0.9.0 Latest release · 9 November 2026
Your agents can get tools built for them: describe one, and Crew's coding agents write it, test it and publish it to Flight Deck, which runs it for your agents in a sandbox on Linux.
Agents fit in one .captain file now, to back up, move or hand to a client. They hand each other work, answer email meant for them, read any IMAP inbox, learn from your corrections once you approve, share one external memory, and know the goals their work is for.
52 pull requests, #436 to #487.
Python tools of your own, built by Crew, tested before they go live, and run by Flight Deck as the person.
Export, import, packages, scheduled backups and signed files, in one .captain file.
Hand-overs, their own address, any IMAP inbox, lessons you approve and a Course of goals.
0.9.0 Also in this release
Seven new tools make 79 built-in tools per agent (78 in the desktop app): custom_tools, ask_user, colleagues, imap_mail, external_memory, course and image_manipulation.
ask_userAn agent asks you one question, or a short form of up to 10, in a Flight Deck dialog in every tab you have open. Send or Don't answer, and either way it says what it assumed.
Questions in Flight DeckCorrect an agent or state a rule, and it drafts one lesson: a Rule, Ask first or Fact, with your words. Approve, Edit or Dismiss; nothing is taught without your yes.
LessonsGoals for the deck, for you and for your agents, in every agent's prompt. Their suggestions wait on your Line, and a goal is never permission to do anything.
CourseMembase or Hindsight linked to all of a person's agents, or one Hindsight server for the whole deck, with a bank per person.
External memoryReplies in your Gmail per agent: Drafts (the default), Send to you or Send to anyone, still through your email-sending switch, allowlist and daily limit.
RepliesWho is using the deck, who has gone quiet, who never started, and whether you've written to them, with Write from your own Gmail or Outlook and Mark as written.
image_manipulationResize, crop, rotate, combine, watermark, annotate, animate and convert pictures, about 45 operations, with no model, no network and no scripts. Inputs are never changed.
A task starts an agent its owner switched off about five minutes before the run, and switches it off again when nothing else needs it. It used to fail with "agent not running".
tmp….docx one, and auto-save never overwrites an existing file.Upgrade To 0.9.0
Console → Updates does it on a deck on a git checkout. jsonschema is now a declared dependency, so install the package again by hand.
Agents need the restart for the seven new tools, the Course prompt and the lesson proposals. Then hard-refresh every browser.
cd ~/captain-claw && git pull
pip install -e . # or: uv sync
# restart Flight Deck, then every agent, then hard-refresh# lesson proposals, per agent, in config.yaml
lesson_proposals:
enabled: false
# agents picking up email that names them, deck-wide
FD_AGENT_MAIL=offgit on the host (uv is optional). On Linux, python3 tests/test_flight_deck/test_xtool_sandbox.py says whether the kernel has Landlock.external_tools/, agent-bundles/ and course.db in the deck's data folder, and bundle_signing_key in its home.Upgrading to 0.9.0 · New settings · Release notes · Installers on GitHub
0.8.11 5 November 2026
A message sent while the agent is busy waits under the transcript, Sends when ‹agent› is free, with Don't send, or Send at next step hands it to the running turn. It used to be refused and lost.
Back after a locked screen, every chat checks its connection and replaces a dead one, and the reply lands once.
Labelled New agent and New group buttons in the kiosk, initials with a state dot, and one-word states: Asleep, Off, Offline. Release notes
0.8.10 5 November 2026
Console → Models → New tier set opens on Quick start: pick Anthropic, OpenAI, Gemini, Grok or DeepSeek, paste the key, and every tier gets that provider's current models.
ocr_pdf and ocr_image read scanned PDFs and photos with a local OCR engine, and the same file gives the same text every time. That made 72 built-in tools.
A turn still making progress no longer stops at 40 steps, and on Windows files keep their own line endings. Release notes
0.8.9 4 November 2026
The Windows desktop app creates and runs agents, gives each its own settings, reads and writes UTF-8, and opens an agent's chat. Its first installer was on this release.
The app is signed ad hoc as a whole, so macOS only asks the first time: System Settings → Privacy & Security → Open Anyway. Release notes
0.8.8 4 November 2026
Captain for Chrome sends the page you're on to an agent, and Captain for Mac anything on your screen. On 4 November we also read what test users on the cohort deck ran into; these are the fixes, ten pull requests in all.
The page you're on, read in your own browser with its files, to the agent you pick.
Hold fn in any app: your selection, the window and your spoken task, and the answer pasted back.
Everything in 0.8.8.
The "you've written this file already" stop counts only whole-file writes, a re-run after a fix goes through, and the next section of a long report isn't refused as a repeat. The same write or command over and over is still stopped.
