Repository navigation
Publish guides to self-hosting the accounts app, answering a Codex Security budget prompt from a phone, resuming a Claude Code session on the same link, fitting an Antigravity CLI swarm board to a phone, opening a Claude Code session on a phone from a short terminal, handing Claude Code an MCP grant off the command line, pinning Claude Code's permission mode before starting it from a phone, and comparing GPT-6.2 Sol and Astra on one Codex review from a phone - #350
Open
artemiia wants to merge 8 commits into
Open
artemiia wants to merge 8 commits into
artemiia wants to merge 8 commits into
Conversation
…your own sign-in Anthropic opened a private beta of self-hosted Claude Managed Agents on 2026-11-12. This guide covers the shell.online accounts app with an OpenID Connect provider and PostgreSQL, pointing the CLI and the background service at your own relay and app, and what the setup does not do. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…om a phone Run a Codex Security 0.10.0 scan inside shell with --persistent, answer the dashboard's 80% budget prompt from the phone key bar, avoid the ctrl-c key that cancels the scan, and resume an interrupted scan on the same link. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…nline link Claude Code 2.1.323 lets a wrapper set the exit hint's resume command with CLAUDE_CODE_RESUME_HINT_COMMAND. The guide builds a claude-shared wrapper around shell --persistent so a resumed conversation returns on the same link and password, and says what the setup does not do. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…screen Adds a tutorial that runs agy under shell, opens the session in the app as its owner, and uses Fit to my screen so the /swarm board from Antigravity CLI 1.5.5 redraws at the phone's size, then restores the size with shell attach. Includes the post body, registry entry and cover. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ort terminal Measures shell.online's one-scan QR code against an 80x24 pane, lists when shell prints no code, and gives shell list, shell password and shell auth as fallbacks. Hooked to Claude Code 2.1.326's Remote Control fixes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…he command line
Adds a Security tutorial by Artemii Amelin, dated 2026-11-18. Claude Code
2.1.327 lets claude mcp add-json read the server JSON from stdin; the guide
issues an observe-only grant into a variable, registers it as a
${SHELL_WATCH_TOKEN} reference, starts Claude with the read tools only and
revokes the grant afterwards.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ing it from a phone Claude Code 2.1.328 starts sessions with no permission mode in auto mode for Opus 4.6 and Sonnet 4.6 on Bedrock, Agent Platform, Foundry and the Claude apps gateway. The web app's Claude Code launcher passes no mode, so the post sets permissions.defaultMode in user settings, explains the daemon environment, and shows how to read the mode back from the phone. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…w from a phone Adds the 2026-11-20 post: run one read-only Codex review on gpt-6.2-sol and gpt-6.2-astra as two named shell.online sessions, confirm each run's model with shell list --json, follow both from the web app, and compare the findings against the diff. Hooked to OpenAI's GPT-6.2 launch and Codex 0.171.0, which makes Sol the default catalog selection. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What changes for the user?
Adds eight blog posts to shell.online/blog/.
2026-11-13
SHELL_ONLINE_SERVER,SHELL_ONLINE_ACCOUNTSandSHELL_ONLINE_WEB. It also has a section on what the setup does not do.blog/posts.json(one new entry),blog/posts/how-to-self-host-shell-onlines-accounts-app-behind-your-own-sign-in.html,public/blog/banners/how-to-self-host-shell-onlines-accounts-app-behind-your-own-sign-in.webp2026-11-14
codex-security scans resume SCAN_IDfor interrupted Standard scans and--max-total-costfor batch scans. The guide runs a scan insideshell --persistent, answers the dashboard's 80% budget prompt from the phone key bar (esc, enter), explains why the ctrl-c key, Ctrl-C andshell killcancel the scan for good, and resumes an interrupted scan on the same link. It ends with what the setup does not do.blog/posts.json(one new entry),blog/posts/how-to-answer-a-codex-security-scans-budget-prompt-from-your-phone.html,public/blog/banners/how-to-answer-a-codex-security-scans-budget-prompt-from-your-phone.webp2026-11-15
