Agent Context Framework

Agent Context Framework (ACF)

Lightweight, file-based framework for bootstrapping, preserving, and handing off project context across AI agent sessions.

Agent Context Framework helps teams keep continuity in AI-assisted software projects, especially across fresh sessions, context resets, and handoffs. Canonical project state lives in Markdown files under context/*, while MCP Memory serves as a supporting semantic recall layer.

For the Polish version, see [PRZECZYTAJMNIE.md](PRZECZYTAJMNIE.md).

Why ACF

• keeps a canonical, file-based source of truth in context/*
• gives agents a repeatable startup and handoff workflow
• reduces repeated re-explanation after compaction or session resets
• keeps MCP as a supporting memory layer instead of the only project record
• stays lightweight in architecture while covering a broad project workflow

Scope Boundary

ACF should stay lightweight by default. For the detailed product boundary and future-tier guidance, see [docs/acfscopeboundary.md](docs/acfscopeboundary.md).

Agent Compatibility

ACF is agent-agnostic at the workflow level.

That means the core model works with:

• Codex
• ChatGPT
• Claude
• Cursor
• Aider
• Continue
• custom CLI or editor agents

The framework itself is based on:

• canonical context files in context/*
• repeatable startup and handoff rules
• MCP as an optional supporting memory layer

Some automation in this repository is Codex-aware, especially MCP registration scripts, but the workflow itself does not depend on Codex.

What You Get

• context/next_context_sync.md for startup and active code-map references
• context/project_map.md for lightweight repository orientation and module layout
• context/procedures/ for procedural memory and repo-specific ways of doing recurring work
• context/handoff_migration.md for technical state, decisions, and runbooks
• context/context_change_history.md for history, tests, and commits
• context/master_plan.md for scope and delivery policy
• tools/acf_setup.sh as the unified interactive entrypoint
• tools/init_project_context.sh for direct project initialization
• tools/bootstrap_existing_project.sh for draft context generation in existing repositories
• tools/generate_project_map.js for draft project-map generation
• tools/sync_project_map.js for refreshing a generated project-map draft without overwriting manual edits
• tools/verify_context_links.js for code-reference validation
• tools/verify_project_map.js for project-map path validation
• tools/run_next_verification.js for running the verification command from next_context_sync.md
• tools/archive_context_history.js for archiving older change-history entries
• tools/acf_status.js for a quick framework status summary
• MCP helper scripts for container setup, registration, and smoke checks

Requirements

• bash
• node and npm
• python3
• podman for the container-based MCP workflow
• codex if you want automatic MCP registration from the setup flow

After Download

If you just downloaded or cloned this repository, use this exact order:

1. Move into the repository directory.
2. Install dependencies:

npm install

1. Start the unified setup:

npm run acf:setup

1. When prompted, choose one mode:

• new if you are starting a fresh project from this framework
• existing if you are attaching the framework to a repository that already has code

1. During setup, ACF can also:

• create and start Podman containers for MCP Memory
• register MCP Memory in Codex
• optionally register a repo-scoped filesystem MCP server

1. After setup finishes, review:

• context/next_context_sync.md
• context/project_map.md
• context/procedures/README.md
• context/handoff_migration.md
• docs/mcp_ai_memory_setup.md

1. Verify the generated context references:

npm run verify:context

1. Continue with:

• [NEWPROJECTCHECKLIST.md](NEWPROJECTCHECKLIST.md) for operational setup
• [docs/systemflow.md](docs/systemflow.md) to understand how file-based context and MCP Memory work together

Start Modes

New Project

Choose new when:

• you are creating a project from scratch
• you want the framework to initialize clean placeholders and MCP naming
• you want to fill in the context files as the project begins

What ACF does in this mode:

• initializes project placeholders
• generates config/project.env
• prepares context files for a new project
• can create and register MCP infrastructure

Existing Project

Choose existing when:

• the repository already contains real code
• you want to attach ACF to an existing codebase
• you want draft context files generated from a repository scan

What ACF does in this mode:

• initializes project-specific config
• scans the repository for manifests, likely stack, top-level directories, and likely entrypoints
• generates draft context files and marks them for human review
• can create and register MCP infrastructure

Important:

• existing mode generates a starting draft
• the user should review and correct the generated context files before treating them as canonical

Quick Start

If you already know the framework and just want the shortest path:

npm install
npm run acf:setup
npm run verify:context

Quick Install Via Agent

If you are already working with an AI coding agent, you can delegate the installation instead of doing the setup manually.

Use a prompt like this inside the target repository:

Install Agent Context Framework in this repository.