A card with Open and Download for every file a turn made, file names instead of server paths, runtime messages that no longer look like yours, and progress lines in your language that fold away after the answer.
A reply reporting other agents' work when no hand-off answered is turned down, and so is one reporting a schedule change no cron call made. Cron acts by job ID, never by list position.
Mail and calendar say "Google isn't connected" and where to connect it, instead of a policy error, and the new-agent picker says up front when an archetype works in mail.
Every agent shares one browser install instead of downloading its own ~700 MB copy. A server missing its system libraries gets one plain answer, and web reading falls back to plain HTTP.
Past a spent daily allowance, one message still goes through if your agents worked in the last 30 minutes, so a job isn't cut off one step from the end. Scheduled tasks never get it.
The fleet list shows only the caller's own agents, and agents on your other decks are for admins unless an admin turns on Every user's agents too in Console → Network.
Public file links last 7 days by default and 30 at most. Old links that never expired get an end date at least a week away, and each owner is told on the Line.
The desktop app no longer shows Captain Swarm: its one user is the local admin. Its agents had 69 built-in tools then, every other deck's 70.
Upgrade From 0.8.5 to 0.8.8
The same steps take a deck from 0.8.5, 0.8.6 or 0.8.7 to 0.8.8. Console → Updates does it on a deck on a git checkout.
Two changes to know first: a Bat run with no cap now stops at $50 (since 0.8.6; FD_BAT_DEFAULT_LLM_USD=0 for none), and public file links expire.
cd ~/captain-claw && git pull
pip install .
# restart Flight Deck, then every agent, then hard-refresh# once, as root, with the deck's Python
python -m playwright install-deps chromium0.8.7 4 November 2026
A new install got a LiteLLM that needs mcp on every turn without it, so every agent turn failed. Captain now depends on litellm[mcp].
The first since 0.8.0, for macOS, Windows and Linux x64 and arm64, attached to the release.
Its package is built from a clean export of the tagged commit. Release notes
0.8.6 4 November 2026
The Line, a way to make agents measurably better at real work, 131 archetypes, and a desktop app that went from 2.5 GB to 75 MB.
One timeline per person; the bell keeps only what needs you.
31 Generalists and 100 Specialists, and an admin can generate the company's own.
75 MB, installs the rest on first start, and updates itself, going back if a version doesn't come up.
Keeps improving one number you care about: several ideas at once, every attempt measured on cases it keeps sealed, and only a gain bigger than the noise kept. Admins only while it's tested.
Checks it runs itself, fix rounds from what reviewers find, up to three steps in parallel, and one budget for the whole run. It asks you at a stalemate.
Families in a side rail, an icon per archetype, a detail pane that's never empty. Forge agents builds the team size you ask for, up to 12.
Charm's Crush as a fifth coding worker, and web search with no key. With line and captain_swarm, 70 built-in tools per agent.
0.8.5 2 November 2026
Coding agents working for you in Crew, a team forged from a plain request in Forge agents, a Taskboard that hands over many tasks at once, and a deck that updates itself.
xlsxOnly what a change touches is rewritten, so charts, pivots and macros stay. Formulas stay live, with results from about 200 functions compared with Excel, and every call lists the cells and totals that moved.
pptxSlides from the deck's own layouts, so a firm's template keeps doing the design. Text is measured and every result says what doesn't fit. With crew, that makes 68 built-in tools.
Folders in the kiosk's agent list, with a name and a colour, that fold away and still show an agent at work or a new reply. Pin a group, and pin agents within it.
A test start, a two-minute warning, a restart in place and a way back. How it works
Scheduled and linked turns now carry the agent's instructions, even if nobody ever opened its chat, on a deck with accounts. A link follows its agent to a new port.
Last login, tokens used this month and all time, and Ask for feedback, an email from your own Gmail or Outlook.
hosting now works behind a proxy and under lockdown, and hands back a public link. A static site really stops, keeps its address, and comes back on Start.
The robot from the icon is the only character now. The StopWatch has a new look and doesn't stop on a pause, and the PaperColor's dashboard is blocks of colour. A weak Wi-Fi link no longer clogs a device. Firmware 0.20.4, PaperColor B 1.0.31.
Updates Console → Updates
On a git checkout, it follows the branch it's on, so a fork follows the fork. It only moves forward, to the commit you saw.
The new commit runs from a temp folder on a spare port, with an empty data folder and none of the deck's keys. If that fails, nothing has been touched.
Everyone signed in gets a toast, then one they can't close, counting down two minutes.
Mid-turn ones can be waited for, up to 5 more minutes. Crew's workers keep going.
New packages went in before the test start, so you don't run pip yourself.
The same process, command line and tmux pane, so it can't lose the environment you started it with.
