⚓

Drydock

A local-first, provider-agnostic terminal coding agent for your own LLM. No accounts, no telemetry, no cloud — your code and credentials never leave your machine.

$ pip install drydock-cli
$ drydock
PyPI → GitHub →

What it does

A full agentic CLI harness in a Textual TUI — scrolling transcript, collapsible tool & reasoning cards, a live activity line. The model calls every tool below autonomously. Everything is clean-room and dependency-free.

Files & shell

Read (with a structure index for huge files), Write, Edit, Bash, Glob, Grep — with advisory write guards.

Vision (multimodal)

Reference an image path in your message — a screenshot, mockup, or diagram — and it's attached for a vision-capable model to see: describe a UI, read text off a screenshot, debug a diagram. Any --mmproj server.

Version control

First-class GitStatus / GitDiff / GitLog / GitCommit — structured, truncated; commits are local and reversible.

Internet search

WebSearch + WebFetch bring live web results and pages into context. Offline-safe — degrades cleanly with no network.

GraphRAG knowledge base

Ingest your docs — text, PDF, Word, and STIG checklists — with /graphrag build <path> (and /graphrag add for more); the agent then retrieves project facts via the Knowledge tool.

Multi-agent

Dispatch runs several read-only sub-agents in parallel, each in a fresh context, for focused investigation.

Second-model advisor

Wire in a stronger model (e.g. Gemini, any OpenAI-compatible endpoint) with /advisor. The agent calls the Consult tool for a second opinion when stuck; you can /ask directly. Opt-in, no extra dependency.

MCP support

Connect to Model Context Protocol servers; their tools appear to the agent as mcp__server__tool. Works with third-party servers like Graphify for a queryable code knowledge graph.

Skills

Create reusable /<name> commands with /skills new <name> <prompt> ($ARGS substitution), or author them as markdown.

Loops

/loop <count> <prompt> runs a prompt iteratively — "fix the next failing test and run pytest". Esc stops.

RMF automation

Ingest the NIST SP 800-53 catalog (/rmf bootstrap) + bundled RMF skills (/rmf-control, -categorize, -review, -poam). 100% local for CUI.

DISA STIG automation

A raw STIG benchmark → completed checklist, all local: /stig new <xccdf> builds a blank .ckl, /stig-assess evaluates each rule vs your evidence, /stig-remediate writes a fix script, /stig graph auto-links rules to NIST controls via DISA's CCI map, and /stig poam exports the open findings to a deterministic eMASS POA&M CSV.

NIST governance skills

Defensive, clean-room skills for AI & cyber risk management: /nist-ai-rmf (AI RMF 1.0 — Govern / Map / Measure / Manage + the seven trustworthiness characteristics) and /nist-csf (Cybersecurity Framework 2.0 — the six Functions, Implementation Tiers, current→target roadmap).

ML engineering skills

Built-in workflows for PyTorch work: /ml-train, /ml-metrics (accuracy · precision · recall · F1 · MCC · ROC-AUC · confusion), /ml-finetune (full & LoRA), /ml-debug, /ml-rl, /ml-data — plus technique recipes retrieved into context per task.

Governed reliability

A deterministic controller governs the loop: the objective & acceptance criteria live in structured state that survives compaction, a verification gate blocks a self-declared "done" on an unrun or failing check, and every run writes a durable event trace you can inspect with /events.

Runs on Windows too

Native on Windows via PowerShell or cmd — no WSL or Git-Bash required — as well as Linux and macOS. /shell shows exactly which shell your commands run in.

Screen capture

The agent can take a Screenshot of your screen and see it with vision — review a GUI, read what's displayed, or debug a rendered result. Works on Windows, macOS, and Linux.

Context control

/context shows and sets the context-window budget (persisted), so a stale config or smaller model server never silently caps your window.

Local-first, by design

Drydock is an original, Apache-2.0, clean-room codebase owned end to end.

  • The only outbound calls are to the model endpoint you configure and (optionally) the web tools you invoke.
  • No telemetry, no phone-home, no hardcoded third-party hosts, no credential transmission — ever.
  • A credential-exfiltration scanner gates every release and blocks anything reaching off-box.
  • Safety is advisory, never blocking: catastrophic commands are refused; sensitive ones prompt Allow / Always / Deny.

Runs on your hardware

Provider-agnostic — any OpenAI-compatible endpoint (llama.cpp, vLLM, Ollama, LM Studio). Tuned against dense Gemma-4-31B (QAT, 64K) on a single workstation; first launch autodetects a local model server and wires it up with no account or API-key prompt.

Tools: Read · Write · Edit · Bash · Glob · Grep · ViewImage · Screenshot · todo · task · Dispatch · GitStatus · GitDiff · GitLog · GitCommit · WebSearch · WebFetch · Knowledge · BuildKnowledge · mcp__<server>__<tool>
Slash commands: /help · /model · /cwd · /undo · /back · /stop · /status · /compact · /context · /graphrag build|add|query|status|clear · /events · /shell · /skills · /skills new · /loop · /mcp · /clear · /quit
Bundled skills: /rmf-* · /stig-* · /nist-ai-rmf · /nist-csf · /ml-train · /ml-metrics · /ml-finetune · /ml-debug · /ml-rl · /ml-data
Or just ask the agent — "how do I add my own docs?" — it knows its own commands.