LLM Providers

LLM Providers

Signal supports two LLM backends. Claude is the default and the only supported provider for automated (scheduled) runs.

---

Provider comparison

	Claude Code CLI	Ollama (qwen2.5:14b)
Cost	Claude Pro/Max subscription (no per-token billing)	Free (local compute)
Speed	~9 sec/article, ~22 min full run	~10 sec/article, ~28 min full run
Analysis quality	Significantly better — more nuanced entity extraction, sharper framing analysis	Good
Requires	Claude Code CLI installed + claude auth login	Ollama server running + model pulled
Safe for launchd	Yes	No — crashes macOS kernel via Metal GPU driver
Network	Requires internet (HTTPS to Anthropic API)	Fully offline
Model	Claude (version determined by CLI)	Configured in sources.yaml

---

Claude Code CLI (default)

How it works

Each LLM call in analyzer.py runs:

subprocess.run(
    ["claude", "-p", prompt, "--print"],
    capture_output=True, text=True, timeout=180
)

This spawns a new claude process per call, which makes an HTTPS request to Anthropic's API and returns the response to stdout. There is no persistent connection or session — each call is fully independent.

Installation

npm install -g @anthropic-ai/claude-code
claude --version   # 2.x.x (Claude Code)

Authentication

Claude Code CLI uses your Claude Pro or Max subscription:

claude auth login

No API key is required. Authentication is tied to your Anthropic account session.

Verify it works

echo "Say hello" | claude -p "Say hello" --print

If that returns a response, Pass 1 will work.

Why Claude is required for scheduled runs

When Ollama is invoked from a launchd background daemon, it accesses Apple Metal for GPU inference in a non-interactive GPU session context. This mismatch triggers a Metal driver kernel panic, which hard-locks the machine. Claude avoids this entirely — it's a network call with no GPU involvement.

See Troubleshooting for the full crash investigation.

---

Ollama (manual/interactive use only)

How it works

Each LLM call uses the ollama Python client:

client = ollama.Client(host=base_url)
response = client.generate(
    model=model,
    prompt=prompt,
    options={"temperature": 0.1, "num_predict": 2048},
)

This connects to the Ollama HTTP server running locally at http://localhost:11434.

Setup

# Install Ollama
brew install ollama

# Start the server
ollama serve

# Pull the model
ollama pull qwen2.5:14b

# Verify
ollama list   # should show qwen2.5:14b

Recommended models for M1/M2/M3 Mac

Model	VRAM	Speed	Quality	Use case
qwen2.5:14b	~9 GB	~10 sec/call	Best	Production (manual runs)
llama3.1:8b	~5 GB	~5 sec/call	Good	Development/testing
mistral:7b	~4 GB	~4 sec/call	Baseline	Quick smoke tests

Use Ollama for a manual run

# Start Ollama first
ollama serve

# Run with Ollama
SIGNAL_LLM_PROVIDER=ollama python3 main.py

Do not run Ollama via run_and_publish.sh or launchctl start — the Metal GPU crash will lock up your Mac regardless of available RAM.

---

Switching providers

Precedence (highest first)

1. SIGNAL_LLM_PROVIDER environment variable
2. llm.provider in config/sources.yaml
3. Hard-coded fallback: "ollama"

Option 1 — Environment variable (one-off override)

# Force Claude
SIGNAL_LLM_PROVIDER=claude python3 main.py

# Force Ollama
SIGNAL_LLM_PROVIDER=ollama python3 main.py

Option 2 — sources.yaml (persistent default)

llm:
  provider: claude # change to "ollama" for local inference

Option 3 — launchd plist (scheduled run default)

The plist already sets SIGNAL_LLM_PROVIDER=claude. To change:

<key>EnvironmentVariables</key>
<dict>
    <key>SIGNAL_LLM_PROVIDER</key>
    <string>claude</string>   <!-- or "ollama" — but see crash warning above -->
</dict>

After editing scripts/com.flexrpl.signal.plist:

launchctl unload ~/Library/LaunchAgents/com.flexrpl.signal.plist
cp scripts/com.flexrpl.signal.plist ~/Library/LaunchAgents/
launchctl load ~/Library/LaunchAgents/com.flexrpl.signal.plist

run_and_publish.sh default

The script sets SIGNAL_LLM_PROVIDER explicitly so it behaves identically to launchd when run manually:

export SIGNAL_LLM_PROVIDER="${SIGNAL_LLM_PROVIDER:-claude}"

This means claude is the default, but you can still override it:

SIGNAL_LLM_PROVIDER=ollama bash scripts/run_and_publish.sh