maxac 0.1.0

A one-shot command-line helper with scoped execution tasks

agent-cli

Your terminal. Any task. One command.

ta "how much disk space do I have remaining"

You describe what you want. ac uses LLM intelligence where it matters — skill matching, tool selection, result verification — so the agent is robust to the variance that breaks classical heuristics. At default verbosity, you just see the answer. Add -v or -vv for internals.

---

Install

uvx/pipx/uv tool maxac

Then point it at any OpenAI-compatible API and ask you TA:

ac -s model "gemma4:9b"
ac -s base_url "http://localhost:11434/"

---

Try it now

# Run a one-shot task
ac "list all files in the current directory"

# No arguments → show current status (tools, skills, config)
ac

ac and agent-cli are interchangeable — ac is the short alias.

---

What happens when you run a task

1. Defines success — the LLM writes a concrete, directly verifiable success condition before touching anything (e.g. "output contains filesystem mount points and their total/used/available space" — not vague things like "output contains information about disk usage").
2. Matches skills — the LLM checks saved skills against your task. It rejects generic matches (a bare cat skill won't fire when you asked about disk space — the LLM knows df is the right tool).
3. Plans — the LLM picks the best tool for the job (df for disk space, free for memory, git for repos) — not limited to already-linked tools. Any tool on PATH is fair game.
4. Resolves tools — needed tools are symlinked automatically (with your approval). If a tool isn't on PATH, the LLM suggests an alternative before falling back to system search.
5. Executes step-by-step — each action is visible at -v; nothing runs silently.
6. Verifies — the LLM checks tool output against the success condition and produces a human-readable answer (e.g. "The CPU is AMD Ryzen 7 and the memory is 16GB").
7. Learns — the LLM names the skill, identifies variable arguments, and parameterizes the plan — all via LLM, not regex.

---

Tool isolation

Instead of reaching into your full system PATH, ac works with a minimal set of symlinked tools under its config directory:

~/.local/agent-cli/tools/
  doc/bin/    whatis  apropos  man  pydoc
  find/bin/   cat  head  tail  ls
  vcs/bin/    git          ← symlinked on first use, after you say yes
  build/bin/  npm  pip  …

When a task needs a new tool, the agent checks PATH first, then asks the LLM for an alternative if needed, then falls back to whatis/apropos as a last resort. It prompts you before symlinking:

  ? allow symlink: git → tools/vcs/bin/git  [y/N]

Use -y / --yes to pre-approve all symlinks for non-interactive runs:

ac -y "clone github.com/user/repo as my-repo"

---

Skills

Every successful task is saved as a skill — an Anthropic-compatible SKILL.md + plan.json pair under your config directory:

~/.local/agent-cli/skills/
  clone-repo/
    SKILL.md      # YAML frontmatter + instructions (importable to other tools)
    plan.json     # parameterized plan, params_map, success condition

The LLM identifies variable arguments (URLs, repo names, paths, versions) and gives them semantic parameter names (e.g. repository_url, branch_name) — so the same skill works on new inputs without re-planning. Skill matching is also LLM-driven: the model decides whether a saved skill genuinely applies to your task, rejecting false matches like a generic cat skill when you asked about disk space.

Skill commands

ac --skills                   # list all saved skills
ac --skills clone-repo        # show detail: instructions, plan, params
ac -d clone-repo              # delete a bad skill so it re-learns from scratch

---

Configuration

Config lives at ~/.local/agent-cli/config.json (path varies by platform — see table below).

Set values

ac -s model "gpt-4o"
ac -s base_url "https://api.openai.com/v1"
ac -s key "sk-..."
ac -s                          # show current values (key is masked)

base_url is auto-corrected — if it doesn't already end with /v1 or /v1beta, /v1 is appended:

ac -s base_url "https://integrate.api.nvidia.com"
# → stored as https://integrate.api.nvidia.com/v1

List available models

Run -m with no value to query the /models endpoint — useful for verifying your key and base_url:

ac -m
#   ✓ models available at https://api.openai.com/v1:
#   · gpt-4o  (openai)
#   · gpt-4o-mini  (openai)

One-shot overrides

Override model, base URL, or key for a single run without changing saved config:

ac -m "gpt-4o-mini" "summarise this repo in 10 bullets"
ac -b "https://my-proxy.example.com/v1" -k "sk-..." "list all files"

Debug: see the raw API call

ac --curlify "say hi"
# prints the equivalent curl command before executing

Config directory

Platform	Default path
Linux	~/.local/agent-cli/
macOS (framework build)	~/Library/Python/3.x/agent-cli/
macOS (non-framework)	~/.local/agent-cli/
Override	-c <path> / --config-dir <path>

---

Quick reference

Command	What it does
ac "<task>"	Run a one-shot task
ac	Show status (tools, skills, config)
ac -s model "gpt-4o"	Set default model
ac -s base_url "…"	Set default API base URL
ac -s key "…"	Set default API key
ac -s	Show current config values
ac -m	List models at current base URL
ac -m "model" "<task>"	Run task with a different model
ac -b "url" "<task>"	Run task with a different base URL
ac -k "key" "<task>"	Run task with a different API key
ac --skills	List saved skills
ac --skills <name>	Show skill detail
ac -d <name>	Delete a skill
ac -v "<task>"	Show sections and steps (-vv for raw tool output)
ac -y "<task>"	Auto-approve all tool symlinks
ac -c <path> "<task>"	Use a different config directory
ac --curlify "<task>"	Print the raw API call as curl

---

Contributing

If you've run a task and thought "that should just work" — open an issue with:

• what you typed
• what you expected
• what actually happened

PRs welcome.

---

License

MIT