Show HN: Ante, a coding agent in a single binary that runs offline

Alpha preview: expect breaking changes and incomplete functionality. macOS and Linux only; connected Windows we propose WSL.

A shade successful your shell. Ante is simply a self-contained coding supplier that lives successful your terminal and self-organizes. One ~15MB Rust binary from Antigma Labs, zero runtime dependencies, built to get the astir retired of immoderate model.

It useful for illustration Claude Code aliases Codex, pinch nary of their limitations aliases exemplary constraints. It tin besides beryllium the optimized halfway for building your ain harness and high-performing assistants.

Every supplier claims to beryllium good. Here are numbers you tin check:

Ante runs Terminal-Bench 2.1 continuously nether charismatic leaderboard constraints: 89 tasks, 5 tests each. Each consequence pins the nonstop build you tin download and links the earthy Harbor tally for independent audit. Latest afloat run: 82.7% pinch open-weight DeepSeek V4 Flash 0731 (368/445 trials, Ante 0.preview.71, astir $68 of inference). DeepSeek reports the aforesaid 82.7 for this model, measured pinch its unreleased DeepSeek Harness successful minimal mode.

Live results → · Methodology →

Ante is hand-written Rust pinch the dense parts (Grep, git, section inference) embedded successful 1 binary, 1 process. Across the aforesaid 20 parallel tasks successful Docker, Ante uses ~7× little highest memory, ~9× little mean CPU, and ~5× little disk I/O than Claude Code.

Raw numbers → · Benchmark specifications →

Ante ships its ain conclusion engine. Point it astatine a GGUF record and the full loop runs connected your machine: nary API key, nary account, nary internet.

Offline mode →

The 3 are 1 creation decision. An supplier you tin verify, afford, and run anywhere is ray capable to tally by the thousands: the substrate for self-organizing intelligence.

See each cookbook guides

Ante is simply a single, self-contained binary pinch nary outer dependencies: download and run.

• Zero vendor lock-in: bring your ain API key, subscription, aliases section model. Switch betwixt 12+ providers freely. No relationship required, not moreover pinch us.
• Multi-agent orchestration: spawn sub-agents and coordinate analyzable tasks crossed independent, decentralized, and centralized architectures. See the patterns →
• Channel integrations: tally Ante arsenic a Slack aliases Discord bot pinch ante gateway.
• Extensible: civilization skills, sub-agents, MCP, and persistent representation crossed sessions.

Ante useful pinch 12+ providers retired of the box:

Configure providers via situation variables (ANTHROPIC_API_KEY, OPENAI_API_KEY, etc.) aliases OAuth. Add civilization providers successful ~/.ante/catalog.json.

We unfastened originated what really matters successful the property of agentic coding, each nether Apache 2.0:

1. Detailed documentation, the descriptive truth. docs-site/ is the root for docs.antigma.ai: a precise explanation of what the harness does and really to thrust it.
2. The protocol, the algorithm of the core. crates/protocol-shape defines the schema and ligament messages spoken by ante serve; crates/agent-sdk is the Rust SDK and customer for building against supplier runtimes.
3. The eval pipeline, constraint and continuous improvement. ante-harbor/ is the Harbor supplier adapter down our Terminal-Bench results: usage it to reproduce immoderate tally astatine antigma.ai/eval. CHANGELOG.md records the improvement, merchandise by release.

The halfway harness itself is developed successful a backstage repository during the alpha and ships arsenic a prebuilt binary via releases. Core libraries from it are included present progressively arsenic they stabilize; crates/exec, standalone process execution, is the first.

The protocol aboveground maps to Ante's client-daemon architecture:

We attraction astir the harness, not the exemplary aliases the prompts.

Documentation is the caller root code.

Ante is designed for cellular-native agents: for illustration cells successful an organism, tiny, expendable, massively replicated. That thesis is why the 3 header claims exist. A cell-scale supplier must beryllium verified (reliability compounds astatine scale), tiny (every byte is multiplied by thousands), and self-contained (no runtime to install, nary work to telephone location to). Read much successful our philosophy and agent statement patterns.

The sanction is the answer: Another Terminal agent, and ante, the liking you put connected the array to play. Ante is fast, lightweight, and the only terminal supplier pinch autochthonal section conclusion built in. We judge a self-contained supplier halfway that self-organizes is the instauration of the coming supplier economy.

Ante has astir of the features you expect from agents for illustration Claude Code aliases Codex: multi-agents, skills, MCP, persistent memory. The quality is the build philosophy.

• Built from scratch successful Rust. Core components for illustration Grep (fully rebuilt and customized) and git are embedded successful the aforesaid ~15MB binary and tally successful the aforesaid process astatine runtime, truthful thing is shelled retired and nary resources leak. Most akin projects vessel connected Node.js aliases CPython and transportation an order-of-magnitude larger footprint.
• We built our ain conclusion motor from the crushed up (see nanochat-rs for a artifact type of that work), truthful a section GGUF exemplary is each Ante needs to tally without immoderate provider.
• No vendor lock-in, not moreover to ourselves: nary relationship needed, reuse your existing API credentials. An opt-in, afloat integrated server-side acquisition lives astatine antix.antigma.ai.
• Every declare is backed by public, reproducible benchmarks of the nonstop builds we ship: antigma.ai/eval.

Beyond the footprint it comes down to supplier architecture, and yet to who is building it and pinch what philosophy. Anyone tin fork a binary; sensation and engineering rigor don't copy. Those differences leak into each item of the product.

For one-on-one supplier interactions, runtime overhead for illustration representation usage and I/O is often little important than exemplary inference.

But our imagination is overmuch bigger: millions of agents self-organizing and communicating astatine monolithic scale. At that point, moreover mini inefficiencies get multiplied millions aliases billions of times, truthful runtime optimization becomes economically significant.

Yes. Ante has a built-in llama.cpp motor that runs GGUF models locally. It handles motor installation, exemplary discovery, and representation guidance automatically. No API keys aliases net relationship required.

Yes. Create a ~/.ante/catalog.json record to adhd aliases override providers and models pinch civilization endpoints, API keys, and configurations. Any OpenAI-compatible API works.

Server mode runs Ante arsenic a long-lived daemon that communicates complete a system JSONL protocol. It's perfect for building editor plugins, web UIs, and civilization integrations connected apical of Ante.

Full archiving is disposable astatine docs.antigma.ai.

Source codification successful this repository (including the SDK and protocol crates) is licensed nether the Apache License 2.0.

The prebuilt ante binary is free to usage — including commercially — during the alpha preview nether the Binary Preview Terms. The core harness is presently developed successful a backstage repository and shipped arsenic a binary; the SDK and protocol aboveground you build against present will remain permissively licensed.