cli

Inline CLI

Install the Inline CLI and have Claude/Codex interact with it.

For setup failures, repeat the command with --verbose (twice for trace detail). See diagnostics and optional error reporting. Official binaries report metadata-only command, bridge-runtime, and panic failures to Sentry. Set INLINE_CLI_TELEMETRY=off to disable reporting. Command arguments, message contents, credentials, local paths, and raw errors are not sent.

Install

Homebrew (cask)

brew tap inline-chat/homebrew-inline
brew install --cask inline

Script

curl -fsSL https://inline.chat/cli/install.sh | sh

Supports macOS and Linux (x86_64/aarch64, glibc and musl). Set INLINE_INSTALL_DIR to choose a custom install directory.

Check for updates

inline update --check
inline update --check --json --compact
inline update

--check reads the release manifest without downloading a bundle, accessing credentials, changing local state, or restarting agents. A successful check exits zero, whether an update is available or not. JSON includes current_version, latest_version, target, supported, and update_available. inline upgrade is an alias for inline update; installing still requires omitting --check.

Updates stream downloads to private temporary storage and verify the checksum and optional manifest size before extraction; downloads are limited to 512 MiB. When installing without elevation across filesystems, the CLI stages a complete replacement beside the destination before renaming it into place. Temporary download files are cleaned up on normal success and failure paths. Homebrew installations can use brew upgrade --cask inline to keep Homebrew in charge of the installed version.

(Optional) Install the Inline skill for Codex

The CLI bundles the complete skill directory from skills/inline/, including its references and OpenAI metadata. Installation is local and does not require Node.js or a network request:

inline skill install

The command installs to $CODEX_HOME/skills/inline when CODEX_HOME is set, otherwise ~/.codex/skills/inline. It is idempotent when the installed files match. If an existing skill differs, review it before using inline skill install --force; force overwrites only bundled Inline files and preserves extra files.

For other supported agents, the broader Skills ecosystem installer can install the public source directly:

npx skills add https://github.com/inline-chat/inline/tree/main/skills/inline --global --yes

Login

inline login

inline login without phase flags remains an interactive terminal flow. On macOS, if a supporting Inline app is installed, the CLI offers to continue with the signed-in app or use email/phone. App login opens a local approval prompt with a verification code; approval creates a separate revocable CLI session and returns its credential over an ephemeral loopback connection. Agents and other non-interactive sessions can start and finish login explicitly:

# 1. Send a code. The V3 challenge is saved locally for the next command.
inline login --email you@example.com --send-code --json --compact

# 2. Verify the code and save the credentials without printing them.
inline login --email you@example.com --code 123456 --json --compact

# Avoid putting the code in argv when it is available on stdin.
printf '%s\n' "$LOGIN_CODE" | inline login --email you@example.com --code-stdin --json --compact

Phone login uses --phone in both steps. Run both commands with the same CLI configuration and API URL so completion can load the saved challenge. V3 login does not return or accept --challenge-token. The successful JSON result reports status: "authenticated", userId, and whether the profile loaded; it never includes the bearer token.

Agents and CI may also pass an existing token with INLINE_TOKEN:

INLINE_TOKEN=... inline me --json

inline me verifies the current auth state. inline logout clears the saved local token. It cannot unset a token provided by the parent environment, so commands remain authenticated while INLINE_TOKEN is set.

Local coding agents

One command can discover an agent already installed on this machine, create or reuse its Inline bot, install the Inline integration, configure owner-safe access, and start its gateway or bridge:

# Machine-readable, read-only discovery for apps and agents.
inline agents discover --json --compact

# Interactive picker; only installed agents are shown.
inline agents setup

# Prompt-free setup for agents and scripts.
inline agents setup --target openclaw --non-interactive --json
inline agents setup --target hermes --profile work --non-interactive --json
inline agents setup --target codex --folder /path/to/project --non-interactive --json

# Equivalent shortcuts for gateway targets.
inline setup openclaw --non-interactive --json
inline setup hermes --profile work --non-interactive --json

Supported installed targets are OpenClaw, Hermes, Codex, OpenCode, Claude, and Amp. Inline does not install those host runtimes. If none are installed, setup stops without creating a bot or changing configuration. Use --dry-run for a read-only discovery check, --no-install to forbid integration installation, or --no-restart to configure without changing the user service. The default access mode is owner-only; automation can use --access allowlist with repeated --allow-user ID flags. Codex/ACP bridge targets use the user's home directory as their workspace unless --folder is provided explicitly.

