OneContext

A shared project memory and coordination layer for teams using different AI coding agents.

Submission category: Developer tools

OneContext helps a team of developers keep ChatGPT, Claude, Codex, GitHub Copilot, Cursor, and VS Code aligned around the same project. It stores useful project knowledge, retrieves only relevant context, shows live teammate intent, and warns when parallel work may overlap.

Watch the OneContext demo on YouTube

When four developers work on one repository with different AI assistants, each assistant sees a different conversation history. Important decisions stay trapped in one chat, teammates repeat work, and two people can unknowingly edit the same area.

OneContext creates a shared project layer between the team and their AI tools:

Project sources, decisions, tasks, and live intent
                         |
                         v
              OneContext memory and gateway
                 /          |          \
                v           v           v
        Web dashboard   Chrome add-context   MCP / VS Code
        Memory Chat     ChatGPT + Claude     Codex + Copilot

The assistants remain independent. OneContext does not replace them or require the whole team to use the same provider.

• Project workspace: Create/select projects and maintain a project goal and sprint brief.
• Connected sources: Upload Markdown/text notes and add a GitHub repository URL for indexing.
• Memory Chat: Ask questions about architecture, decisions, files, tasks, and project history.
• Structured memory: Keep concise decisions, handoffs, tasks, and project activity instead of blindly copying every conversation.
• Knowledge Graph: Visualize relationships between files, concepts, and decisions.
• Decision timeline: Review why important project choices were made.
• Chrome extension: Add relevant project context to ChatGPT and Claude with one button.
• OneContext Live: Create a Team Code, publish task intent, see active teammates, and detect overlapping work.
• VS Code extension: Join a team, automatically publish active-file/saved-file presence, ask Codex with team context, and save handoffs.
• MCP server: Give compatible coding agents a common interface for context, conflict checks, progress updates, and handoffs.
• Codex-first CLI: Print a context-enriched prompt for a terminal-callable Codex workflow.

1. A user adds project sources, notes, decisions, or live intent.
2. OneContext stores project-scoped information in PostgreSQL.
3. A question is routed to the active project and relevant memories are retrieved.
4. The dashboard, Chrome extension, MCP server, or Codex wrapper receives a compact context block.
5. A teammate or AI agent can continue from the same decisions and active work.

The application works without an OpenAI API key by using local routing and retrieval fallbacks. OpenAI-powered features are optional and configured only on the server.

Install these before running the project:

• Node.js 18 or newer (Node.js 20 or 22 LTS is recommended)
• npm
• PostgreSQL 14 or newer
• Git
• Google Chrome, for the browser extension demo
• VS Code, for the live team and MCP demos
• Optional: Docker Desktop for Redis-backed realtime presence
• Optional: an OpenAI API key for AI routing, reranking, answers, or memory distillation

git clone https://github.com/varunkumar-2005/Onecontext.git
cd Onecontext
npm install

Create a database named onecontext in PostgreSQL. The database user must have permission to create tables and extensions used by the schema.

For a local PostgreSQL installation, a connection string looks like this:

postgresql://postgres:<YOUR_PASSWORD>@localhost:5432/onecontext

Do not put a real password in this README or commit it to Git.

Copy the template:

Copy-Item .env.example .env.local

Then edit .env.local. A safe local example is:

DATABASE_URL=postgresql://postgres:<YOUR_PASSWORD>@localhost:5432/onecontext
SESSION_SECRET=replace-with-a-long-random-secret

# Shared gateway authentication. Keep this private.
ONECONTEXT_GATEWAY_KEY=replace-with-a-local-gateway-key

# The project used by the demo.
ONECONTEXT_PROJECT_ID=atlas-project

# Optional AI features. Leave OPENAI_API_KEY empty to use local fallbacks.
OPENAI_API_KEY=
ONECONTEXT_USE_OPENAI_EMBEDDINGS=false
ONECONTEXT_USE_AI_ROUTING=true
ONECONTEXT_CONTEXT_ROUTER_MODEL=gpt-4o-mini
ONECONTEXT_USE_AI_RETRIEVAL=false
ONECONTEXT_USE_AI_ANSWERS=true
ONECONTEXT_USE_AI_CONVERSATION_MEMORY=false
ONECONTEXT_AI_MODEL=gpt-4o-mini

# Realtime service.
REALTIME_PORT=8787
ONECONTEXT_REALTIME_URL=ws://localhost:8787/live
REDIS_URL=redis://localhost:6379

The exact values in your .env.local are private. Never commit database passwords, OpenAI keys, session secrets, or gateway keys.

