Skip to content

AI-assisted development

Ghuri is new, so the models behind coding agents have not seen it in training. Left alone they fall back to what they know: Python Django, DRF, Prisma or raw Hono. Ghuri ships four things that give an agent the real APIs instead. Use as many as your tool supports; they stack.

Layer What it gives the agent Works with
AGENTS.md Project layout, commands and the core API cheat-sheet, in every new project Every agent that reads AGENTS.md or CLAUDE.md
Ghuri skill Step-by-step workflows and per-area reference, loaded only when relevant Claude Code (plugin), Codex, Cursor, Gemini CLI and others (npx skills)
ghuri mcp Live access to your project: models, routes, settings, migrations, management commands, docs search Any MCP client
llms.txt The whole documentation as one Markdown file Anything that can fetch a URL

ghuri startproject (and npm create ghuri) writes three files for coding agents:

File Purpose
AGENTS.md Project layout, the ghuri commands, a cheat-sheet of the core APIs (models, QuerySets, serializers, viewsets, admin, URLs and views) and the rules an agent should follow, such as running makemigrations after every model change and never touching D1 or R2 bindings from app code
CLAUDE.md A one-line @AGENTS.md import, so Claude Code reads the same file
.mcp.json Registers the ghuri mcp server for Claude Code, using your package manager (npx, pnpm exec, bunx or yarn)

Codex, Cursor, GitHub Copilot, Gemini CLI and most other agents read AGENTS.md automatically. Edit it as your project grows: add your apps, your domain vocabulary and your own conventions.

For a project created before these files existed, run ghuri startproject into a temporary directory and copy them over.

The skill teaches an agent Ghuri in depth: a Django-to-Ghuri mapping, step-by-step workflows (add an app, add a model and migration, expose a REST API, register it in the admin, add GraphQL or WebSockets, deploy, move from D1 to Postgres), the mistakes a Django-trained model makes, and one reference file per area. The agent loads it only when the task involves Ghuri, so it costs nothing the rest of the time.

Claude Code — install the plugin, which bundles the skill and the ghuri mcp server:

/plugin marketplace add ghurijs/ghuri
/plugin install ghuri@ghuri

Codex, Cursor, Gemini CLI, OpenCode and others — install the skill with the skills CLI:

Terminal window
npx skills add ghurijs/ghuri

The skill lives in the ghurijs/ghuri repository under plugins/ghuri/skills/ghuri/.

ghuri mcp runs a Model Context Protocol server over stdio that gives an agent live access to your project, so it checks your real models and routes instead of guessing. See ghuri mcp in the CLI reference for every option.

Tool What it does
project_info Project root, runtime, bindings, DEBUG, INSTALLED_APPS, DATABASES and STORAGES (engine and name only), MIDDLEWARE and @ghuri/* package versions. Secrets are hidden
list_apps Each installed app: label, path, models, commands, migrations
list_models Every app.Model with its table and fields
describe_model Fields with their options, relations, reverse relations, Meta and managers
list_routes Every URL pattern with its name and view, like ghuri routes
list_commands Every management command, including those added by apps
show_migrations Applied and unapplied migrations per app
openapi_schema Your REST API’s real OpenAPI document: a list of operations by default, or the JSON for a path prefix or tag
run_command Runs a management command (makemigrations, migrate, test, check, startapp, …) and returns its output
search_docs, get_doc Search and read this documentation (from llms-full.txt)

run_command will not run commands that are interactive, long-running or aimed at production: runserver, shell, dbshell, deploy, flush, startproject, createsuperuser without --no-input, test --watch, and anything with --remote. Project state is re-read whenever a source file changes, so the agent sees models and routes it has just added.

New projects already register the server in .mcp.json. To add it elsewhere, run the command from the project directory:

Claude Code

Terminal window
claude mcp add ghuri -- npx ghuri mcp

Codex (~/.codex/config.toml, or .codex/config.toml in the project)

[mcp_servers.ghuri]
command = "npx"
args = ["ghuri", "mcp"]

Cursor (.cursor/mcp.json)

{ "mcpServers": { "ghuri": { "command": "npx", "args": ["ghuri", "mcp"] } } }

VS Code (.vscode/mcp.json)

{ "servers": { "ghuri": { "type": "stdio", "command": "npx", "args": ["ghuri", "mcp"] } } }

The documentation is published as plain Markdown for tools that read URLs:

URL Contents
/docs/llms.txt Index of the sets below
/docs/llms-full.txt All of the documentation (about 175k tokens)
/docs/llms-small.txt Abridged version (about 140k tokens)
/docs/_llms-txt/orm.txt, rest-framework.txt, admin.txt, auth.txt, agents.txt, deployment.txt One area each (7k–20k tokens)

The focused sets fit easily in a context window. Paste one into a chat, or point a tool such as Cursor’s @Docs at it.