grandma 0.2.2

Explain it like you'd tell grandma.

👵 grandma

The digest layer for noisy AI agents.

---

Your AI agent just wrote 2,000 words. You need the next step. grandma extracts it.

Pipe any LLM output through grandma and get a clean terminal card: what happened, the bottom line, what to do next. No API key needed if you use Claude Code.

pipx install grandma
grandma --demo

---

Try it in 30 seconds

# zero-config if you have Claude Code installed
echo "$(cat some-agent-output.txt)" | grandma

# local model via Ollama (no cloud key needed)
GRANDMA_MODEL_BACKEND=ollama GRANDMA_MODEL=llama3.1 grandma --demo

# any provider
GRANDMA_MODEL_BACKEND=openai GRANDMA_MODEL=gpt-4o-mini OPENAI_API_KEY=sk-... grandma --demo

---

Before / After

Raw agent output (~90 words, ~120 tokens):

I inspected the repository and found that the authentication flow now routes through
the new async session adapter. I updated three files, added one regression test, and
confirmed that the login path still returns the expected token shape. There is one
compatibility consideration: the adapter relies on asyncio.TaskGroup, so Python 3.10+
is required. Overall, this should reduce request latency by ~70%, but the deployment
notes should mention the runtime floor bump.

After grandma (default mode — ~30 words):

👵 grandma
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
📖 Second turn reviewing an auth refactor PR.

What happened: Auth moved to the async session adapter.
Bottom line:   Faster login path, but Python ≥3.10 required.
Do next:       Update deployment docs before shipping.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

After grandma (deep mode — full impact table):

👵 grandma
┌──────────────┬──────────────────────────────────────────────┐
│ 📖 story arc │ Second turn reviewing an auth refactor PR.   │
│ 👀 happened  │ Auth moved to async session adapter.         │
│ 🧶 changed   │ 3 files, 1 regression test added.            │
│ ✅ positive  │ Login latency reduced ~70%; token shape OK.  │
│ ⚠️  negative  │ Requires Python ≥3.10 (asyncio.TaskGroup).  │
│ ➖ neutral   │ API surface unchanged.                        │
│ 💡 net gain  │ Win — ship it with a runtime floor note.     │
│ 📋 actions   │ - Bump requires-python to >=3.10             │
│              │ - Update deployment docs                      │
└──────────────┴──────────────────────────────────────────────┘

---

Install

pip install grandma        # pip
pipx install grandma       # isolated (recommended)
uv tool install grandma    # uv
pip install git+https://github.com/ypollak2/Grandma.git  # from source

Requires Python ≥ 3.10.

---

Usage

echo "agent output..." | grandma              # default 3-line card
cat output.txt | grandma --mode deep          # full impact table
cat output.txt | grandma --json               # raw JSON verdict
cat output.txt | grandma --mode off           # passthrough (for hooks)
grandma --demo                                # try it with no input
grandma --demo --mode deep
export GRANDMA_MODE=deep                      # set default for session

---

The modes

Mode	Output	Best for
default	3-line card: happened / bottom line / do next	Most agent output
deep	Full impact table with positive/negative/neutral	PRs, arch decisions, refactors
off	Passthrough — original text unchanged	Hook pass-through control

---

💸 Why this matters: token savings

Every intermediate agent response fed into the next prompt costs tokens. Grandma compresses the signal.

	Words	Tokens (approx)
Raw LLM response	~400	~530
grandma card	~40	~55
Saving	~90%	~90%

Across a 20-turn session: 8,000–12,000 context tokens saved.

---

Model backends — bring your own

Grandma auto-detects the best available backend. No config needed to get started with Claude Code.

Priority	When	Backend
1	GRANDMA_MODEL_BACKEND set	Explicit provider
2	GRANDMA_MODEL / GRANDMA_API_KEY / GRANDMA_BASE_URL set	OpenAI-compatible
3	OPENAI_API_KEY set	OpenAI
4	GROQ_API_KEY set	Groq
5	GEMINI_API_KEY / GOOGLE_API_KEY set	Gemini
6	ANTHROPIC_API_KEY set	Anthropic SDK
7	(nothing set)	claude -p - (Claude Code subscription, no key)

Model names are fully dynamic — no hardcoded vendor strings in grandma. You choose the model; grandma uses whatever you set. For backends that don't require an explicit model (claude_cli, anthropic SDK), grandma lets the provider pick its own default.

