A minimalist coding agent with hashline-based editing. Supports multiple providers, including local LLMs.
Lines are tracked by content hashes. Edits reference LINE:HASH anchors — if the file changed, mismatches are caught and retried automatically.
The default system prompt and tool definitions fit in less than 3k tokens. Extremely fast time-to-first-token on cloud APIs, and small enough to run well on local models.
Works with OpenAI, Anthropic, Google Gemini, Azure OpenAI, Ollama, LM Studio, and OpenRouter. Cloud or fully local.
Discuss an idea, save a plan, split it into ordered tasks, then work through them one by one. Persistent across sessions.
macOS Seatbelt integration restricts file writes to the project directory. Reads and network stay unrestricted.
Seven focused tools — read, edit, write, skim, filegrep, shell, shell_status. Background process management built in.
Hashline editing prevents conflicts by tracking content, not line numbers.
nav reads files with hashline-prefixed output. Each line gets a short content hash.
42:a3|const foo = "bar";
43:f1|const baz = 42;Edits reference LINE:HASH anchors instead of reproducing old content. The model says what to change, not what the file looks like.
edit lines 42:a3-43:f1 with:
const foo = "updated";
const baz = 99;If the file changed between read and edit, hashes won't match. The edit is rejected with corrected anchors — the model retries without re-reading the entire file.
Cloud APIs, local Ollama, Azure deployments, or OpenRouter. Auto-detected from model name.
Discuss ideas interactively, save plans, split into tasks, then run them sequentially or pick your own order.
When context gets long, nav summarizes progress and continues in a fresh window. Automatic or manual.