Unified discovery and setup JSON include protocolVersion and documentationUrl. Setup target, option and authentication failures also include status, failedPhase, changes, and retry; status: "partial" means setup may have changed state, while changes contains only confirmed completed work. Errors use stable codes and safe hints so an app or agent can offer a retry or send the user to the setup guide without parsing terminal output. Preflight failures report no completed changes, and target is null when target selection has not completed. Retry commands preserve supplied flags; correct invalid arguments identified by the error before retrying. Output omits bot tokens and discovered local paths; the retry command can include workspace paths explicitly supplied by the user. Native app protocol errors stay on one JSON line even without --compact. Global argument-parsing and CLI-configuration errors retain the generic CLI error envelope.

For gateway targets, ~/.inline/config.toml stores only the target, profile instance, bot user ID, and bot username needed for idempotent reconciliation. Tokens stay in OpenClaw or Hermes credential storage. Codex/ACP mappings remain in the existing bridge account state instead of being duplicated in this file.

Inline can also keep Codex and ACP agents available as private, local-first bots in your chats. The older provider-specific setup commands remain supported and use the same setup core. Setup is persistent; you do not run a new bridge per thread.

inline setup codex --folder /path/to/project
inline setup opencode --folder /path/to/another-project
inline setup claude --folder /path/to/project
inline setup amp --folder /path/to/project
inline bridge status

Codex 0.146.0+ is eligible; there is no exact-version allowlist or upper version limit. Setup checks the executable and authentication. Service launch performs a bounded, read-only catalog contract check; operation-specific capabilities are validated when used and incompatible operations fail closed. It can use the signed ChatGPT runtime without a standalone Codex CLI and leaves sign-in to Codex/ChatGPT. Agent directions work in the bot DM, direct mentions, replies, and followed threads, but only for the owner by default. Other stable Inline user IDs must be added locally with inline bridge operators add USER_ID; provider-specific overrides use --provider. OpenCode, Claude, and Amp are experimental provider paths for developers. Amp setup installs Inline's source-revision-pinned ACP adapter and uses the exact authenticated host Amp CLI accepted during setup. All configured providers can coexist under one supervised background service. See Local coding-agent bridge for provider support, folder switching, lifecycle, and security details.

During a turn the bot silently edits one open Working disclosure. Codex commentary and tools stay in order, with the working directory in its footer; normal mode includes reasoning summaries and exact structured action details. Verbose mode adds the complete retained textual provider record: reasoning content, output, progress, commands, paths, arguments, results, and raw item payloads. The bridge fills the existing Working message to Inline's text limit of 100,000 UTF-16 units before adding a densely packed silent continuation; it never starts a new message merely because a command or provider item began. Normal mode remains one bounded progress message with an explicit overflow notice. Completion collapses progress and sends the final answer separately with normal notifications. Approvals, questions, and attachments keep their existing delivery paths; generated-image bytes stay in the attachment instead of being repeated as base64 text. The persisted final-send identity is reused across retries and service restarts so an ambiguous response does not create another final answer.

Output

Interactive start page

Run inline with no arguments in a terminal for a quick introduction, useful commands, and links to help. The page is local and exits successfully: it does not read credentials, start an agent, check for updates, or connect to Inline. Commands and flags are highlighted, with a stacked layout in narrow terminals. The Inline wordmark uses a three-tone silver bevel within the letters and a full-height logo mark, with no outer shadow. Narrow terminals fall back to text; monochrome output keeps the same clean letter shapes without shading.