CLAUDE_CODE_RESUME_HINT_COMMAND, which lets a program that starts Claude Code make the exit hint name its own resume command; 2.1.325 (2026-11-13) fixed recent turns going missing from a resumed session after an unclean stop. The guide builds a one-wordclaude-sharedwrapper aroundshell --persistent ... -- claude "$@", reads the exit hint, resumes on the same link and password, finds a conversation with--resumeor--continuewhen the hint is not to hand, and explains what the relay does on a persistent exit (session marked disconnected for 30 days, viewers see "Temporarily offline" with the last screen, MCP grants cleared). It ends with what the setup does not do.blog/posts.json(one new entry),blog/posts/how-to-resume-a-claude-code-session-on-the-same-shell-online-link.html,public/blog/banners/how-to-resume-a-claude-code-session-on-the-same-shell-online-link.webp2026-11-16
/swarmas a preview with/swarm board, a full-screen live view of the swarm's members and posts. The guide runsshell --name swarm -- agyon a linked machine, opens the session in the app on a phone as its owner, and uses Fit to my screen (and the app's automatic fit on open) so the board redraws at the phone's size. It explains what the fit does (owner-only button, relay drops it for--read-onlysessions, host clamps it to an attached terminal,TIOCSWINSZandSIGWINCH), restores the size withshell attach, and ends with what the setup does not do.blog/posts.json(one new entry),blog/posts/how-to-fit-an-antigravity-cli-swarm-board-to-your-phones-screen.html,public/blog/banners/how-to-fit-an-antigravity-cli-swarm-board-to-your-phones-screen.webp2026-11-17
/remote-controlQR code being cut off in short windows such as 80x24, and Remote Control showing a session as running while it waited on a prompt that can only be answered in the terminal. The guide starts Claude Code withshell claude, measures shell's own one-scan QR code with the encoder the repository pins (103-character payload, QR version 5, 23 rows and 47 columns, 27 rows from heading to prompt), explains when the card has no QR at all (NO_COLOR,TERM=dumb, standard error not a terminal, bitmap over 57 modules), givesshell list,shell password <ID>andshell authas fallbacks, and shows why a terminal-only prompt reaches the phone. It ends with what the setup does not do.blog/posts.json(one new entry),blog/posts/how-to-open-a-claude-code-session-on-your-phone-from-a-short-terminal.html,public/blog/banners/how-to-open-a-claude-code-session-on-your-phone-from-a-short-terminal.webp2026-11-18
claude mcp add-json <name> -, which reads the server's JSON from standard input "so tokens can stay off the command line". The guide issues a one-hour observe grant withshell mcp grantinto an exported variable, registershttps://shell.online/mcpwith Claude Code through a quoted here-document so.mcp.jsonholds a${SHELL_WATCH_TOKEN}reference rather than the bearer, approves and checks the server withclaude mcp list, starts Claude with--allowedTools "mcp__shell-watch__*"and a deny rule forshell_send, and ends the access withshell mcp list,shell mcp revoke,shell mcp revoke-allandshell password rotate. Its limits section says the bearer is visible to Claude Code's own environment, that docs/MCP.md records Claude model-backed behavior as unverified and that the pairing was not run for the post.blog/posts.json(one new entry),blog/posts/how-to-hand-claude-code-a-shell-online-mcp-grant-off-the-command-line.html,public/blog/banners/how-to-hand-claude-code-a-shell-online-mcp-grant-off-the-command-line.webp2026-11-19
--permission-mode(only--resume,--modeland the Skip permission checks toggle's--dangerously-skip-permissions), so the guide pinspermissions.defaultModein~/.claude/settings.jsonwith jq, explains why a provider exported in a shell profile does not reach a session the installed service starts (the service file carries onlySHELL_ONLINE_*variables), starts the session from the phone, reads the mode from Claude Code's status line (which the Chat renderer hides, so Adaptive or Fixed grid), and starts auto mode on purpose withshell --name ... -- claude --permission-mode auto. It ends with what the setup does not do.blog/posts.json(one new entry),blog/posts/how-to-pin-claude-codes-permission-mode-before-you-start-it-from-a-phone.html,public/blog/banners/how-to-pin-claude-codes-permission-mode-before-you-start-it-from-a-phone.webp2026-11-20
