late-cli
AI coding agent that always stays sharp. 200k+ tokens of work in a 64k context via parallel isolated subagents. Sandboxed overnight runs. Sub-10ms Go CLI.
What is late-cli?
AI coding agent that always stays sharp. 200k+ tokens of work in a 64k context via parallel isolated subagents. Sandboxed overnight runs. Sub-10ms Go CLI.
Overview of late-cli
What it is
Late is an AI coding agent implemented as a Go CLI. It is designed to keep the model's context clean during long workflows by isolating execution steps. The tool works with any cloud provider or local model and can be dropped into any project. It is distributed as a single, statically compiled binary with zero dependencies.
Key features
- Architecturally enforced isolation: a Lead Orchestrator plans and verifies, while ephemeral Coder and Researcher subagents are spawned into isolated contexts whose scratchpads are wiped after each atomic task.
- Disposable test-time compute: subagents can spend more aggregate inference than would fit cleanly into one reasoning trajectory, trading cheap compute for a bounded, high-signal orchestrator context.
- Empirical logit biasing: real-time logit biasing for llama.cpp/llama-server suppresses redundant thinking tokens on the fly, reclaiming wasted chain-of-thought compute.
- Physical tool registry pruning: the Orchestrator has no file-writing tools, has repo-mutating bash commands blocked, and cannot bypass delegation; subagents have no orchestration tools and cannot recursively spawn agents.
- Parallel subagents: independent plan steps can run as concurrent, fully isolated subagents via spawn_subagents, with all results returning to the orchestrator in a single turn (off by default, enabled with --max-async-subagents).
Use cases
Late is suited for complex, multi-step coding tasks that would otherwise cause context degradation in standard agents. It can handle codebase scans, compiler errors, file reads, failed diffs, and retries without accumulating execution noise in the primary context. The tool is also applicable for overnight sandboxed runs and workflows where a model with a 64k context window needs to complete jobs of 200k+ tokens.
Getting started
bash# Linux / macOS (Homebrew)brew tap mlhher/late && brew install late``````bash# Universal Fallback (Linux / macOS / Windows WSL)curl -sfL https://raw.githubusercontent.com/mlhher/late-cli/main/install.sh | bash``````bash# Launch interactively in any projectcd your-projectlate
License and pricing
The license is not stated in the provided metadata or README. The pricing type is listed as open_source.
Based on the project's public README and repository metadata.