Release notes · Kaptain 0.0.8 · September 2026

Kaptain 0.0.8

CLI runtimes and handoffs handle long-running agents better, local models can take on your image processing, and chats load faster. Available now through Settings › General › Update Kaptain.

What is new in 0.0.8

CLI runtimes and handoffs. Claude and Codex conversations continue across tool toggles, compaction and settings changes instead of restarting. Long tool calls are no longer abandoned by the CLI. BlackBox shows the path each turn took.
Image processing with local models. Local tools for image files are enabled. Pick the model that processes images in Settings › Model Handling › Vision tool: a local model, a cloud model, a Claude or Codex CLI, or a model on one of your other machines.
Speed. Chat loading is optimised for local inference: a long chat shows its first message in about a second, and older tool-heavy turns fold into one line you can open.
Your machines together. A paired machine's models appear in your pickers under its name, can process your images, and its model settings (context window, output cap, thinking) are followed here.
Deview. Markdown files render the way they were written: quotes, wrapped lines, headings, task lists, tables, code.

CLI runtimes and handoffs

Turning Tools or Delegate on or off continues the same Claude or Codex thread.
When the thread's bookmark goes stale after compaction, Kaptain hands over a summary plus the recent messages into a fresh thread and continues from there.
Editing a persona or switching the project root hands the conversation over the same way, without re-sending the history.
Kapsule chats keep their thread when their tool grants change; only editing the instructions starts a new one.
The CLIs wait up to 15 minutes for a Kaptain tool (configurable), so a long image or delegation step completes.
BlackBox names the reason for each transport decision.

Image processing

Settings › Model Handling › Vision tool lists the models that can process images: local models that declare it, cloud models, your Claude and Codex CLIs, and models on paired machines.
Thinking for the vision model: Auto, Off or On, with two answer budgets. Applies live.
A CLI chosen as the vision model uses its own tools to process the image.
A GGUF model processes images when its projector file (the publisher's mmproj-*.gguf) sits in the same folder; Kaptain loads it with the model. A GGUF without one is not offered as a vision model, and the tool says which file to add.
The time limit for image processing is 10 minutes by default and configurable.

Your machines together

Pair Kaptains with a connection code under Settings › Network Access. Chats, files and settings stay private to each machine.
A paired machine's models show under its name in every picker. Chat headers show the machine name, not an address.
A paired machine can process your images with its own model, key or CLI.
Model settings made on the paired machine are followed here.
A machine on 0.0.7 or older lists its models but cannot process images or share model settings until updated.

The workbench

Deview renders Markdown with the same CommonMark and GFM engine editor previews use. Links open in a new tab; images load only from trusted sources.
Type "/" in any chat for commands and installed skills, including a CLI's own skills when that CLI is the model.
Krew › Skills imports a skill from a folder, a zip, a git URL or owner/repo.
Kapsule chat uses the same renderer as the main chat. Each turn's trace sits under its answer. Deleting a chat or a kapsule asks first.
Long tool-heavy turns can continue past the context window (Settings › Model Handling › Context & Compaction, off by default).

Everything Kaptain includes

The full surface as of 0.0.8. The User Guide covers each part.

Surface	What you get
Chat	One conversation with the model you choose: local Ollama models, GGUF files through llama.cpp, cloud APIs with your keys, or the Claude, Codex and Gemini CLIs. Tools, delegation, attachments, steering mid-turn, saved sessions.
Kapsules	Standing agents with their own model, memory, tools and chat thread. Schedules, run now, a reviewer for correctness, per-kapsule tool grants.
Krew	Built-in tools with a per-tool switch, your MCP servers, imported skills. Approvals before anything changes.
Deview	Project tree, editor with Read, Edit and Preview, Markdown and HTML preview, images, project-wide search.
God's Eye	Your project as a graph of files, functions, routes and symbols, indexed by KodeGraph.
BlackBox	Per-turn traces: context sent, tool calls, reply, tokens, transport decisions. Model usage totals.
Channels	Outside chat surfaces, each one your own bot run from this device.
Settings	General, Model Handling, Runtimes, Network Access.
Network Access	Localhost or Tailscale, device allowlist, per-device access code, access log, connection codes between your Kaptains.
Updates	Signed releases, checked and applied from Settings; one binary per platform.

How to update

Settings › General › Update Kaptain › Check for updates, then Update now. Kaptain swaps the binary in place, keeps a backup and asks you to restart. New installs follow the install guide.

Earlier releases

0.0.7: traces that read like the chat, reviewer improvements, Kapsules approvals, faster long chats, the connection-code mesh, Windows fixes. 0.0.5 and 0.0.6: the Go binary, the setup wizard, licensing, the first Kapsules.