@prof-bilal/atlas-chatgpt-plugin

A small ChatGPT / Codex plugin that exposes CodeAtlas repository intelligence over the Model Context Protocol (MCP). It is a thin adapter: it reuses @prof-bilal/atlas-mcp's handlers, validation, budget, freshness, and result discipline, and re-reads nothing itself — every answer comes from the CodeAtlas index through createContextSDK.

Understand your codebase before you change it.

Tools

Five focused, goal-oriented tools (the raw eleven-tool CodeAtlas surface is deliberately not exposed, to keep assistant tool-routing small):

| Tool | Answers | Backed by | | --- | --- | --- | | analyze_repository | "Analyze this repo", structure, counts, areas, findings | MCP project_overview | | search_repository | "Where is X implemented?", relevant files/symbols | MCP search_symbols + search_files | | explain_repository | "How does X work?", architecture + relationships | MCP find_relevant_context + graph reads | | impact_analysis | "What breaks if I change X?" | MCP analyze_impact + graph reads | | get_context | "Show me the code for X" | MCP inspect_symbol / read_file_range |

Each tool declares zod input/output schemas and returns structuredContent. Repository summary, search, explanation, impact, and code-context results also bind to optional MCP Apps widgets — the tools are fully useful without UI.

Run it

Build once, and index the project you want to serve:

pnpm --filter @prof-bilal/atlas-chatgpt-plugin build
atlas init          # or: atlas build   (from codeatlas-cli)

Codex CLI / local MCP clients (stdio)

codeatlas-chatgpt-mcp --root /path/to/your/project

Codex ~/.codex/config.toml:

[mcp_servers.codeatlas]
command = "codeatlas-chatgpt-mcp"
args = ["--root", "/path/to/your/project"]

ChatGPT (Streamable HTTP)

ChatGPT connectors need a reachable HTTPS URL, so run the server over HTTP and tunnel it during local development:

# 1. Start the HTTP server (loopback by default)
codeatlas-chatgpt-mcp --http --port 8787 --root /path/to/your/project

# 2. Expose it (any tunnel works; example uses cloudflared)
cloudflared tunnel --url http://127.0.0.1:8787

Then add https://<tunnel-host>/mcp as a connector in ChatGPT (developer mode). For anything other than loopback, require a token:

codeatlas-chatgpt-mcp --http --host 0.0.0.0 --port 8787 --token "$SECRET" --root /path/to/your/project

Health check: GET /healthz.

Programmatic use

import { createChatGptServer, startStdioServer, startHttpServer } from "@prof-bilal/atlas-chatgpt-plugin";

const plugin = createChatGptServer({ root: "/path/to/project" });
await plugin.connect(someTransport);

Packaging & submission

Static metadata for directory/app submission lives in plugin/:

• plugin/manifest.json — name, descriptions, developer, logo, categories, tool list, and MCP server config.
• plugin/mcp.json and plugin/codex.toml.example — ready-to-paste client config.
• plugin/icons/logo.svg, plugin/icons/logo-dark.svg — branding.
• plugin/PRIVACY.md — privacy policy.

See docs/reference/CHATGPT.md in the repository root for the manual OpenAI-dashboard steps (the dashboard does not read these files automatically).

Security

• Loopback by default; a bearer token is required for non-loopback binds.
• Read-only; no arbitrary filesystem access; the CodeAtlas secret deny-list is enforced on every read.
• No secrets are logged.

Limitations

• Runs only against a local CodeAtlas index (.codeatlas/context.db); there is no CodeAtlas cloud backend in this release.
• Parser coverage is TypeScript/JavaScript; other languages are indexed as content only.
• Widgets require a host that implements the Apps SDK; without it, tool output is still complete.