Copy .env.example → .env and uncomment your provider:

cp .env.example .env

Ollama (local, no key):

GRANDMA_MODEL_BACKEND=ollama
GRANDMA_MODEL=llama3.1
GRANDMA_DEEP_MODEL=deepseek-r1

Groq (fast inference):

GRANDMA_MODEL_BACKEND=groq
GRANDMA_MODEL=llama-3.1-8b-instant
GRANDMA_DEEP_MODEL=llama-3.3-70b-versatile
GROQ_API_KEY=gsk_...

OpenAI:

GRANDMA_MODEL_BACKEND=openai
GRANDMA_MODEL=gpt-4o-mini
GRANDMA_DEEP_MODEL=gpt-4.1
OPENAI_API_KEY=sk-...

Gemini:

GRANDMA_MODEL_BACKEND=gemini
GRANDMA_MODEL=gemini-2.5-flash
GRANDMA_DEEP_MODEL=gemini-2.5-pro
GEMINI_API_KEY=AIza...

Any OpenAI-compatible provider (Together, Fireworks, LM Studio, etc.):

GRANDMA_MODEL_BACKEND=openai_compatible
GRANDMA_BASE_URL=https://your-provider.example.com/v1
GRANDMA_API_KEY=your-key
GRANDMA_MODEL=your-model-name

Custom subprocess (anything that reads stdin):

GRANDMA_MODEL_BACKEND=custom_command
GRANDMA_MODEL_COMMAND=ollama run llama3.1

---

IDE & agent integrations

Auto-install all detected tools:

./install.sh

Tool	How	Effect
Claude Code	Stop hook	Auto-card after every long response
Codex CLI	Stop hook	Auto-card after every long response
Gemini CLI	AfterAgent hook	Auto-card after every agent turn
Cursor	MCP + rules	Agent calls grandma_summarize automatically
Cline	MCP + rules	Agent calls grandma_summarize after tasks
Continue	MCP + slash command	On demand or automatic
Windsurf	MCP + rules	Agent calls grandma_summarize automatically
Zed	MCP (context_servers)	Tool in Zed AI panel
Goose	MCP extension	Tool in Goose sessions
Aider	pipe wrapper	Pipe aider output through grandma
OpenHands	pipe wrapper	Pipe headless output through grandma

MCP server — works with any MCP-capable IDE:

grandma serve

Add to .mcp.json:

{
  "mcpServers": {
    "grandma": { "command": "grandma", "args": ["serve"] }
  }
}

Tools exposed: grandma_summarize(text, mode, story_context) and grandma_summarize_json(...).

---

Troubleshooting

"No model configured for backend X" Set GRANDMA_MODEL=<model-name> in your .env. See .env.example for provider examples.

"claude CLI not found" Install Claude Code or set a different backend via GRANDMA_MODEL_BACKEND.

Output is not JSON / parsing error The model returned something unexpected. Try --mode off to see the raw response, then check your model/backend config.

Provider returns 401 / 403 Check that the right API key env var is set for your backend (e.g. OPENAI_API_KEY, GROQ_API_KEY).

---

How the mascot rotates

Each portrait in assets/grandmas/ was generated with Gemini Image. A daily GitHub Action picks one at random and replaces assets/grandma.png with a [skip ci] commit.

Want to add your own grandma? Drop a PNG into assets/grandmas/ and open a PR.

---

Contributing

See CONTRIBUTING.md for full details. Quick start:

git clone https://github.com/ypollak2/Grandma.git
cd Grandma
pip install -e ".[dev]"
pytest tests/        # run tests
ruff check src       # lint

Good first issues are tagged good first issue.

---

FAQ

Does grandma make it dumber? No dear. She keeps the facts. She removes the fog machine.

What model does it use? Whatever you configure via GRANDMA_MODEL. There are no hardcoded model names. If you set nothing, claude_cli lets Claude Code pick, and anthropic lets the SDK pick its default. API-based backends (openai, groq, ollama, gemini) require GRANDMA_MODEL to be set.

Why is there an off mode? Because hooks need a clean passthrough option. When GRANDMA_MODE=off, grandma becomes transparent — useful for temporarily disabling the Stop hook without unregistering it.

What's the story_so_far line? Grandma reads the last 3 turns of the conversation to track where you are in the arc. Turn 8 of the same auth bug? She'll say so. This is the dementia prevention feature.

Why Python ≥ 3.10? The mcp package (used for grandma serve) requires 3.10+.

How is this different from llm / fabric / shell_gpt? llm gives you model access. fabric gives you prompt patterns. shell_gpt gives you a REPL. grandma does none of those things. It sits at the output end — it is the digest layer you add after any of those tools to stop reading 2,000-word agent responses.

---

Changelog

See CHANGELOG.md.

---

_{Made with 💜 by Yali Pollak · MIT License · PyPI}