Use inline --help for the full command reference and inline <command> --help for details. Redirected no-argument calls and flags-only errors keep their existing behavior; the introduction is never mixed into JSON or piped output. NO_COLOR disables styling. Human output also respects CLICOLOR=0 and TERM=dumb; a nonzero CLICOLOR_FORCE forces color unless NO_COLOR is set.

Everyday shortcuts and scope

Singular command names work alongside the original groups: chat, thread, message, user, space, and bot. List commands accept ls; chat, message, and user get commands also accept view. Help displays these aliases. Common flags have short forms: -c for a chat, -u for a DM user, -L for a limit, -q for a search query, -f for a list filter, and -m for message text. -v still prints the version.

inline chat ls --space-id 7 --unread -L 20
inline chat ls --home --type thread
inline chat ls --type dm --pinned
inline search "release checklist" -c 123 --offset-id 500 -L 50
inline search -c 123 --filter documents --ids
inline message ls -c 123 --has-media --ids

Common chat state and organization operations use the same -c/-u peer flags and the same V2/V3 authenticated connection as reads and sends:

inline chats subthread --parent-chat-id 123 --message-id 456 --title "Follow-up"
inline chats move --chat-id 789 --space-id 7
inline chats archive -c 123
inline chats follow -c 789
inline notifications set-chat -c 123 --mode mentions
inline messages pin -c 123 --message-id 456
inline auth sessions
inline auth revoke-session --session-id 42

chats subthread is also available as create-reply, reply-thread, and create-subthread. follow-default removes an explicit follow preference. Per-chat notification modes are all, mentions, none, and inherit. Session revocation asks for confirmation; JSON automation must pass --yes. Use inline logout to end the current session; auth revoke-session ends another account session.

Scope filters combine with --filter and run before --limit/--offset in both human and JSON output. --home means chats outside spaces, including DMs and Home threads; use --type thread to exclude DMs. --unread includes manual unread marks. --id requires exactly one match and cannot be combined with pagination, which could otherwise conceal ambiguous matches.

Search accepts one quoted positional phrase or repeated --query/-q flags; the two forms cannot be mixed. --offset-id uses the existing search cursor. Words within one query are ANDed; repeated queries are ORed. Search also accepts --filter photos|videos|photo-video|documents|links|voice-memos, with optional query text. Media filtering runs on the server before its result limit. Time/media filters on message lists, and time filters on search, apply to the fetched page; they do not scan all history. --ids on message lists and search prints peer-local message IDs without fetching the chat catalog for names. It cannot be combined with --json or --translate.

Agent discovery and Codex plugin

inline capabilities [COMMAND...] and inline schema commands [COMMAND...] emit the actual public Clap command metadata as JSON. They run before config, telemetry, update checks, or auth, so agents can inspect one narrow command without network access or stale embedded docs:

inline capabilities --compact
inline capabilities messages send --compact
inline schema commands chats subthread --compact

Install the full Codex integration with inline plugin install. It delegates to Codex's idempotent plugin manager with fixed arguments and installs the Inline skill plus OAuth MCP definition. --dry-run prints those commands without changing Codex. inline skill install remains the skill-only path.

Shell completions

Generate completions from the CLI's current command tree. Generation is local: it does not read account configuration, connect to Inline, or start the runtime.

inline completion bash > inline.bash
inline completion zsh > _inline
inline completion fish > inline.fish

Bash can source inline.bash; put _inline in a directory on Zsh's fpath before running compinit; put inline.fish in Fish's completion directory. powershell and elvish are also supported. completions is an alias. The command writes the script to stdout and never edits shell startup files. Hidden internal commands and flags are omitted. A pipeline consumer closing early is handled quietly instead of printing a panic. --json is rejected because the output is a shell script.

The upstream generator has shell-specific limits: Fish does not offer flags for commands nested beyond two levels, such as bridge workspace add. PowerShell and Elvish stop detecting the command path at the first flag, so place global flags after the command path when using their completions (for example, inline message send --json).

