Agent Smith

A Claude Code skill that offloads bulky, repetitive work to a cheaper model — then has Claude verify and finish it — so you spend Claude's tokens on judgment, not grunt work.

Gemini (or a local model) drafts the words: research, document digests, bulk extraction/transform, plans, marketing copy, config boilerplate, first-draft code. Claude scopes the task, cross-checks the output, and integrates it. The model drafts; Claude verifies. Nothing the model writes is treated as a deliverable until it's been reviewed.

Named for the Matrix agent who copies himself across the system — this skill fans heavy work out to a fleet of model "copies" while the One keeps the judgment.

What it's good for

• Web research with source links (Gemini's Google Search grounding)
• Digesting big inputs — summarize/extract/classify long PDFs, transcripts, logs, CSVs
• Bulk transforms — classify or rewrite many records
• First drafts — plans, proposals, marketing copy, config/IaC boilerplate, code + tests

It is not for short/interactive work, correctness-critical debugging, or the execution half of a task (deploying, committing, posting) — those stay with Claude or a script you control.

Install

As a plugin (recommended):

/plugin marketplace add negativetime/agent-smith-plugin
/plugin install agent-smith@agent-smith-marketplace

As a personal skill (no plugin system): copy plugins/agent-smith/skills/agent-smith/ into ~/.claude/skills/agent-smith/ (macOS/Linux) or %USERPROFILE%\.claude\skills\agent-smith\ (Windows).

Setup — pick ONE backend (or several)

Option A — Gemini cloud (default, fastest & strongest)

Get a free API key at https://aistudio.google.com/apikey, then:

export GEMINI_API_KEY=your_key_here     # macOS/Linux (add to your shell profile to persist)
# Windows:  setx GEMINI_API_KEY "your_key_here"   (open a new shell after)

Hitting free-tier rate limits (429s)? If you have the Gemini CLI installed and logged in (gemini → "Login with Google"), use --backend gemini-cli to run the same Gemini models on your subscription/account quota instead of the metered API key — no rate-limit ceiling. It auto-disables the CLI's tools (text-only, never edits files) and runs ~25% leaner.

Option B — No account, fully local (Ollama)

Don't have a Gemini account? You don't need one. Install Ollama, make sure ollama serve is running, then run the disk-aware installer — it sizes the model to your free space:

bash plugins/agent-smith/skills/agent-smith/scripts/setup_local_model.sh

It offers, by available disk:

Then use --backend ollama (defaults to qwen3-coder:30b, or set OLLAMA_MODEL).

Option C — Apple Foundation Models (advanced, opt-in)

The fm backend runs on-device on macOS 26+ with Apple Intelligence. No binary ships with this plugin (don't run opaque executables from strangers) — you supply your own fm_helper: a tiny Swift CLI wrapping Apple's FoundationModels framework that reads {"messages":[...],"system":...} JSON on stdin and prints {"answer": "..."}. Point the skill at it with export FM_HELPER=/path/to/fm_helper. If unset, the fm backend simply errors and you stay on gemini/ollama.

Usage

Once a backend is set up, just ask Claude to do offload-shaped work ("research X with sources", "summarize this 200-page PDF", "draft a wrangler.toml", "turn this announcement into posts"). The skill triggers automatically. You can also be explicit: "use Gemini for this."

Under the hood Claude calls the helper, e.g.:

SKILL="${CLAUDE_PLUGIN_ROOT:+$CLAUDE_PLUGIN_ROOT/skills/agent-smith}"
SKILL="${SKILL:-$HOME/.claude/skills/agent-smith}"

python3 "$SKILL/scripts/gemini.py" --search "What's new in the latest Python release?"
python3 "$SKILL/scripts/gemini.py" --file report.pdf "Summarize the findings as bullets"
python3 "$SKILL/scripts/gemini.py" --backend ollama --model qwen3-coder:30b "Draft a function that ..."

Which model for what (from the bundled coding bake-off)

• Best coder overall: Gemini --model pro (cloud) — swept correctness + design.
• Best local/offline coder: qwen3-coder:30b (30B MoE, 3B active — ties qwen2.5-coder:14b on correctness but ~2× faster + better design). qwen2.5-coder:14b is the lighter runner-up.
• General text, no account: Gemma (gemma3:12b/27b).
• Always: the model drafts, you verify. Every model tested shipped at least one bug a review caught.

Platform support

• Gemini / Ollama backends: macOS, Linux, Windows (pure-stdlib Python helper).
• Apple FM backend: macOS 26+ only, and only with your own fm_helper.

License

MIT — see LICENSE.