Run this once after PostgreSQL is available:

npm run db:init

The initializer reads DATABASE_URL, creates the configured database if needed, and applies infra/db/schema.sql.

Use two PowerShell terminals from the repository root.

Terminal 1 - web application:

npm run dev

Terminal 2 - live team presence:

npm run realtime

Open:

• Dashboard: http://localhost:3000
• Login: http://localhost:3000/login
• Live team page: http://localhost:3000/team

If the local seed is present, the demo account is:

Email:    suresh@example.com
Password: demo1234

Change or remove demo credentials before deploying the project publicly.

After signing in, use the project navigation as follows:

For a first demo, upload or add notes containing content such as:

# Sprint planning

Goal: Build one shared memory for a four-person AI-assisted team.
Current sprint: Capture activity and align the team's AI coding agents.

Decision: Keep retrieval provider-agnostic so each developer can keep using
their preferred coding assistant.

Then ask Memory Chat:

What is the current project goal, what decision has the team made about retrieval, and what should a new teammate know before editing the code?

The browser extension is a local Manifest V3 extension. It adds an Add context action beside supported ChatGPT and Claude prompts.

1. Start the web app with npm run dev.

2. Open Chrome and visit chrome://extensions.

3. Enable Developer mode.

4. Click Load unpacked.

5. Select this folder:

   C:\Users\<YOU>\Desktop\onecontext\browser-extension
   

6. Click the OneContext extension icon and configure:

   • API URL: http://localhost:3000
   • Project ID: atlas-project
   • Gateway key: the value of ONECONTEXT_GATEWAY_KEY from .env.local

7. Click Save & verify.

8. Refresh ChatGPT or Claude after saving.

For a second laptop, use the host laptop's reachable address instead of localhost, for example:

API URL: http://192.168.31.79:3000
Project ID: atlas-project
Gateway key: the same private gateway key

The IP address above is only an example from the local demo. Replace it with the host laptop's current Wi-Fi IPv4 address.

In ChatGPT, ask a project question such as:

We are building OneContext. What is the team's current goal and what should I consider before planning the next feature?

Click Add context. The extension retrieves relevant project memory and places a block like this into the prompt:

[PROJECT CONTEXT - OneContext]
Project: Atlas project
Relevant context: ...
[END CONTEXT]

Now switch to Claude, ask:

Continue this project work. Based on the shared context, propose the next implementation step and explain which files are relevant.

Click Add context in Claude as well. Claude receives the same project memory without requiring you to copy the entire ChatGPT conversation.

Casual prompts such as Hi, how are you? are intentionally not treated as project questions, so the extension reports that no project context was added.

The current packaged version is:

apps/vscode-extension/onecontext-vscode-0.1.2.vsix

1. Open VS Code.
2. Open the Extensions view with Ctrl+Shift+X.
3. Click the ... menu in the Extensions panel.
4. Choose Install from VSIX....
5. Select apps/vscode-extension/onecontext-vscode-0.1.2.vsix from the cloned repository.
6. Reload VS Code when prompted.

The VSIX can be installed on each teammate's VS Code. The source repository is not required on a teammate's laptop just to install the extension, but the local repository is required for the MCP server workflow described below.

1. Start the web app and realtime service.
2. Open /team in the browser and copy the Team Code.
3. In VS Code, open the Command Palette with Ctrl+Shift+P.
4. Run OneContext: Join Team.
5. Enter the Team Code and your display name.
6. Run OneContext: Configure Gateway Key once and enter the same gateway key used by the server. The extension stores it in VS Code Secret Storage.
7. Open the OneContext Team view in the Explorer sidebar.
8. Run OneContext: Start Task, enter the file/area and task description, and broadcast the intent.

The extension can also publish presence automatically when the active editor changes or a file is saved. Presence expires automatically after 15 minutes unless refreshed.

This is only needed when developing the extension itself:

cd apps\vscode-extension
npm install
npm run compile

Open apps/vscode-extension as its own VS Code window and press F5. This opens a separate Extension Development Host window. In that new window, open a project folder and use the OneContext Team view. Do not press F5 from the normal repository window if you are trying to test the packaged extension.

To rebuild the installable package:

cd apps\vscode-extension
npm run package

The package script compiles the extension and creates a versioned .vsix file. Install the newest version after rebuilding; VS Code can otherwise continue running an older installed package.