If it doesn't come up, the old commit goes back and the deck starts on that, by itself. Every admin gets an email saying how it went.
Off by default, or Tell me, or Install by itself, where whatever is merged to the branch goes live once it passes the test start. A desktop build or a container is updated the way it was installed.
Upgrade From 0.8.4 to 0.8.5
There are no new Python packages, and BotPort hubs need nothing new. This one is by hand; after it, the console does it. Approving someone for Crew trusts them with the server, so read this first. Crew stays off for non-admins until you tick it.
cd ~/captain-claw && git pull
pip install .
# restart Flight Deck, then every agent, then hard-refresh# for Crew: install the CLIs on the deck's host, then
# tick Crew per person on Users, and save logins in Crew → Logins
# new device firmware, from firmware/device
python tools/publish.py <env> --deck https://<deck>Before you approve anyone for Crew · Upgrading to 0.8.5 · 0.8.5 release notes
0.8.4 30 October 2026
Teach an agent your work and exam it, then see everything it was asked on one Taskboard.
Documents, lessons, exams and practice on real tasks.
Every task on one board, and a hand-over that checks the agent did it.
A small paper for your desk, with or without touch.
With sign-in on, every path an agent names is checked against its folder, its workspace, its owner's VFS and the call's shared folders. .env is out of reach.
They lose shell, code, bat and the other tools that run code or drive the host, and their web tools take public addresses only.
Non-admins see a team key as ···· and its last four, and it goes only to the endpoints your published team sets use. Anything else is re-keyed or refused.
Anthropic, OpenAI, Gemini, OpenRouter or xAI, checked and sealed. While it's on, your agents run on your keys and no daily limits apply.
Fill it in for me reads your documents and links, looks you and your company up, and drafts your texts. Nothing reaches your agents until you save it.
visualize and knowledgeHTML in, the page and a PNG out, with Chart.js, ECharts, Plotly, D3, Mermaid or Leaflet. That made 65 built-in tools per agent.
Pictures reach the chat and the channels, a Files menu keeps the last five files, Insert now hands a queued message to the running turn, and automated results stay small.
The Waveshare ePaper 1.54, with or without touch. And on every device, the agent's dot says why it can't be reached: blue while it sleeps, red when it's off.
This is containment, not an OS sandbox: an admin's agent still has a shell that reaches the whole host. · The team checklist
Coming from 0.8.3? The 0.8.4 steps take away the tools that run code from non-admins' agents, so read those first. · 0.8.4 release notes
0.8.3 27 October 2026
Agent links, your own WhatsApp, a register of every tool call, notifications everywhere, and idle agents that sleep.
0.8.2 Also in that release
Users put the agents they own on the network from the agent's ⋯, after a dialog that says plainly what other decks can and can't do with them. Admins keep a switch and a Stop button.
The devices' voice plays through a ring instead of one buffer, so it no longer stops at about 66 seconds. The PaperColor goes all white once an hour to keep its colours clean.
web_fetch keeps the pictures a page shows that matter to the research, and drops logos, ads and icons. Markdown links them, HTML embeds them, Word places them.
An admin token for the dashboard and API, constant-time key checks with a failure limiter, deck identity bound to its key, and agent tokens sent only to the agent's own address.
Coming from 0.8.1? The hub first, then the decks.
· The BotPort guide, from pip install on a Raspberry Pi to the first request between two decks
0.8.1 24 October 2026
Jev (TypeSafe AI) answers yes/no, choice and rubric questions with probabilities. Each user connects their own key; agents get a jev tool, flows a tool: jev step, the MCP server jev_ask, and Connections a playground.
A standalone B2B research desk built purely on Flight Deck's HTTP API: streams of research with a plan gate, reports with a receipts panel (verdict, confirmed and refuted claims, cost with ROI), standing briefs and a Basna second opinion — plus Kalup, a Pack Studio that turns a desk description into a published vertical desk.
Thinking models stop 400-ing; OpenAI reasoning effort per tier; local servers shrink an output budget their window can't hold instead of failing; Google subscription text through the Antigravity CLI; You.com web search. Agents start fast offline: 75 s → 0.1 s with the Hub unreachable.
Opt-in for long-form runs: deterministic checks that can block, model passes that only revise, advisory scores — and a per-tier Weak / fast model checkbox so the pipeline tends a small model instead of trusting it.
A Reports view has a Deep-Researcher narrate how the village of living beings worked over any period — Quick or Deep, numbers straight from the ledger, saved and readable in a proper reader.
The Task Planner distributes a reviewed plan round-robin across lanes; chat stays fast on older machines; an archetype grid composes reviewer.legal-style roles from a function and a domain; each user binds their own WhatsApp number; six opt-in Loops & Graphs levers.