Steps:
1. Clone or copy ACF into this repo.
2. Run `npm install` in the ACF directory.
3. Run `npm run acf:setup`.
4. If this repository already contains code, choose `existing`.
5. After setup, run `npm run verify:context`.
6. Summarize what was created and what still needs manual review.

For a more explicit request:

Attach Agent Context Framework to this existing repository, use the `existing` setup mode, verify the generated context links, and then tell me which context files I should review first.

Shortest copy-paste version:

Install ACF here, choose `existing` if this repo already has code, run the setup, verify context links, and tell me what I should review next.

Non-Interactive Example

./tools/acf_setup.sh \
  --mode new \
  --project-name "New Project" \
  --project-slug "new-project" \
  --db-port 55735 \
  --redis-port 56810 \
  --with-filesystem

By default, the initializer:

• derives project-slug from the parent repository folder name
• builds project-name from that slug
• uses default ports 55735 and 56810

Typical Workflow

LLM Agent
   |
   v
context/next_context_sync.md
   |
   v
context/project_map.md
   |
   v
MCP semantic recall
   |
   v
verify with canonical context
   |
   v
work on codebase
   |
   v
update context + memory

In ACF, the recommended session flow is:

1. The agent starts from context/next_context_sync.md.
2. The agent reads context/project_map.md to orient on modules, entrypoints, and important files.
3. If the task matches a known repository workflow, the agent reads the relevant file in context/procedures/.
4. The agent recalls relevant MCP Memory entries.
5. The agent verifies MCP recall against canonical files in context/*.
6. The agent works on the codebase.
7. The work slice is closed with:

• a short code review
• an implementation commit
• updates to context/*
• one concise MCP write

Primary rule:

• context/* wins over MCP.

Multi-Agent Use

ACF works well when multiple agents collaborate on the same project.

Typical examples:

• one agent works on backend changes
• another agent works on frontend changes
• another agent reviews or updates documentation

This works because:

• context/* is shared canonical state
• context/next_context_sync.md is the common startup point for every new session
• context/project_map.md gives each session a lightweight repo map before deeper analysis
• context/procedures/ captures repo-specific ways of carrying out recurring changes
• MCP Memory is used for concise recall, not as the only source of truth

For multi-agent work to stay reliable:

• every closed work slice should update context/*
• every closed work slice should write one concise MCP record
• every new session should start from context/next_context_sync.md and then read context/project_map.md
• when MCP and context/* differ, context/* wins

In short:

• ACF supports multiple agents on one project
• the shared context files are what keep those agents aligned

MCP Setup

Use [docs/mcpaimemorysetup.md](docs/mcpaimemorysetup.md) for MCP registration and environment details.

Useful commands:

npm run acf:setup
npm run generate:project-map
npm run project-map:sync
npm run context:run-verification
npm run context:archive-history -- --keep-latest 100
npm run acf:status
npm run mcp:init
npm run verify:context
npm run verify:project-map
npm run mcp:register
npm run mcp:status
npm run mcp:smoke

Documentation Map

• [NEWPROJECTCHECKLIST.md](NEWPROJECTCHECKLIST.md): human setup checklist
• [docs/acfscopeboundary.md](docs/acfscopeboundary.md): what belongs in core ACF vs future advanced modes
• [docs/systemflow.md](docs/systemflow.md): architecture and memory flow
• [docs/agentworkflow.md](docs/agentworkflow.md): day-to-day agent workflow
• [docs/mcpaimemorysetup.md](docs/mcpaimemorysetup.md): MCP registration and runtime setup
• [docs/mcpmemoryrules.md](docs/mcpmemoryrules.md): source-of-truth and write/retrieval rules

Recommended Startup Prompt

Act as a senior engineer for project .

Startup:
1. Read `context/next_context_sync.md`.
2. Read `context/project_map.md`.
3. Read a relevant file from `context/procedures/` if the task matches a known procedure.
4. Recall prior decisions from MCP `` using project-specific keywords.
5. Verify recall against:
   - `context/handoff_migration.md`
   - `context/context_change_history.md`

Rules:
- `context/*` is the source of truth.
- If MCP and `context/*` differ, `context/*` wins.
- After each closed slice:
  - short code review,
  - commit,
  - update `context/*`,
  - 1 concise MCP record.

Contributing

Contributions are welcome. Please read [CONTRIBUTING.md](CONTRIBUTING.md) before opening a pull request.

Security

If you find a security issue, please follow [SECURITY.md](SECURITY.md).

License

This project is available under the [MIT License](LICENSE).

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: filipsawicki
• Source: filipsawicki/agent-context-framework
• License: MIT

Install and usage instructions live in the source repository linked above.