The MCP server is the agent-facing bridge. It exposes four tools:

The server is launched by Node from scripts/onecontext-mcp.mjs. It uses the workspace .env.local to reach the OneContext gateway; it does not store an API key in the MCP configuration file.

This repository includes .vscode/mcp.json:

{
  "servers": {
    "onecontext": {
      "type": "stdio",
      "command": "node",
      "args": ["scripts/onecontext-mcp.mjs"],
      "cwd": "${workspaceFolder}"
    }
  }
}

In VS Code with an MCP-capable chat/agent:

1. Open the repository folder.

2. Open the Command Palette with Ctrl+Shift+P.

3. Run MCP: List Servers.

4. Start or restart onecontext.

5. In the agent's Configure Tools menu, enable the OneContext tools.

6. Ask:

   Use the OneContext MCP tool onecontext_get_context before answering.
   What is the team currently working on, and which files should I avoid changing?
   

The expected result is a context block containing the same project goal, recent team activity, decisions, and source information visible to other connected agents.

If the server reports fetch failed, confirm that the host web app is running, the gateway URL in .env.local is reachable, and the gateway key is the same on both devices. If it reports Unexpected end of JSON input, restart the MCP server and verify that the repository contains the latest scripts/onecontext-mcp.mjs.

From the repository root:

npm run onecontext:agent -- codex "What is our current project goal and sprint focus?"

This prints a context-enriched prompt. The safe default is a preview, so you can review the context before sending it to an agent.

This demonstrates that two devices use the same project memory.

Find the Wi-Fi IPv4 address with ipconfig. In the original local demo it was 192.168.31.79; use your current address if it has changed.

From the repository root, run:

npm run dev -- -H 0.0.0.0

In a second terminal:

npm run realtime

Allow Node.js through Windows Firewall on private networks if Windows asks. Keep both terminals running.

Clone the repository or copy the project folder, install Node.js 18+ and run:

cd C:\path\to\Onecontext
npm install

Laptop 2 should not run npm run dev or npm run realtime for this test. Its .env.local should point to Laptop 1:

ONECONTEXT_API_URL=http://192.168.31.79:3000
ONECONTEXT_REALTIME_HTTP_URL=http://192.168.31.79:8787
ONECONTEXT_PROJECT_ID=atlas-project
ONECONTEXT_GATEWAY_KEY=<the-same-private-key-as-laptop-1>

Also set the VS Code extension settings on Laptop 2:

onecontext.apiBaseUrl = http://192.168.31.79:3000
onecontext.realtimeUrl = ws://192.168.31.79:8787/live

Both laptops must use the same project ID and Team Code. On Laptop 2, install the VSIX and join the Team Code. To prove shared memory, publish a handoff or task on Laptop 1, then ask the MCP-enabled Copilot/agent on Laptop 2 to retrieve it with project_id set to atlas-project.

If Laptop 2 cannot open the dashboard, test the connection from Laptop 2:

Test-NetConnection 192.168.31.79 -Port 3000
Test-NetConnection 192.168.31.79 -Port 8787

Both tests should show TcpTestSucceeded : True. Make sure both devices are on the same private Wi-Fi network and that the host firewall permits ports 3000 and 8787.

OneContext is designed to store useful project knowledge, not indiscriminately copy private chats.

• Keep memory project-scoped.
• Do not store API keys, passwords, JWTs, OAuth tokens, private keys, or .env contents.
• Do not store unrelated personal conversations.
• Prefer concise decisions, tasks, summaries, and file references over raw verbose transcripts.
• Keep DATABASE_URL, OPENAI_API_KEY, SESSION_SECRET, and ONECONTEXT_GATEWAY_KEY in .env.local only.
• The Chrome extension stores its connection settings locally and does not receive the server's database credentials.
• The VS Code extension stores the gateway key in VS Code Secret Storage.
• Team Codes are suitable for a demo, not a complete production authorization system.

Before a public deployment, add production authentication/authorization, rotate all demo credentials, use HTTPS/WSS, deploy the realtime service, and use Redis or another durable presence store.

OpenAI-powered routing and answers are optional. If enabled, the key stays on the server:

OPENAI_API_KEY=<your-key>
ONECONTEXT_USE_AI_ROUTING=true
ONECONTEXT_USE_AI_RETRIEVAL=true
ONECONTEXT_USE_AI_ANSWERS=true
ONECONTEXT_USE_AI_CONVERSATION_MEMORY=true
ONECONTEXT_AI_MODEL=<your-supported-model>

If an AI request fails, the application falls back to local routing and retrieval. The project can therefore be demonstrated without additional API usage or cost by leaving the key empty.

