Ante

What it is in one line

A ~15MB single-binary coding agent written in Rust that crams an embedded llama.cpp inference engine, grep, and git into one file, so the whole read-edit-test loop can run offline against a local GGUF model — or against any of 12+ cloud providers.

Who built it

Antigma Labs (antigma.ai). The author says the focus is "the harness, not the model." Public repo is Apache-2.0, created 2025-12-23, 1,680 stars / 52 forks, latest release v0.preview.78. The core harness ships as a prebuilt binary from a private repo; the open-sourced parts are the docs, protocol (protocol-shape), SDK (agent-sdk), and eval pipeline (ante-harbor).

Read: making packaging the product and public benchmarking the trust signal are two bets on the same thing — in a commoditized agent-tool market, win developers on engineering and transparency instead of on model claims.

What it actually does

• One binary, four shapes → interactive TUI (ante), headless one-shot (ante -p, for CI), long-lived server (ante serve, JSONL protocol for editor plugins), and gateway (ante gateway, as a Slack/Discord bot)
• Native local inference → a pinned llama.cpp build; point it at any GGUF file and the loop runs with no API key, no account, no network. Also 12+ cloud providers (Anthropic, OpenAI, Gemini, DeepSeek, etc.) with easy switching
• Embedded tooling → grep and git compiled into the binary instead of shelling out
• Multi-agent and memory → sub-agent orchestration, MCP, persistent cross-session memory
• Named profiles → --profile swaps system prompt, toolset, skills, and memory; a minimal bare profile is built in, and the community can share new ones
• Verifiable evals → continuously benchmarked on Terminal-Bench 2.1 under official leaderboard constraints, every result pinned to a downloadable build

What it deliberately does not do: no forced account ("No account required, not even with us"), no model lock-in. Telemetry is on by default but disableable via ANTE_TELEMETRY=off and sends only an anonymous install label.

What old behavior it replaces

Running a coding agent used to mean installing a runtime stack: a Node or Python environment, editor plugins, a package tree, plus API keys or a local inference daemon like Ollama. Environment drift was the norm — last week's container silently fails this week.

Ante replaces that whole install-and-run assumption: one file, curl | bash, and if you want local inference you point it at a GGUF and there is no daemon to manage.

The offline half replaces another old behavior: air-gapped industries — defense, healthcare, finance — were effectively locked out of coding agents. "Your code never leaves your machine" is not a selling point to them, it is the entry ticket. They previously edited code by hand; now there is an option that runs an agent with the network down.

Business model

Free in alpha. The prebuilt binary is free during alpha preview, commercial use included, per BINARY-TERMS.md; the open-source parts are Apache-2.0. The README mentions an opt-in hosted service at antix.antigma.ai, hinting at future commercialization, but there is no pricing.

Read: the classic open-core path — free binary to build the base, future monetization on the hosted layer. But the company hasn't even committed to whether a cloud is needed, so the model itself isn't standing yet.

Hard numbers

• 1,680 stars / 52 forks. Repo created 2025-12-23, latest release v0.preview.78 (2026-08-14)
• HN: 159 points, 88 comments, front page on launch
• Terminal-Bench 2.1: 82.7% (DeepSeek V4 Flash 0731, 89 tasks × 5 trials, 368/445 passed, ~$68 of inference). Official leaderboard #1 is Claude Code + Fable 5 at 83.8%; Ante has not submitted, so 82.7% is a vendor-reported number
• Resource footprint: on the same 20 parallel Docker tasks, ~7× lower peak memory, ~9× lower average CPU, ~5× lower disk I/O vs Claude Code (self-reported)
• Offline config is self-reported at ~56.2% (Qwen-class 27B quantized, a figure from the HN thread) — buying "offline" costs roughly 25 points of task success
• Team size and paying users: not disclosed

Four-way read

The call

This is BusyBox/SQLite philosophy applied to agents: the packaging is the product claim. Everyone else competes on models and prompts; Ante competes on "no environment, no network, one file and it runs." For teams orchestrating dozens or hundreds of agents on one box, harness weight stops being a rounding error and starts being the bill. The direction holds.

The real disagreement with the Claude Codes is the trust model. The core harness is closed source, shipped as a prebuilt binary — while the entire pitch is "your code never leaves your machine." The audience that needs offline agents is exactly the regulated, security-sensitive audience that will not run an unauditable binary on faith. The SDK, protocol, and eval pipeline are open, but the thing you actually run is not fully inspectable. Until the core crates move publicly, the pitch and the license are fighting each other.

Discount the benchmark. The 82.7% ran on a cloud DeepSeek model — a different configuration from the headline offline story. The self-reported offline figure is ~56%, about 25 points below the front-end config. Anyone buying the offline story is paying a real price, and that ledger has to be on the table.

The transferable rule: single-file distribution is the most complete way to zero out install cost. No installer, no dependency tree, no daemon — one curl pipe. Install cost is systematically underrated as a conversion factor in developer tools.

What to watch next

① Whether the core harness source actually ships as promised ("progressively") — open-source progress is trust-repair speed ② Whether they submit to the official Terminal-Bench leaderboard — 82.7% only counts once it's on the board ③ The pricing and shape of the antix hosted service — the first step from free acquisition to a business model

What you can take from it

Product logic: if your tool targets "run many agents in parallel" or "air-gapped environments," single-binary plus zero runtime deps plus optional local inference is a copyable architecture decision. But decide the closed/open trust cost up front — a product that promises offline privacy cannot hand users an unauditable binary.

Pricing structure: free-in-alpha plus a future hosted layer is the current default for agent tools. Watch it, don't copy it; the real question is whether anyone pays for the hosted half.

Verdict

Worth watching. The packaging philosophy and eval transparency are rare in this category, and the HN heat was real (159 points, 88 comments). But the trust model (closed core) and the offline performance gap (-25 points) are two promises still to be honored. Revisit in three months against the three checks above — especially the first, which decides whether this is a true open-source project or an open shell.