<2-3 sentences summarising what the project addresses and solves, the target audience, and the benefit.> Late (Lightweight AI Terminal Environment) addresses the inefficiencies of standard AI coding assistants, which suffer from token bloat, amnesia, and degraded reasoning due to large context windows. It provides a deterministic orchestrator for solo developers, enabling them to achieve the throughput of an entire engineering team with minimal VRAM. The core benefit is efficient, cost-effective AI-assisted development through intelligent delegation and precise code manipulation.

How It Works

Late employs a "Lead Architect" orchestrator that maps the codebase and spawns ephemeral, isolated "Atomic Subagents." Each subagent receives a strictly scoped, fresh context window containing only the exact task instructions, preventing context pollution and KV cache bloat. This approach mirrors real engineering teams, ensuring fast processing, zero context degradation, and significantly reduced VRAM and API costs. It guarantees deterministic, exact-match code edits, avoiding silent failures.

Quick Start & Requirements

• Install: Download the pre-compiled single binary for Linux, macOS, or Windows from the Releases page and ensure it's in your system's PATH.
• Prerequisites: An OpenAI-compatible API endpoint (local or cloud).
• Hardware: Designed for consumer hardware, capable of running on 5GB VRAM with local models.
• Build from Source: Requires Go. Clone the repository (https://github.com/mlhher/late.git), navigate to the directory, and run make build followed by make install.

Highlighted Details

• Zero Token Bloat: Core system prompt is optimized to ~1,000 tokens, maximizing context for code.
• Exact-Match Diffs: Subagents use strict search/replace, failing loudly on mismatches to trigger self-healing loops.
• Human-in-the-Loop: Auto-approves read-only commands; requires explicit confirmation for state-changing operations.
• Pure Go & No Dependencies: Statically compiled binary eliminates external runtime requirements.
• Local-First & Model Agnostic: Supports any OpenAI-compatible endpoint, including local models via llama.cpp.
• Stateful Resilience: Maintains session history on disk for context retention across runs.
• Git Worktree Support: Enables parallel, isolated instances across multiple Git worktrees.
• Agent Skills: Supports reusable instruction sets and scripts.

Maintenance & Community

No specific details regarding maintainers, community channels (e.g., Discord, Slack), or roadmap were found in the provided text.

Licensing & Compatibility

• License: Business Source License 1.1 (BSL 1.1).
• Restrictions: Prohibits monetizing Late itself (e.g., as a paid AI service) or deploying it as enterprise internal infrastructure without a commercial agreement.
• Compatibility: Free for individual use, including commercial startups writing code. Converts to GPLv2 on February 21, 2030.

Limitations & Caveats

The BSL 1.1 license imposes commercial restrictions for specific enterprise and service-based deployments. While designed for speed, the human-in-the-loop confirmation for state-changing commands may require user interaction. The agent operates strictly within the project directory, blocking cd commands.