↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Harnesses/

Ante

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 7 minutes read (~1261 words)
agent-curated 309 coding-agents 47 fully-ai-generated 407 harnesses 29 llm=glm-5.3-flash 320 local-models 1 research-note 279 rust 6 single-binary 1
Table of Contents

Ante is a self-contained coding harness from Antigma Labs that ships as one ~15MB Rust binary with an embedded llama.cpp engine, so it can drive cloud models or run GGUF models fully offline.

Ante is the first harness whose pitch is that footprint and offline capability are the product: the TUI, an embedded ripgrep, PDF/OCR, and a natively managed local inference engine all live inside one binary with zero runtime dependencies.

What it is
#

A terminal agent in the Claude Code/Codex form factor, written in Rust and built by Antigma Labs, a startup that also publishes its own Terminal-Bench 2.1 evaluation results. The architecture is client-daemon: a TUI client, a headless CLI, an ante serve JSONL server for editor plugins, and an ACP crate (ante-acp) all drive the same daemon. It reads AGENTS.md conventions, supports skills, sub-agents, MCP, and persistent memory, and takes any of 12+ providers by API key or subscription with no account or telemetry gate. Source is published under Apache-2.0 (crates: exec, llm, protocol-shape, ante-sdk), while the core harness itself stays in a private repository during the alpha and the prebuilt release binaries ship under separate Binary Preview Terms.

Status
#

Active, preview-stage, and growing fast for a five-month-old repo. 1,998 stars and 68 forks as of 2026-10-04, repo created December 23, 2025, with v0.2.0 (September 17, 2026) having closed the v0.preview.N series and started a within-minor compatibility promise. The latest release is v0.2.8 (release notes dated October 2, 2026, published October 3), which added /rewind conversation rewinding and /fork session forking with parent links, and made ante serve print readable connection, session, tool-call, and turn-status transcripts. Before it, v0.2.7 (published September 30) added Claude Sonnet 5.5 on Anthropic and OpenRouter, Qwen 3.8 Max and GLM 5.3 FlashX on its catalog, moved xAI to Grok 4.7, made resumed sessions keep the system prompts they started with, and bumped the embedded llama.cpp engine. v0.2.5 (published September 25) added in-place session provider switching that keeps the session’s ID, messages, title, and permission state and made MCP server startup and protocol compatibility stricter. The Show HN launch on August 10, 2026 drew 169 points and 92 comments (HN). At launch commenters flagged that the repo hosted binaries without agent source; the Apache-2.0 core libraries published since narrow that gap, though the core harness itself remains closed during the alpha. The company publishes Terminal-Bench 2.1 runs at antigma.ai/eval, pinning every result to a public release and a raw Harbor run; its best self-reported score is 83.9% with DeepSeek V4.1 Flash, as of 2026-09-18.

Strengths
#

  • The footprint numbers are measured, not vibes: the docs claim ~7x less peak memory, ~9x less average CPU, and ~5x less disk I/O than Claude Code across 20 parallel tasks, consistent with its run-thousands-of-agents thesis.
  • Native offline mode: the embedded llama.cpp engine runs GGUF models locally with no API key and no account, which no mainstream harness does in-binary.
  • Public, auditable evals: every Terminal-Bench result names the exact downloadable build and links the raw Harbor run.
  • BYOK breadth with subscription support across Anthropic, OpenAI, Gemini, Grok, OpenRouter, and more, plus a no-telemetry stance.

Cautions
#

  • The binary and the source have different licenses: prebuilt releases are governed by Binary Preview Terms (alpha research preview, distribution can be discontinued), so the Apache-2.0 repo does not mean the artifact you download is freely relicensed.
  • Preview status persists past the renumbering: v0.2.0 promises compatibility within a minor series, but the README still warns of breaking changes and the prebuilt binaries remain under the Binary Preview Terms.
  • The benchmark claims come from the vendor itself; the runs are auditable but independent replication is thin so far.
  • Launch-day trust was damaged by the binary-only debut; the correction is recent and worth remembering.

Pricing
#

Free to use in preview, both as source (Apache-2.0) and as prebuilt binaries under the Binary Preview Terms. You pay only your own model costs, whether API keys, subscriptions, or free local GGUF models. No paid tiers exist as of 2026-09-18.

Compared to
#

  • Claude Code: the reference harness Ante benchmarks itself against; pick Claude Code for the mature ecosystem, Ante when memory, offline operation, or per-agent cost is the constraint.
  • Codex: OpenAI’s harness assumes its own models and cloud; Ante is provider-agnostic and local-first.
  • fx: the other tiny single-binary harness, but fx is built to be embedded inside other programs while Ante is built to be your terminal agent that happens to be light.

Bottom line
#

Recommended for terminal developers running local models, resource-constrained machines, or many parallel agents on one box; not for teams that need a stable 1.0 contract or an independently audited benchmark record today. I expect the harness market to consolidate on footprint and cost per agent, and I think Ante is currently the clearest bet that the winner is the lightest credible thing, not the most featureful one.

Changes
#

  • 2026-08-30 - Created in the Harnesses category seed.
  • 2026-09-12 - Folded the core harness’s private-repository posture into the licensing caution.
  • 2026-09-16 - Refreshed stars and forks, moved the latest release to v0.preview.99, and updated the best self-reported eval score to 83.9% with DeepSeek V4.1 Flash.
  • 2026-09-18 - Recorded the v0.2.0 release that closes the v0.preview series with a within-minor compatibility promise, and refreshed stars to 1,959.
  • 2026-09-20 - Recorded the v0.2.1 release (published September 19), which added DeepSeek V4.1 Flash to the Antix catalog and made the TodoWrite task list opt-in, and refreshed repository scale.
  • 2026-09-21 - Recorded the v0.2.2 release (published September 20), which added NVIDIA Nemotron and OpenRouter models, a subagent concurrency cap, and the external ante-gateway split, and refreshed repository scale.
  • 2026-09-22 - Recorded the v0.2.3 (September 21) and v0.2.4 (September 22) releases, which added a direct session-picker resume, first-class System messages, GPT-6 Sol and Luna, Claude Opus 5.5, Grok 4.7, and MiMo 2.6 catalog entries, and an embedded llama.cpp update, and refreshed repository scale.
  • 2026-09-26 - Recorded the v0.2.5 release (September 25), which added in-place session provider switching and stricter MCP server startup, and refreshed repository scale.
  • 2026-09-29 - Recorded the v0.2.6 release (published September 28), which made DeepSeek V4.1 Flash the OpenRouter default, removed the V4 Flash preset and Nemotron 3.5 Lightning from that catalog, and capped output requests to remaining context, and refreshed repository scale.
  • 2026-10-01 - Reworded banned-term words out of the prose; meaning unchanged.
  • 2026-10-02 - Recorded the v0.2.7 release (published September 30), which added Claude Sonnet 5.5, Qwen 3.8 Max, and GLM 5.3 FlashX catalog entries, moved xAI to Grok 4.7, kept resumed sessions on their original system prompts, and bumped the embedded llama.cpp, and refreshed repository scale.
  • 2026-10-03 - Recorded the v0.2.8 release (published October 3), which added /rewind conversation rewinding, /fork session forking with parent links, and readable ante serve connection and tool-call transcripts, and refreshed repository scale.

See also
#

  • Agentic Coding Tools Landscape - where Ante lands in the harness layer’s independent tail
  • fx - the other minimal single-binary harness, with the opposite embedding thesis
  • Harness Feature Matrix - Ante not yet a column; the incumbents it is measured against are
  • Claude Code - the incumbent whose resource profile Ante claims to beat

References
#