• seer checks each command itself before running it. It doesn't just take the model's word that a command is safe. In auto mode only an allowlist of read-only commands runs without asking. Commands like find -exec, xargs, sed and awk count as read-only only when what they run or edit is also read-only.
• The model also marks commands it knows will change something, and those always ask. The model can add confirmations but never skip them, because text in a file or a web page could try to talk it into something harmful.
• Commands run in your current directory with no input and a 60s timeout, for at most 6 steps. Ctrl+C stops at any point.
• seer --no-brave <question> turns brave mode off for one question. seer help and seer do never use it.
• Piped input (git log -5 | seer -b …) is read to the end before brave mode starts, however long the command takes. For a live stream (tail -f … | seer), use watch mode: no -b.
• Brave mode needs a capable model. It was tested with Claude Sonnet. Very small local models (~2B) tend to lose track after a few steps.

If you have Claude Code or Codex installed and signed in, seer can use it instead of an API key. This is off by default:

provider: claude-cli     # use it always…
auto_cli: true           # …or only when no local server is running (claude-cli, then codex-cli)

providers:
  claude-cli:
    type: claude-cli
    model: sonnet        # haiku / sonnet / opus, or a full model id
  codex-cli:
    type: codex-cli
    model: [gpt-6.1-sol, auto]  # tried in order; auto = your ~/.codex/config.toml model
    reasoning_effort: low

• seer runs the official claude -p / codex exec binary, which signs in with your own account. seer never touches your credentials.
• The CLI runs with its tools switched off, so it only answers. When brave mode runs commands, seer runs them, never the CLI.
• Requests count against your plan's limits. claude-cli answers in a few seconds; codex-cli takes about 10s.
• This use is subject to Anthropic's and OpenAI's terms. If you want no ambiguity, use the anthropic / openai providers with an API key.

provider: myserver
providers:
  myserver:
    type: openai                     # any OpenAI-compatible API
    base_url: http://myserver:8080/v1
    api_key: none
    model: my-model                  # or auto: the first model the server lists
    max_tokens: 16000                # optional: output cap per call, reasoning included
    reasoning_effort: low            # optional: only for servers and models that support it

Reasoning models think before they answer. While they do, the status line counts it (thinking… (reasoning, 1,234 words)), and max_tokens stops one that would think indefinitely: seer then says it hit the limit instead of waiting. If a server rejects the default cap because the model's context window is smaller (vLLM with an 8K model, say), seer drops it and retries; a max_tokens you set is always sent.

Specific version or branch (any git ref):

curl -fsSL https://raw.githubusercontent.com/zeddius1983/seer/main/install.sh | bash -s -- --version v1.3.0

Different Ctrl+G key (zsh bindkey notation, e.g. ^@ for Ctrl+Space):

curl -fsSL https://raw.githubusercontent.com/zeddius1983/seer/main/install.sh | SEER_IMPLICIT_BIND='^@' bash

From source: git clone https://github.com/zeddius1983/seer && cd seer && uv tool install .

Uninstall:

curl -fsSL https://raw.githubusercontent.com/zeddius1983/seer/main/install.sh | bash -s -- --uninstall

The installer backs up ~/.zshrc before changing it (~/.zshrc.YYYYMMDD_HHMMSS.bak).

Development:

uv sync                  # dev environment
uv run pytest            # tests
.venv/bin/seer --help    # run without installing