Local presence works in memory. For a Redis-backed realtime setup, use the included compose file if Docker Desktop is installed:

docker compose -f docker-compose.live.yml up

Then set REDIS_URL in .env.local to the Redis connection string you are using.

From the repository root:

npm test
npm run build
npm run vscode:compile

From apps/vscode-extension:

npm run compile
npm run package

The repository test suite covers retrieval and context behavior. npm run package creates the installable VSIX.

What is the team's current goal, which decisions are already recorded, and which files should I avoid changing?

Use OneContext: Complete Task and Save Handoff:

Title: Shared memory privacy decision
Summary: Keep project memory project-only. Store concise decisions, handoffs, tasks, and file references, but never store secrets, credentials, unrelated chats, or raw verbose transcripts.
Next task: Verify that the MCP-enabled agent on Laptop 2 can retrieve this handoff with project_id atlas-project.

Use the OneContext MCP tool onecontext_get_context before answering.
What privacy decision did my teammate save, and what is the next task?

ChatGPT:

We are building a shared project memory for a four-person AI-assisted team. Propose the next feature and identify the important project decisions that should be shared with another AI assistant.

Click Add context, then open Claude and ask:

Continue the OneContext project from the shared memory. Turn the previous planning into a concrete implementation checklist and mention the relevant files.

Click Add context in Claude. The result demonstrates provider-independent project continuity.

Codex with GPT-5.6 was used throughout the development of OneContext as the primary coding and reasoning partner. It helped transform the initial product specification into a working architecture, generate the Next.js and PostgreSQL implementation, build the Chrome and VS Code extensions, create the realtime team-presence service, and implement the MCP integration for AI coding agents.

Codex also accelerated debugging and iteration. It helped diagnose local-network issues between two laptops, VS Code extension activation problems, stale VSIX packages, MCP initialization failures, and authenticated handoff errors. It also helped create tests, improve the dashboard UI, write documentation, and prepare the final demo workflow.

Key technical decisions made during development included:

• Use provider-agnostic shared memory instead of directly synchronizing ChatGPT, Claude, or Codex conversations.
• Use PostgreSQL for persistent project memory and activity history.
• Use MCP as the bridge between OneContext and AI coding assistants.
• Use Team Codes and realtime presence for multi-device collaboration.
• Store concise project decisions, handoffs, and tasks instead of raw private chats.
• Add privacy protection so secrets, credentials, and unrelated conversations are not stored.
• Use conflict warnings as coordination signals rather than hard file locks.

The final system combines a web dashboard, Chrome extension, VS Code extension, realtime collaboration, persistent memory, and MCP-based context retrieval.

Use this repository as the code URL and select Developer tools as the category. The submission should include:

1. A concise description of the shared-memory problem and the OneContext solution.
2. A public demo video shorter than three minutes with audio explaining the product, Codex, and GPT-5.6 usage.
3. This repository URL and the setup instructions above.
4. The packaged VSIX path: apps/vscode-extension/onecontext-vscode-0.1.2.vsix.
5. Chrome extension installation instructions from this README.
6. A test path: local dashboard, demo account if retained, Team Code demo, and MCP retrieval prompt.
7. The Codex session ID returned by /feedback for the session where most core functionality was built.
8. The exact section How Codex and GPT-5.6 Accelerated the Build pasted into the Devpost submission.

The local repository cannot submit the external Devpost form automatically. Copy that section from this README, or use the ready-to-paste file DEVPOST_SUBMISSION.md.

src/app/                    Next.js pages and API routes
src/lib/                    Storage, retrieval, AI, auth, and live helpers
infra/db/schema.sql         PostgreSQL schema
browser-extension/          Chrome extension for ChatGPT and Claude
apps/vscode-extension/      VS Code source, README, and packaged VSIX
apps/realtime-sync/         WebSocket presence service
scripts/                    DB setup, Codex wrapper, MCP server, demo helpers
.vscode/mcp.json            VS Code MCP server configuration

This is a working local MVP intended for evaluation and demonstration.

• GitHub indexing is suitable for the demo and may need production hardening for large/private repositories.
• Team Codes and gateway keys need stronger multi-tenant authorization for a public service.
• Localhost works on one computer; teammates on another network need a deployment or secure tunnel.
• The conflict radar warns about overlap; it does not replace Git branches, code review, or merge tooling.

OneContext is released under the MIT License. See LICENSE for the full terms.