access.
  • 16 shipped project-memory schemas for durable v2 entities such as WorkItem, DecisionRecord, Artifact, Note, LogEntry, Migration, Actor, Role, Binding, Scope, Gate, Discussion, and related primitives.
  • 5 package tiers: minimal, managed, software, research, and product.
  • A provider-neutral MCP gateway that can expose processkit through one stdio server, one streamable HTTP daemon, or a stdio proxy.

Design goals​

processkit separates process semantics from harness behavior:

  • The schemas define durable project memory.
  • The skills describe repeatable workflows and domain gotchas.
  • The MCP tools validate writes, enforce state transitions, and keep the context searchable.
  • The gateway gives harnesses one processkit entry point without knowing about Claude, Codex, OpenCode, Hermes, Aider, or any other provider-specific runtime.

That split keeps processkit forkable, installable by hand, and usable by multiple harnesses. Integrations can automate install and lifecycle, but they do not own the processkit contracts.

How to use it​

The direct path is:

  1. Download a release tarball from GitHub Releases.
  2. Copy the shipped context/, .processkit/, and AGENTS.md files into your project.
  3. Register processkit-gateway or selected per-skill MCP servers with your harness.
  4. Ask the agent to use processkit tools for entity reads and writes.

Managed installers can do those steps for you. For example, aibox can fetch a pinned processkit release, choose a package tier, write harness MCP configuration, and supervise a gateway daemon in a devcontainer.

Where to go next​

  • Getting Started explains the manual and managed install paths.
  • MCP Servers explains gateway, daemon, stdio-proxy, aggregate, and per-skill layouts.
  • Primitives explains the project-memory entity model.
  • Skills explains the skill package format and catalog.
  • Packages explains the five package tiers.
  • v2 Contracts explains the current deliverable boundary and demoted legacy primitives.

Current status​

The current release line is pre-1.0. Breaking changes may still land in minor releases, and the changelog calls them out explicitly.

v0.25.0 is a breaking pre-1.0 release. It completes the SmoothTiger/SmoothRiver v2 deliverable boundary, adds the provider-neutral processkit-gateway, removes legacy first-class primitive schemas from the shipped src/context/ surface, and turns the release checks into executable gates.