CoProducer is a portable, vendor-neutral shared root of agentic infrastructure: experts, skills, tools, and connectors, governed by the standards in standards/. A small core team authors it once; it deploys identically to every organization that uses it.
Nothing client-, organization-, or person-specific ever lives inside. That separation is what makes a single copy servable to everyone: this root holds capability, and the roots a customer owns hold their context, memory, and secrets. The two meet in a workspace and stay separate on disk.
CoProducer is proprietary software. It is not open source. Copyright © 2026 Anthony Franco. All rights reserved.
Use is permitted only under a separate written license agreement with Anthony Franco (or a party he has authorized in writing to grant licenses). Access to this repository or presence of these files on a machine is not a license. See LICENSE for the full terms.
| Directory | Contents |
|---|---|
standards/ |
How things are written: the instruction quality bar, the Play, Playbook, and primitive formats, and shared conventions |
skills/ |
Capabilities a user invokes by name to get an output |
experts/ |
Personas that carry a perspective and judge work through it |
tools/ |
Deterministic operations that skills and experts call |
connectors/ |
Authenticated access to outside platforms |
system/ |
Version, changelog, and the update path a deployed copy follows to reach a newer release |
The four primitive types and how they invoke one another are defined in standards/primitives.md. Family indexes live in each family's AGENTS.md (skills/, experts/, tools/, connectors/).
infrastructure/ is core-team build machinery: plans, the validator, the release scripts. It is stripped from every release, so a deployed copy never contains it.
A workspace composes this root with one or more roots the user owns. This root is read-only in use; the owned roots are where work, memory, and additions live. Each owned root declares its own type (personal, org, client, department, industry) and states which memory it provides, so a skill can ask for something like a voice or a company profile without knowing whose workspace it is running in.
Agentic hosts disagree about which file they read on startup, so this root ships small adapter files (CLAUDE.md, .cursor/rules/, .windsurf/rules/, and others as needed) that contain no instructions of their own and only point at AGENTS.md. From there, each directory's own AGENTS.md carries the chain forward.
Extensions and overlays live in the owned roots. A customer who needs a skill this root does not have, or a different version of one it does, adds it on their side. Nothing here has to be edited or forked to make room.
Releases are versioned and files-only. The core team uses git as transport between its own machines, but a deployed copy at a customer destination never contains a .git directory. Those destinations sit on drive-sync platforms, and a synced .git gets corrupted. So a deployment is a set of files copied into place, and an update is that same operation against a newer version.
Agents start at AGENTS.md and follow the chain.
Authors start at standards/instruction-quality.md. It sets the bar every instruction file in this root is held to, and reading it first is cheaper than rewriting against it later.
Core-team clones: after cloning (or connecting git on a synced tree), run sh scripts/install-hooks.sh once, or just run node infrastructure/validate/generate.js / validate.js — either enables the committed .githooks so plugin/ and other generated outputs stay in lockstep on every commit. Details: infrastructure/validate/AGENTS.md.