LLM API Scope (apiscope)

a tool for LLM agents to read and cache structured documents from remote.

install

use pipx for isolated installation:

pipx install llm-api-scope

how it works

apiscope has two ideas: a registered source, and a tree you navigate by address. text is the default output; --json prints the data layer instead.

manage

apiscope add registers a source with a name, a location, and a type. apiscope remove deletes it. apiscope list shows what is registered, filtered by type.

a source link is registered once; a repeated link fails, even under another name or type. a private source in the home configuration uses an absolute path, and a project source uses a path relative to the project root.

six document types share one command surface:

• filesystem reads a local file or directory
• repo reads a directory inside a git repository
• openapi reads an OpenAPI specification
• rfc reads an IETF document
• llmstxt reads a site index that lists documentation pages
• arxiv reads one arXiv paper by its canonical identifier

apiscope add docs https://example.test/docs/llms.txt --type llmstxt
apiscope add api https://github.com/example/api.git/docs@main --type repo
apiscope add paper 1706.03762 --type arxiv
apiscope list all

use

apiscope sync fetches sources into a local cache. apiscope view shows the cached structure, and apiscope read reads one node.

view and read share the same address: a source name plus an optional route.

apiscope sync all
apiscope view docs
apiscope read docs 1.2

every source projects into one tree. an ordinary node has children; a leaf has none. a view line is - [index] key: description, where the index is a tree address. an index from view always works with read.

skill

apiscope skill prints or installs a combined command reference and strategy guide for AI agents. run apiscope skill show to print it, or apiscope skill install to install it under ~/.agents/skills/apiscope.

agents read this once at onboarding instead of running --help repeatedly.

configuration

apiscope reads three configuration layers in order. the global file is ~/.apiscope/config.json. the project file is .apiscope/config.json, and the local file is .apiscope/local.json. --global uses the global layer only and skips project discovery.

the public setting holds the default source ttl in days. the local setting holds the proxy and the no_proxy bypass list.

{
  "setting": {
    "public": {"doc_ttl": 7},
    "local": {
      "proxy": "http://proxy.example.test:8080",
      "no_proxy": "internal.example.test,localhost"
    }
  }
}

the HTTP client uses only the configured [proxy] and [no_proxy] and ignores environment proxy variables. for repository operations, apiscope passes configured proxy settings to git and gh; when [proxy] is unset, it does not set or clear their proxy. [no_proxy] is a comma-separated list of hosts, domains, or * that bypasses the configured proxy.

the cache lives in the app directory of the selected layer, and the rest of its rules sit under ## commands.

commands

add

register a source:

apiscope add <name> <source> --type <type> [--ttl <days>]

arxiv accepts a canonical identifier such as 1706.03762 or hep-th/9901001; append vN for a fixed version. URL forms are not accepted.

remove

delete a source:

apiscope remove <name>

list

list registered sources:

apiscope list <selector> [--limit <n>] [--offset <n>]

the selector is one of all, filesystem, repo, openapi, rfc, llmstxt, or arxiv.

sync

fetch sources into the cache:

apiscope sync <selector> [<name>] [--force]

the selector is the same as in list, and the optional name narrows the range to one source. a source refreshes when its cache is older than its ttl, and --force ignores the ttl. repo clones with shallow depth, and a subpath selects files with a sparse checkout. llmstxt reads the index page, caches the markdown version of each page, and skips a page that only serves HTML. arxiv prefers structured HTML, falls back to TeX source, and does not parse PDF-, PostScript-, or DVI-only submissions.

view

show the cached structure of an address:

apiscope view <address> [--depth <n>]

depth caps the levels below the scope; the default is unlimited.

read

read one node:

apiscope read <address> [<index>]

the index is optional when the address already points at a leaf.

bookmark

save a named reference and reopen it later:

apiscope bookmark add <id> <mode> <target>... --description <text>
apiscope bookmark list [<group>]
apiscope bookmark use <id>
apiscope bookmark remove <id>
apiscope bookmark prune [--invalid]

mode is file, view, read, or group; a group takes five to nine member ids. bookmarks live in a global file and a project file, and the project wins.

skill

print or install the agent skill:

apiscope skill show
apiscope skill install [<target>]

install defaults to ~/.agents/skills/apiscope.

future

• read academic papers from arxiv
• more formal document formats as the need arises

license

MIT