GitHub - turlockmike/infer: Pipe-friendly LLM agent harness with a bash tool. Works with any OpenAI-compatible provider.

infer "what directory am i in"
cat crash.log | infer "why did this fail"
infer -f main.py "explain this"
infer -j '{"name":"string","pid":0}' "current process info"

curl -fsSL https://raw.githubusercontent.com/turlockmike/infer/main/install.sh | sh

git clone https://github.com/turlockmike/infer
cd infer
bun install
bun run infer.ts "hello"

# OpenAI
export INFER_API_KEY=sk-...
export INFER_URL=https://api.openai.com/v1
export INFER_MODEL=gpt-4o

# Anthropic (via OpenAI-compatible proxy, e.g. LiteLLM)
export INFER_API_KEY=sk-ant-...
export INFER_URL=http://localhost:4000/v1
export INFER_MODEL=claude-opus-4-5

# Groq
export INFER_API_KEY=$GROQ_API_KEY
export INFER_URL=https://api.groq.com/openai/v1
export INFER_MODEL=llama-3.3-70b-versatile

# Local (Ollama)
export INFER_URL=http://localhost:11434/v1
export INFER_MODEL=gemma4:latest
export INFER_API_KEY=ollama

infer "what directory am i in"

infer [OPTIONS] [PROMPT]

  -m MODEL          Model name (default: gemma4:latest)
  -u URL            Provider base URL (default: http://localhost:11434/v1)
  -k KEY            API key (default: "ollama" for local providers)
  -s TEXT           System prompt override
  -r ROLE           Named role — loads ~/.config/infer/roles/<name>.md
  -f FILE           File to use as context (prepended to prompt)
  -i FILE           Attach an image (repeatable); requires a vision-capable model
  -j [SHAPE]        Output JSON, optionally validated against a shape
  -n N              Max tool-call iterations per run (default: 10)
  -v                Verbose — show tool calls and token stats on stderr
  --stream          Stream tokens as they arrive (default: off)
  --no-sandbox      Use real bash (default: sandboxed via sandbox-exec)
  --no-network      Block network access inside the sandbox (default: allowed)

cat logs.txt | infer "summarize the errors"
infer -f config.yaml "is this valid?"

# Start a session
infer -S /tmp/review.jsonl "here is the file" < src/auth.py

# Do something programmatic between turns
issues=$(run_linter src/auth.py)

# Continue the same conversation
echo "$issues" | infer -S /tmp/review.jsonl "the linter found these issues — which ones matter?"

# One more turn
infer -S /tmp/review.jsonl "now write the fix for the most critical one"

infer -S /tmp/review.jsonl repl
# infer repl  gemma4:latest  (6 messages loaded from /tmp/review.jsonl)

jq '.role + ": " + .content' /tmp/review.jsonl   # read conversation
cp /tmp/review.jsonl /tmp/review-fork.jsonl       # branch from this point

infer "list files here"                   # sandboxed, network allowed
infer --no-network "ls only, no curl"     # block network in sandbox
infer --no-sandbox "unrestricted bash"    # bypass sandbox entirely

# Plain JSON (any shape)
infer -j "current date and time"

# Object shape
infer -j '{"name":"string","pid":0}' "current process info"

# Array of strings
infer -j '["string"]' "list files in /tmp"

# Array of objects
infer -j '[{"file":"string","size":0}]' "files in /tmp with sizes"

infer -j '[{"name":"string","tags":["string"],"score":0}]' "top 3 processes"

infer -j '["string"]' "list files in /tmp" | jq '.[]'

# Describe a single image
infer -i frame.jpg "describe what you see"

# Compare two images
infer -i before.png -i after.png "what changed?"

# Extract structured data from a screenshot
infer -i invoice.png -j '{"total":0,"currency":"string"}' "extract the total"

# Stdin text + image
cat context.txt | infer -i chart.png "does this chart match the claim in the text?"

# Image-only invocation with a role that encodes the task — no prompt needed
infer -r security-camera -i frame.jpg

# Summarize recent commits
git log --oneline -20 | infer "summarize what changed this week"

# One-liner per merged PR
gh pr list --state merged --limit 10 --json number,title,mergedAt | infer "one sentence per PR"

# Review a PR diff
gh pr diff 42 | infer "what's the risk level of this change?"

# List tools on DeepWiki (GitHub repo documentation, no auth required)
murl https://mcp.deepwiki.com/mcp/tools | infer "what can this service do?"

# Read the architecture of any repo indexed on deepwiki.com
murl https://mcp.deepwiki.com/mcp/tools/read_wiki_structure \
  -d repoName=langchain-ai/langchain | infer "what are the main sections?"

infer -s "$(cat prompts/security-reviewer.md)" < src/auth.py

# 1. Write the prompt
mkdir -p ~/.config/infer/prompts
cat > ~/.config/infer/prompts/review.md << 'EOF'
You are a senior engineer doing a code review. Focus on: correctness, security, and performance.
Flag anything that could fail in production. Be direct. No praise.
EOF

# 2. Write the script
cat > ~/bin/review << 'EOF'
#!/bin/bash
infer -s "$(cat ~/.config/infer/prompts/review.md)" "$@"
EOF
chmod +x ~/bin/review

# Now use it anywhere
cat src/auth.py | review
gh pr diff 42 | review "is this safe to merge?"
git diff HEAD~1 | review "any regressions?"

infer config set url http://localhost:11434/v1
infer config set model gemma4:latest
infer config set api_key ollama
infer config show
infer config get model
infer config unset model

echo "Output only code. No explanation." > ~/.config/infer/roles/coder.md
infer -r coder "write a fizzbuzz in python"