gpt-6.2-solandgpt-6.2-astrato its catalogs with Sol as the default catalog selection. The guide stores one review prompt in a variable, starts two named background sessions withshell --name <name> -- codex --model <model> --sandbox read-only "$PROMPT", explains why options after--reach Codex, confirms each run's model fromshell list --json(and warns that the JSON carriese2ee_password), follows both from the web app or the printed links, reads token counts with Codex's/status, checks each finding againstgit diff main...HEAD, and closes with/quitorshell kill. It ends with what the setup does not do.blog/posts.json(one new entry),blog/posts/how-to-compare-gpt-6-2-sol-and-astra-on-one-codex-task-from-your-phone.html,public/blog/banners/how-to-compare-gpt-6-2-sol-and-astra-on-one-codex-task-from-your-phone.webpVerification
The branch now holds all eight posts and is merged with current
upstream/main(already up to date, no conflicts). Checks run on 2026-11-20 for the new post and on the branch as it now stands, with the earlier posts' results kept below.2026-11-20 post:
python3 tools/lint_post.py how-to-compare-gpt-6-2-sol-and-astra-on-one-codex-task-from-your-phone: OK, 1996 words, 25 links (12 outside across 8 sites, 3 to pilotprotocol.network), no warnings.python3 tools/check_links.py blog/posts/how-to-compare-gpt-6-2-sol-and-astra-on-one-codex-task-from-your-phone.html: 25 links, 24 ok, 1 blocked, 0 dead. The blocked link is the Let's Data Science article (403 to scripts), which was read through WebFetch during the run.npx vitest run tests/blog.test.ts: 15 tests passed (run on the branch with all eight posts).node scripts/test-landing-seo.mjs: passed.npm run build:web: exit 0. The only output was Vite's existing chunk-size warning.cmd/shell/help.go(start options,--,shell list,shell password rotate,shell kill,shell auth),cmd/shell/main.go(Goflagparsing),cmd/shell/session_output.go(start card fields,Closes when the task exits),cmd/shell/sessions.go(shell listcolumns, 48-character command truncation, compact layout under 132 columns,--jsonfields includingcommandande2ee_password, quoting of recorded arguments),app/src/terminal/tabs.ts(up to eight tabs, labelled by name) anddocs/content.json(mobile page). The start card andshell listcode match the v0.25.0 tag. The mobile page's fixed phone grid card was not used, as it is known to be stale. Codex behavior was checked against the rust-v0.171.0 release notes, pull request 62106 (model slugs, Sol as default catalog selection),codex-rs/exec/src/cli.rsandcodex-rs/tui/src/slash_command.rsat the rust-v0.171.0 tag, the Codex README and the Codex command reference on learn.chatgpt.com. Codex was not run for the post. OpenAI's announcement page refused automated fetches, so the post links the two outlets that were read and names OpenAI's announcement as their source.2026-11-19 post:
python3 tools/lint_post.py how-to-pin-claude-codes-permission-mode-before-you-start-it-from-a-phone: OK, 1987 words, 22 links (13 outside across 8 sites, 3 to pilotprotocol.network), no warnings.python3 tools/check_links.py blog/posts/how-to-pin-claude-codes-permission-mode-before-you-start-it-from-a-phone.html: 22 links, 22 ok, 0 blocked, 0 dead.npx vitest run tests/blog.test.ts: 15 tests passed (run on the branch with all seven posts).node scripts/test-landing-seo.mjs: passed.npm run build:web: exit 0. The only output was Vite's existing chunk-size warning.cmd/shell/help.go(shell auth --allow-remote-start,shell daemon status,shell list,shell attach, Ctrl-X D,--name),app/src/lib/session-kinds.ts(Claude Code command builder and the Skip permission checks toggle),app/src/components/NewSessionModal.tsx(New session sheet, Run on, command preview, Start session),cmd/shell/browser_command_unix.go(browser commands run as argv),cmd/shell/service.go,service_linux.goandservice_darwin.go(sixSHELL_ONLINE_*passthrough variables),app/src/terminal/chat/agents/claude-code.ts(status lines, including the mode line, left out of the Chat view),app/src/terminal/chat/keys.ts(Tab sent without Shift) andapp/src/routes/Workspace.tsx(Renderer options). The launcher has been unchanged since 2026-09-09. Claude Code behavior was checked against the v2.1.328 release page and the permission modes, settings and Amazon Bedrock docs on code.claude.com. Claude Code was not run for the post.2026-11-18 post:
python3 tools/lint_post.py how-to-hand-claude-code-a-shell-online-mcp-grant-off-the-command-line: OK, 1835 words, 23 links (14 outside across 9 sites, 2 to pilotprotocol.network), no warnings.python3 tools/check_links.py blog/posts/how-to-hand-claude-code-a-shell-online-mcp-grant-off-the-command-line.html: 23 links, 23 ok, 0 blocked, 0 dead. (An earlier run reported the GNU Bash manual page as dead with status 000 after a slow response; the post now cites the POSIX here-document section instead.)npx vitest run tests/blog.test.ts: 15 tests passed (run on the branch with all six posts).node scripts/test-landing-seo.mjs: passed.npm run build:web: exit 0. The only output was Vite's existing chunk-size warning.cmd/shell/help.go(shell help mcp),cmd/shell/sessions_unix.go(shell mcp grantprints only the bearer on standard output;list,revokeandrevoke-alloutput),cmd/shell/sessions.go(six-character ID prefixes),shared/mcp-grants.ts(observe default 3600 s, cap 12 hours or the remaining session lifetime),worker/index.ts(/mcpPOST handler),app/src/terminal/TerminalPane.tsx(MCP · server-side decryption authorized) anddocs/MCP.md. The grant commands were read in the Unix code path; the post says macOS and Linux. Claude Code behavior was checked against the v2.1.327 release page and the MCP and permissions docs on code.claude.com (stdin form,--scope project,${VAR}expansion inurlandheaders, pending approval and health statuses,mcp__<server>__*allow globs,claude mcp remove --scope). Claude Code was not run against a grant; the post says so.2026-11-17 post:
python3 tools/lint_post.py how-to-open-a-claude-code-session-on-your-phone-from-a-short-terminal: OK, 2019 words, 22 links (13 outside across 9 sites, 2 to pilotprotocol.network).python3 tools/check_links.py blog/posts/how-to-open-a-claude-code-session-on-your-phone-from-a-short-terminal.html: 22 links, 22 ok, 0 blocked, 0 dead. (A first run reported the GitHub CHANGELOG.md page as dead with a 503; the post now links the v2.1.326 release page instead.)npx vitest run tests/blog.test.ts: 15 tests passed (run on the branch with all five posts).node scripts/test-landing-seo.mjs: passed.npm run build:web: exit 0. The only output was Vite's existing chunk-size warning.cmd/shell/help.go,cmd/shell/qr.go(password added to the fragment, Low error correction, 57-module limit, half blocks, SGR 30;107),cmd/shell/session_output.go(card on standard error, QR only whenNO_COLORis empty,TERMis notdumband the writer is a character device),cmd/shell/sessions.go(shell passwordtakes an ID or a prefix of at least six characters; 32-character session IDs; ten-character short ID),cmd/shell/persistent.go(22-character salt) anddocs/content.json. QR sizes were measured by encoding payloads of those lengths with the pinnedgithub.com/skip2/go-qrcodeversion in a scratch program. Claude Code behavior was checked against the v2.1.326 release page, CHANGELOG.md, the npm registry publish time and the Remote Control and security docs on code.claude.com.2026-11-16 post:
python3 tools/lint_post.py how-to-fit-an-antigravity-cli-swarm-board-to-your-phones-screen: OK, 1990 words, 24 links (14 outside across 8 sites, 3 to pilotprotocol.network), no warnings.python3 tools/check_links.py blog/posts/how-to-fit-an-antigravity-cli-swarm-board-to-your-phones-screen.html: 24 links, 24 ok, 0 blocked, 0 dead.npx vitest run tests/blog.test.ts: 15 tests passed (run on the branch with all four posts).node scripts/test-landing-seo.mjs: passed.npm run build:web: exit 0. The only output was Vite's existing chunk-size warning.cmd/shell/help.go,cmd/shell/terminal_grid.go(10 to 500 columns, 4 to 300 rows, local-terminal limit, background launch grid),cmd/shell/session_unix.go(grid_request handling, attach sets and detach clears the