Agent tools

NoTrack Coder turns the model's answer into actions by parsing tool tags it emits and running them on your machine. You don't call tools yourself — the model does, and you see each one as a card. This is the full toolset.

Files

• read_file — read a file's contents.
• write_file — create or overwrite a file.
• append_file — add to the end of a file.
• edit_file — replace a specific piece of text in a file.
• delete_file, make_dir, move_file, copy_file — the usual file operations.
• list_dir, grep, glob — list a directory, search file contents, match by pattern.

All paths resolve inside the chat's working folder unless absolute.

Shell

• shell — run a command. On Windows the shell is PowerShell. You can still write common Unix idioms — NoTrack Coder translates them on the fly: top‑level && → ;, -o /dev/null → NUL, 2>/dev/null → $null (only outside quotes, so remote ssh "… && …" is left intact). Success is judged by the real process exit code, so a curl that got a 200/403 while writing to stderr is not flagged as an error (no more pointless retries). Dangerous commands (recursive force‑delete, disk wipes, piping remote scripts into a shell) are blocked, and risky ones prompt in Ask mode.

Secrets in tool output (private keys, .env passwords, bearer tokens) are automatically redacted before they reach the model's context.

Web

• web_search — searches two sources in parallel and returns both, labeled: DuckDuckGo (general web) and freeserp.ai (a keyless index of fresh / AI‑built sites, each with a Domain Rating). The model picks the relevant result(s) itself.
• http_get — fetch a URL's contents.

Media & docs

• gen_image — generate a real image into the chat folder — see Image generation.
• ocr — read text from an image using the built‑in Windows OCR.

Orchestration

• task — spawn a parallel sub‑agent for a self‑contained subtask.
• run_on_model — hand a task to another connected model/endpoint (e.g. route a hard step to a bigger model).
• skill — load a SKILL.md playbook on demand.
• mcp — call a tool exposed by a connected MCP connector (see Connectors).
• list_capabilities — list the connected models and what they can do.

Knowledge & memory

• search_knowledge — search the local knowledge base (past sessions + distilled notes).
• remember — save a fact to memory. This only runs when you explicitly ask the agent to remember something.

Why you sometimes see retries

Weaker (uncensored / low‑quant) models occasionally misuse a tool — wrong tag syntax, Unix commands on Windows, or an over‑long reply that gets cut off. NoTrack Coder's guards catch these, nudge the model, and retry, so the task still completes — just with a few visible correction cards. A stronger model produces cleaner runs.

Next: Endpoints & models.