Anya. Headless LLM agent runner: scheduled jobs, skills/flows, email reports, blotter. Read-only in intent—no destructive actions. Supports Anthropic Claude and OpenAI-compatible APIs (e.g. mlx_lm.server).

Your Anya has jobs. Each job is a DAG that combines Deterministic Actions (D) and (multimodel) LLM inference (I). The general shape is D → I → D: deterministic fetch/gather, then inference, then deterministic output handling. Potentially destructive actions stay out of the I stages—the LLM produces text only; it never has direct tool-calling. Anya's executor enforces an air gap: inference output flows through a controlled pipeline (blotter, email, memory) rather than to arbitrary shell, file, or browser actions. This contrasts with agent frameworks like OpenClaw, where the LLM directly invokes tools—broad permissions plus prompt injection or malicious skills can chain into real-world impact. Anya's design reduces that surface: the executor is the single, auditable boundary between inference and the outside world.

Anya (air-gapped):

flowchart LR
    D1["D: Fetch<br/>fetch.py, rss"]
    I["I: LLM<br/>text only"]
    D2["D: Output<br/>blotter, email"]
    D1 --> I --> D2

Loading

Typical agent (e.g. OpenClaw):

flowchart LR
    I2["LLM"] --> T1["shell"]
    I2 --> T2["file I/O"]
    I2 --> T3["browser"]
    I2 --> T4["messaging"]

Loading

Install

uv init
source .venv/bin/activate
uv pip install -U .

Quick start

1. Set env: ANTHROPIC_API_KEY (for Claude), or for OpenAI-compatible: LLM_PROVIDER=openai, LLM_MODEL=<model>, LLM_BASE_URL (default http://localhost:8080/v1). Also UNOSEND_API_KEY, UNOSEND_FROM (e.g. Anya <anya@yourdomain.com>)
2. Create jobs in job/ — one dir per job, each with MAIN.md
3. Run once: anya run --email_to=you@example.com
4. Or serve (daily): anya serve --email_to=you@example.com --interval=86400
5. Run example job too: anya run --phases=default,ignore

In reality I'd use 1password, so

op run --no-masking --env-file=.env -- anya run --email_to=uche@ogbuji.net 

LLM providers

Provider	Env	Capability
anthropic (default)	ANTHROPIC_API_KEY, optional LLM_MODEL	Full support; future features (tools, extended thinking) when added
openai	LLM_PROVIDER=openai, LLM_MODEL (required), LLM_BASE_URL or OPENAI_API_BASE (default http://localhost:8080/v1)	Core text-in/text-out; no Anthropic-specific features

Example with mlx_lm.server:

# Terminal 1: start mlx server
mlx_lm.server --model mlx-community/Mistral-7B-Instruct-v0.3-4bit

# Terminal 2: run Anya
LLM_PROVIDER=openai LLM_MODEL=mlx-community/Mistral-7B-Instruct-v0.3-4bit anya run --email_to=you@example.com

CLI overrides: --provider, --model, --llm_base_url.

op run --no-masking --env-file=.env -- anya run --provider=openai --llm_base_url=http://localhost:8080/ --model=mlx-community/Qwen3-Coder-30B-A3B-Instruct-4bit --email_to=uche@ogbuji.net

Job structure

job/
  my-job/
    MAIN.md      # Core prompt, instructions, optional fetch: / rss: lines
    .env         # Optional; per-job env (API keys, etc.)
    fetch.py     # Optional; runs before Claude, stdout → context
    *.py         # Any .py files run (read-only data gathering)

Use anya.fetchers in job .py files for pluggable web fetching (HTML→Markdown):

import asyncio
from anya.fetchers import fetch_url, create_fetcher, FetchResult

async def run():
    result = await fetch_url('https://example.com')
    print(result.markdown)  # FetchResult has .markdown, .title, .success

    # For bot-blocked sites (e.g. Reddit): use crawl4ai
    fetcher = create_fetcher('crawl4ai')  # or 'plain' for simple HTTP
    result = await fetcher.fetch('https://reddit.com/...')
asyncio.run(run())

Fetch methods: plain (default) — simple HTTP + ogbujipt; crawl4ai — Crawl4AI service for JS-heavy or bot-blocked sites. For crawl4ai, run docker run -p 11235:11235 unclecode/crawl4ai:basic and optionally set CRAWL4AI_BASE_URL.

More fully:

docker run -d -p 11235:11235 --name crawl4ai --shm-size=1g unclecode/crawl4ai:latest

MAIN.md

• id: optional; overrides dir name for blotter/email (default: dir name)
• phase: default (default) or ignore; ignore jobs are skipped unless --phases includes them
• frequency: daily (default), weekly, sundays, saturday, weekday
• fetch: https://... — fetch URL, body → context
• rss: https://... — RSS/Atom feed → context
• Inline actions — ---ACTION--- blocks expand before the LLM runs: ---ACTION---\nfetch('https://example.com/')\n---END ACTION--- → replaced with page content as Markdown
• Rest is the prompt for the LLM

Blotter & memory

• Blotter (data/blotter.txt by default): append-only log for review. Set BLOTTER_FILE env or --blotter CLI to share with other agent systems. Uses file locking ({blotter}.lock); BLOTTER_LOCK_TIMEOUT (default 30s) — if lock cannot be acquired, reports a system issue for you to investigate (e.g. stale lock file)
• Memory (data/memory.txt): long-term; the LLM can append via ---MEMORY--- block, or prune resolved issues via ---RESOLVED--- block

Scheduler

Default: asyncio loop (anya serve --scheduler=asyncio). Optional: uv pip install anya[scheduler-apscheduler] then anya serve --scheduler=apscheduler. Scheduler is modular — implement anya.scheduler.base.Scheduler to plug in cron, python-crontab, schedule, etc.

Conventions

See AICONTEXT-PYLIB.md for Python style and tooling.