Project Corpus

Русская версия

Persistent project state and authority for AI agents.

Project Corpus is a Markdown-first, vendor-neutral protocol that lets ChatGPT, Codex, Claude, and other AI agents resume a project from clean context without relying on chat memory.

It records the project's canonical identity, current state, evidence, exact next action, and authority boundaries in files that remain available when an AI session ends. The Protocol needs only Markdown; the optional Runtime adds technical enforcement.

flowchart LR
  subgraph without_corpus["Without Project Corpus"]
    direction TB
    A1["AI Session A"] --> M["chat memory / ad-hoc notes"]
    M --> E["session ends"]
    E --> B1["AI Session B"]
    B1 --> I["reconstruct from incomplete context"]
  end
  subgraph with_corpus["With Project Corpus"]
    direction TB
    A2["AI Session A"] --> C["Project Corpus"]
    C --> P["PROJECT.md — what the project is"]
    C --> S["STATUS.md — where the project is now"]
    C --> T["Tasks — authorized work"]
    C --> R["Reports — evidence"]
    C --> Y["Policy — project-side limits"]
    C --> B2["AI Session B"]
    B2 --> L["load canonical state"]
    L --> N["continue from exact next action"]
  end

Loading

AI sessions are disposable. Project state is not.

More than a memory bank

A memory bank helps an agent remember information. Project Corpus additionally defines what is canonical, what is current, what counts as evidence, what comes next, and who may change state.

60-second Quick Start

No Runtime, MCP server, or package installation is required.

1. Download or clone this repository.

2. Copy the contents of templates/v2/minimal/ into the root of your project. From a repository checkout:

   macOS/Linux:

   cp -R templates/v2/minimal/. /path/to/your-project/

   PowerShell:

   Copy-Item -Force .\templates\v2\minimal\AGENTS.md C:\path\to\your-project\
   Copy-Item -Recurse -Force .\templates\v2\minimal\.project-corpus C:\path\to\your-project\

3. Fill in .project-corpus/state/PROJECT.md with the project ID and durable project identity.

4. Fill in .project-corpus/state/STATUS.md with the same project ID, verified baseline, blockers, evidence references, and exact next action.

5. Give ChatGPT, Codex, Claude, or another agent access to the project folder. In a manual chat, upload AGENTS.md, PROJECT.md, and STATUS.md. If you are uploading files manually, also upload the active Task and any Reports referenced by STATUS.md. Then say:

   Load Project Corpus for this project. Follow AGENTS.md; read .project-corpus/state/PROJECT.md and .project-corpus/state/STATUS.md completely; if an active Task or cited Reports are available, load only those relevant artifacts; state the project ID, current status, active Task, and exact next action; then continue from that action.

Want enforcement too?

The optional Runtime is installed via PyPI:

pip install project-corpus
project-corpus validate /path/to/your-project
project-corpus doctor /path/to/your-project

It adds validation/doctor, external owner trust, policy enforcement, expected-hash transactions, audit/recovery, and optional local stdio MCP. Controlled writes require an external owner trust grant; follow the Runtime CLI guide rather than treating project content as authority.

Protocol and Runtime

Project Corpus Protocol	Optional Project Corpus Runtime
Markdown-first and vendor-neutral	Validation and controlled CLI
No installation or database	External owner trust and policy enforcement
Manual workflow supported	Verified writes and audit/recovery
MCP optional	Optional local stdio MCP

The Protocol is the portable product contract. The Runtime implements it but does not define or silently amend it. Enforced guarantees apply only in controlled modes on the qualified local filesystems listed in the platform matrix.

Maintained V1 workflow

The original V1 workflow remains available for existing projects. Its template contains:

AGENTS.md
OPERATOR_PROFILE.md
PROJECT_ROADMAP_CURRENT.md
CORPUS_ACCESS_CURRENT.md
LOADER_PROMPT_CURRENT.md
SESSION_HANDOFF_CURRENT.md
SESSION_HANDOFF_FULL_CURRENT.md
Tasks/
Report/

AGENTS.md defines the rules. The roadmap and handoffs preserve current state. Tasks/ contains scoped work orders; Report/ contains evidence of completed work.

Choose one V1 access mode

Mode	Best for	What happens
Direct folder	Codex, Claude Code, ChatGPT Work, and other local agents	Give the client access only to the Corpus folder and let it read or update the files directly.
Your own MCP	Clients that support an MCP server or file connector	Connect a trusted server of your choice, restrict it to the Corpus root, and record its real capabilities.
Manual session	Any AI chat, with no setup	Upload the seven current files at the start; at the end, save only the replacement files the AI returns.

The three modes use the same protocol. An unavailable MCP server is not a blocker: switch to direct-folder or manual access.

Read the detailed guides:

• Direct folder
• Your own MCP
• Manual sessions

V1 five-minute start

1. Download or clone this repository.
2. Copy one language folder to a safe location and name it Corpus:
   • template/en for English;
   • template/ru for Russian.
3. Open CORPUS_ACCESS_CURRENT.md and choose an access mode.
4. Give your AI the matching text from PROJECT_INSTRUCTION_TEMPLATE.md.
5. Say what you want to build in normal language.

Example:

Start a new project. I want to build a local photo organizer. First help me define the architecture. Do not install, delete, publish, or contact anyone.

Client guides

• ChatGPT
• Codex
• Claude Code
• Claude Desktop
• Other AI clients

Client interfaces and plan availability can change. Each guide keeps volatile client setup separate from the stable Corpus protocol and links to official documentation.

V1 manual mode really works

If you do not want to configure local folders or MCP:

1. upload the seven current files in a new session;
2. upload only the Tasks and Reports relevant to the request;
3. ask the AI to read AGENTS.md first and issue a loading receipt;
4. work normally;
5. ask for a manual synchronization package;
6. back up the old local files and save the returned replacements.

Ordinary synchronization must not rewrite AGENTS.md or OPERATOR_PROFILE.md. The AI should return only files that actually changed and must not claim that it saved them on your computer.

Important limits

Manual and direct-folder modes are not enforced Runtime modes. Controlled Runtime guarantees are scoped to documented operations on qualified local NTFS, ext4, and APFS filesystems; read the security model for the complete boundaries.

• Project Corpus is a documentation protocol, not a security sandbox.
• Real access is controlled by your AI client, filesystem permissions, or your chosen MCP server.
• V1 direct-folder and manual workflows do not gain Runtime enforcement.
• The optional V2 Runtime provides local CLI and stdio MCP only; it does not provide HTTP/remote MCP or audit third-party servers.
• A saved Report is not proof that a service or external system is currently healthy.
• Do not store passwords, tokens, cookies, seed phrases, or API keys in the Corpus.
• Keep one active project per Corpus.

More: how it works, FAQ, security model, languages, and moving or removing a Corpus.

Current status

Project Corpus V2 is the current open-source generation. The bilingual templates, three access modes, documentation links, repository integrity and optional Runtime are checked automatically. Release changes are recorded in the changelog.

Feedback

Use GitHub Issues for bugs, setup questions, and improvement ideas. Do not post secrets or private Corpus contents.

Support the project

Project Corpus is free and MIT-licensed. Voluntary support is available through USDT on TON; see the support page and wallet address. A donation is not a purchase and does not provide additional rights or guarantees.

License

MIT. See LICENSE.