limit),cmd/shell/process_unix.go(pty.Setsize),worker/index.ts(forwardGridRequestdrops requests on read-only sessions),app/src/terminal/TerminalPane.tsx,app/src/terminal/adaptive/adaptive-terminal.tsandapp/src/routes/Workspace.tsx(owner-only Fit to my screen, Adaptive renderer only, 11 px whole-grid and 13 px reading font). Dynamic grids shipped in v0.23.2. The automatic fit on open is listed under Unreleased in CHANGELOG.md; the post says so, and I confirmed the bundle served at app.shell.online today contains it (terminalAutoFit). The post says Windows was not checked.docs/content.json's mobile card still describes the older fixed 80x40 phone grid; the post follows the code and does not repeat that detail. Antigravity behavior was checked against the 1.5.5 release notes, the/swarmand model docs on antigravity.google, and the Google and Antigravity announcement posts.2026-11-15 post:
python3 tools/lint_post.py how-to-resume-a-claude-code-session-on-the-same-shell-online-link: OK, 1731 words, 25 links (14 outside across 8 sites, 2 to pilotprotocol.network), no warnings.python3 tools/check_links.py blog/posts/how-to-resume-a-claude-code-session-on-the-same-shell-online-link.html: 25 links, 25 ok, 0 blocked, 0 dead.npx vitest run tests/blog.test.ts: 15 tests passed (run on the branch with all three posts).node scripts/test-landing-seo.mjs: passed.npm run build:web: built, anddist/blog/how-to-resume-a-claude-code-session-on-the-same-shell-online-link/index.htmlis present. The only output was Vite's existing chunk-size warning.cmd/shell/help.go,cmd/shell/main.go,cmd/shell/persistent.go(state file fields, 700/600 modes, access-mode mismatch error),cmd/shell/session_reservation.go(second run refused),cmd/shell/summary_claude.go(--settingsand--session-idadded on a linked machine),cmd/shell/background_unix.go,process_unix.goandsession_unix.go(environment passed to the wrapped process),worker/index.ts(persistent exit handler),shared/session-lifetime.ts(30 days and 12 hours),shared/host-presence.ts(offline notice text) anddocs/content.json. None of these differ from v0.25.0 in a way the post relies on. Claude Code behavior was checked against the v2.1.323 and v2.1.325 release pages, CHANGELOG.md and the env-vars and CLI reference docs.2026-11-14 post:
python3 tools/lint_post.py how-to-answer-a-codex-security-scans-budget-prompt-from-your-phone: OK, 1969 words, 24 links (14 outside across 9 sites, 2 to pilotprotocol.network), no warnings.python3 tools/check_links.py blog/posts/how-to-answer-a-codex-security-scans-budget-prompt-from-your-phone.html: 24 links, 24 ok, 0 blocked, 0 dead.npx vitest run tests/blog.test.ts: 15 tests passed (run on the branch with both posts).node scripts/test-landing-seo.mjs: passed.npm run build:web: built, and both posts' pages are indist/blog/<slug>/index.html. The only output was Vite's existing chunk-size warning.cmd/shell/help.go,cmd/shell/main.go,cmd/shell/persistent.go(state file pins access mode),cmd/shell/session_unix.goandprocess_unix.go(SIGTERM to the process group, SIGKILL after 2 s),web/main.ts,web/terminal-keyboard.tsandweb/style.css(mobile key bar). Codex Security behavior was checked against its 0.10.0 release notes,sdk/typescript/docs/cli.md, and the source (cli-signals.ts,cli.ts,scan-dashboard.ts,api.ts) at commit 173483d. The post describes only behavior present in shell v0.25.0.2026-11-13 post (from the earlier run):
lint_post.py: OK, 1957 words, 28 links (16 outside across 11 sites, 3 to pilotprotocol.network), one warning for/self-hosting/linked twice.check_links.py: 27 links, 27 ok, 0 blocked, 0 dead (postgresql.org answered on the rerun).cmd/shell/help.go,cmd/shell/account.go,cmd/shell/service.go,cmd/shell/agent_loop.go,internal/account/,app/README.md,app/.env.example,app/docker-compose.yml,app/Dockerfile,app/server/lib/relay-proxy.ts,docs/self-hosting.mdanddocs/content.json.Access and privacy
No product change. All eight posts use only invented example hosts, paths, IDs, costs and passwords (the 2026-11-18 post shows the public endpoint
https://shell.online/mcpand no bearer), and all eight covers show only invented values, including the QR pattern drawn on the 2026-11-17 cover.Rollout
Website only. The posts go live after the next deploy of the site.
🤖 Generated with Claude Code