Configuration

Most people never edit a file: Settings in the app covers models and keys. This page is for everything else.

Where things live

~/.local/bin/abstract  the program (when installed with the installer)

~/.abstract/
  config.json          keys, chosen model, port, recent workspaces, budgets
  chatgpt-auth.json    your ChatGPT sign-in (readable only by you)
  skills/              your skills, shared by every workspace
  rules/               your rules (optional)

<workspace>/
  sources/  drafts/  figures/  analysis/
  .openpaper/          library database, conversations, memory, reading notes

The standalone program also unpacks its interface once per version into your cache folder (~/Library/Caches/abstract on macOS, ~/.cache/abstract on Linux). It’s safe to delete.

config.json

{
  "providers": {
    "anthropic": { "apiKey": "sk-ant-…" },
    "ollama": { "baseURL": "http://localhost:11434" }
  },
  "roles": {
    "orchestrator": "anthropic/claude-sonnet-5-5"
  },
  "port": 4477,
  "recentWorkspaces": ["/Users/you/Abstract/thesis"],
  "budgets": {}
}

Field	Meaning
providers	API keys (apiKey) and the Ollama address (baseURL). Set these in Settings.
roles.orchestrator	The model picked in the chat box, as provider/model. It runs every task.
roles.embeddings	Optional: force an embedding model, e.g. "openai/text-embedding-3-small".
port	Default port (4477).
recentWorkspaces	Shown on the Workspaces screen; the first existing one opens by default.
budgets	Limits for long runs (see below).

If the file can’t be read, abstract starts with defaults and prints a warning.

Budgets

Long tasks are bounded by these settings, all optional. Set any of them to 0 to turn that limit off.

Setting	Default	What it does
turnSoftInputTokens	1500000	Past this many input tokens, the agent is asked once to pace itself and wrap up.
turnHardInputTokens	3000000	The ceiling for one segment of work.
maxAutoContinues	2	When the ceiling is hit with the plan still open, earlier steps are summarized and work continues in a fresh segment, up to this many times.
compactAtInputTokens	110000	When the conversation context grows past this, it’s summarized mid-turn to keep fitting the model. Raise it for long-context models.
nativeTaskBudgetTokens	0	Anthropic only: also send a native task budget to the model.

{ "budgets": { "turnSoftInputTokens": 800000, "maxAutoContinues": 0 } }

Independently of these, one message can run for at most 3 hours.

Environment variables

Variable	Purpose
ANTHROPIC_API_KEY, OPENAI_API_KEY, GOOGLE_GENERATIVE_AI_API_KEY, OPENROUTER_API_KEY	Provider keys. They take precedence over keys saved in Settings.
OLLAMA_HOST	The Ollama server address (default http://localhost:11434). Takes precedence over Settings.
UNPAYWALL_EMAIL	The email sent to Unpaywall when looking for open-access copies. Set your own to follow Unpaywall’s usage etiquette.
ABSTRACT_SKILLS_DIR	Use a different skills folder.
ABSTRACT_RULES_DIR	Use a different rules folder.
ABSTRACT_CHATGPT_STORE	Use a different file for the ChatGPT sign-in.
ABSTRACT_NO_PLAN	Set to anything to turn off the agent’s task plan.

Rules

Rules let you forbid specific agent actions. Put JSON files in ~/.abstract/rules/. Each file holds one rule or a list of rules:

{
  "name": "no-arxiv-fetch",
  "tool": "fetch_paper",
  "input_matches": "arxiv\\.org",
  "action": "block",
  "message": "this project forbids fetching from arXiv mirrors"
}

Field	Meaning
name	A label shown when the rule fires (defaults to the file name).
tool	The action to match: an exact tool name, * for all, or a prefix ending in *.
input_matches	Optional. A case-insensitive regular expression tested against the action’s input.
action	Must be "block". Rules can only forbid, never allow.
message	The reason. The agent is told this and works around it.

Rules are read at the start of each message. When one fires, the agent sees blocked by your rule "no-arxiv-fetch": … and takes another route. Common tool names: fetch_paper, search_scholar, snowball, run_code, edit_source, draft_section, export_draft, delegate.