Devour

Feed your AI real docs so it stops repeating the same mistakes.

Why Devour exists

I built Devour because AI tools kept drifting away from official docs, then repeating wrong patterns in later prompts.

Devour fixes that loop with a local-first workflow:

1. scrape official docs,
2. keep them in your project,
3. search and ask against that local corpus,
4. sync updates when sources change.

What works today

• devour init creates a complete local workspace
• devour scrape supports URL, language-specific docs, local files, GitHub repos, OpenAPI specs, and --sources
• devour get supports language/framework shortcuts (now including Next.js, Svelte, Angular, Remix, SolidJS, Express)
• devour query is now functional (local lexical index)
• devour ask is hybrid local-first with live fallback when local relevance is weak
• devour sync updates configured sources and reindexes
• devour status reports real local health/statistics
• devour push <path> imports local docs into the workspace and reindexes
• devour serve local mode runs JSON-RPC over stdio (devour_query, devour_status, devour_scrape, devour_ask, devour_sync)
• devour auto routes natural-language intent to the right command
• devour verify smoke runs opt-in live checks and writes reports

Experimental

• devour serve --remote is available as an experimental HTTP RPC mode
• remote push workflows are intentionally not enabled as stable behavior yet

Quick start

1) Build

go mod download
go build -o devour ./cmd/devour

2) Initialize

./devour init

3) Get docs fast

./devour get go net/http --format markdown
./devour get nextjs routing
./devour get express middleware

4) Search locally

./devour query "json unmarshal"

5) Ask a docs-grounded question

./devour ask --lang go "how to parse json" --format text

6) Let Devour route intent automatically

./devour auto "how to use useEffect in react"
./devour auto "https://pkg.go.dev/net/http"

Core commands

Command	Purpose
devour init [path]	Create config + local storage directories
devour get <language> <keyword>	Shortcut fetch from curated official docs
devour scrape <source>	Scrape one source directly
devour scrape --sources sources.yaml	Scrape multiple configured sources
devour query <text>	Search local indexed docs
devour ask --lang <lang> <question>	Structured answer using local-first + live fallback
devour sync	Sync configured sources and rebuild index
devour status	Show docs/index/source health
devour push <path>	Import local docs into workspace
devour serve	Start local stdio JSON-RPC server
devour auto "<intent>"	Auto-route intent to command
devour verify smoke	Live smoke verification report
devour quality ...	Code quality scan/triage/fixes

Supported get / ask languages and frameworks

• Go (go, golang)
• Rust (rust)
• Python (python, py)
• Java (java)
• Spring (spring)
• TypeScript (typescript, ts)
• React (react)
• Vue (vue)
• Nuxt (nuxt)
• Docker (docker)
• Cloudflare (cloudflare, cf)
• Astro (astro)
• C# (csharp, cs)
• Kotlin (kotlin, kt)
• PHP (php)
• Ruby (ruby, rb)
• Elixir (elixir, ex)
• Next.js (next, nextjs)
• Svelte (svelte)
• Angular (angular, ng)
• Remix (remix)
• Solid (solid, solidjs) via github.com/solidjs/solid-docs
• Express (express, expressjs)

Run devour languages for examples, or devour languages --format json for automation.

Config

Devour reads devour.yaml (or --config).

New additive sections:

• indexing: local lexical index defaults
• verification: live smoke timeout defaults

Starter config: devour.example.yaml.

Real-world verification

Run live smoke checks (opt-in):

./devour verify smoke

Reports are saved to devour_data/verify/smoke-<timestamp>.json.

JSON-RPC local server

devour serve (local mode) accepts JSON-RPC 2.0 methods:

• devour_query
• devour_status
• devour_scrape
• devour_ask
• devour_sync

Development

Test

go test ./...

Typical integration loop

./devour init
./devour scrape https://pkg.go.dev/net/http --type godocs
./devour query "http client"
./devour ask --lang go "timeout example"
./devour sync
./devour status

License

MIT (LICENSE).