blitz

Fast one-off AI from the terminal, powered by OpenAI Codex subscription auth.

blitz is a small Go CLI for quick, non-agentic requests. It reads input from stdin or arguments, sends it to the Codex subscription endpoint, and prints the result without starting a bloated agent harness.

The binary starts quickly, has no runtime package manager, and uses only the Go standard library. It is designed for scripts, text filters, launchers, and other places where you want one model call and then exit.

Use your OpenAI/Codex subscription login:

blitz login

Open the printed URL, enter the code, and blitz saves OAuth tokens to:

~/.blitz/auth.json

blitz also falls back to an existing Codex CLI login at:

~/.codex/auth.json

So this also works:

codex login
cat input.txt | blitz

blitz is intentionally subscription-only. It does not support API keys or custom endpoints.

blitz "Tell me a joke"
printf '%s' "$LONG_TEXT" | blitz --summarize
cat notes.md | blitz --prompt "Extract action items. Output bullets only."
blitz status

Built-in defaults use gpt-5.4, reasoning off, and fast mode on. Codex requires streamed Responses requests, so blitz always streams internally and prints output as it arrives. See effective defaults with:

blitz
blitz status
blitz --help

Persist defaults in ~/.blitz/config.json with blitz config set:

blitz config set model gpt-5.4
blitz config set reasoning off
blitz config set fast true
blitz config set timeout 10m
blitz config unset reasoning

Supported config keys: model, codex-home, skills-dir, prompt, reasoning, max-output-tokens, timeout, and fast.

Skills are saved system prompts. Put Markdown files in ~/.blitz/skills; the filename without .md becomes a flag:

mkdir -p ~/.blitz/skills
cat > ~/.blitz/skills/summarize.md <<'EOF'
Summarize the input clearly. Keep only the most important points. Output bullets.
EOF

blitz --summarize "Long text goes here"

Use -skills-dir or BLITZ_SKILLS_DIR to point at a different skill folder.

A transcript cleanup workflow is just a skill:

cat > ~/.blitz/skills/transcript.md <<'EOF'
Improve this transcript for readability and accuracy. Fix obvious ASR mistakes,
punctuation, capitalization, speaker-flow issues, and paragraphing. Preserve
meaning. Output only the enhanced transcript.
EOF

blitz --transcript "Hello world ähm hmm here i am"

The user-facing abstraction is fast.

As of the current Codex implementation, Codex maps Fast mode to the Responses API request field:

{"service_tier":"priority"}

blitz follows that behavior: when fast is true, it sends service_tier: "priority". There is intentionally no user-facing service-tier setting; if Codex changes how Fast mode is represented, update this section and the codexRequestBody implementation.

Codex also currently builds Responses requests with stream: true, and the Codex backend rejects stream: false with Stream must be set to true. blitz therefore always sends stream: true and reads server-sent events (SSE).

Reference checked against openai/codex:

• ServiceTier::Fast.request_value() returns "priority".
• ServiceTier::Flex.request_value() returns "flex".
• ResponsesApiRequest is constructed with stream: true.
• Fast mode is feature-gated in Codex and defaults enabled there.

-model             model name, default gpt-5.4
-prompt            replacement system prompt
-skills-dir        directory of skill markdown prompts
-fast              request Fast mode via service_tier=priority; default true
-reasoning         reasoning effort, default off; use low/medium/high or off/none
-max-output-tokens optional output cap
-timeout           request timeout, default 10m
-codex-home        Codex home containing auth.json
-blitz-home        Blitz home containing auth.json and config.json

Environment variables:

BLITZ_MODEL
BLITZ_PROMPT
BLITZ_SKILLS_DIR
BLITZ_REASONING_EFFORT
BLITZ_HOME
CODEX_HOME

This repo was built with Go via mise:

mise x go@1.26.4 -- go test ./...
mise x go@1.26.4 -- go build -o blitz ./cmd/blitz

If Go is installed normally:

go test ./...
go build -o blitz ./cmd/blitz

MIT

The Codex subscription path follows the current OpenAI Codex and Pi patterns:

• Codex stores ChatGPT login tokens in $CODEX_HOME/auth.json.
• Pi documents ChatGPT Plus/Pro Codex subscription auth through /login, with tokens in ~/.pi/agent/auth.json.
• Both use ChatGPT OAuth tokens and call the Codex Responses endpoint at https://chatgpt.com/backend-api/codex/responses.

blitz keeps its own login tokens in ~/.blitz/auth.json, refreshes them when the access token is near expiry, and uses the Codex-compatible headers:

Authorization: Bearer <access_token>
chatgpt-account-id: <account_id>
OpenAI-Beta: responses=experimental
originator: blitz