JSON and text

Use --json for automation and --compact for pipelines:

inline messages list --chat-id 123 --json --compact

Use batch selectors when an agent already knows the relevant message IDs:

inline messages get --chat-id 123 --message-id 91,92,100 --json
inline messages download --chat-id 123 --message-id 80-100 --dir ./media
inline messages download --chat-id 123 --from-msg-id 600 --limit 50 --dir ./media

Selectors support single IDs (91), comma lists (91,92,100), ranges (91-100), and repeated --message-id flags. Batch downloads skip messages without media, report skipped/missing/failed counts, and prefix local filenames with the message date, MSG ID, media type, and media ID. For contiguous windows, use --from-msg-id ID --limit N with export/transcript/download.

For reviewable conversation bundles, prefer transcript/export:

inline transcript --chat-id 123 --limit 500 --download-media --output ./feedback-bundle
inline transcript --chat-id 123 --limit 500 --download-media --media-dir ./feedback-media --output feedback.md
inline transcript --chat-id 123 --from-msg-id 600 --limit 50 --download-media --output ./feedback-bundle
inline messages export --chat-id 123 --limit 500 --format json --output feedback.json
inline messages export --chat-id 123 --limit 500 --format jsonl --output feedback.jsonl
inline messages export --chat-id 123 --limit 500 --format csv --output feedback.csv

inline transcript is a shortcut for inline messages transcript, a markdown export optimized for reading: sender names, timestamps, replies, forwards, media links, file links, and hidden message IDs. Markdown uses CDN URLs by default; pass --download-media to download photos/files in one pass and rewrite transcript links to local paths. Messages without media are skipped during media download. If --output is a directory, or a no-extension path with --download-media, the CLI writes transcript.md/transcript.<format> there and defaults downloaded media to a media/ folder inside it.

--translate works in JSON output for messages list, messages search, the top-level search shortcut, and messages get:

inline messages get --chat-id 123 --message-id 91,92,100 --translate en --json

List commands can pre-filter JSON payloads before printing:

inline chats list --json --filter "launch"
inline users list --json --filter "sam"
inline bots list --json --filter "deploy"
inline messages list --chat-id 123 --has-media --json --compact
inline messages list --chat-id 123 --empty-text --forwarded

For advanced ad hoc analysis, jq is still useful on compact JSON:

inline messages list --chat-id 123 --limit 500 --json --compact | jq '.messages | length'
inline messages list --chat-id 123 --limit 500 --has-media --json --compact | jq -r '.messages[].id'

For message content from standard input, pipe or redirect the input. --stdin fails fast when stdin is an interactive terminal:

echo "hello" | inline messages send --chat-id 123 --stdin

Use --text-file with send/edit to preserve indentation and trailing newlines:

inline message send -c 123 --text-file ./update.md
inline message edit -c 123 --message-id 456 --text-file ./update.md
printf '  indented text\n' | inline message send -c 123 --text-file -

File input accepts nonempty UTF-8 up to 1 MiB and cannot be combined with --text or --stdin. The server's message limits still apply. Existing --text/--stdin trimming and precedence are unchanged; use --attach for files that should be uploaded instead of read as message text.

Human tables detect the current terminal width on Unix; COLUMNS overrides the detected width and also works in pipes. Chat/user/message tables and chat/ message details escape terminal control characters from content. JSON and exports preserve original strings. Set NO_COLOR=1 to disable color, or CLICOLOR_FORCE=1 to force color in a non-TTY. Non-JSON runtime errors print a short human report with an error code, and may include status, API error, response preview, hint, and examples.

inline chats list gives chat titles extra room and wraps long titles onto a second table row before truncating, so filtered chat searches stay readable.

Diagnostics

inline doctor --json reports system, client identity, config, path, and auth state diagnostics. Client identity includes the CLI client type/version, user-agent, OS version, and metadata header names sent to the API/realtime server.

Development

The workspace uses one Cargo target directory and disables incremental compilation to avoid large per-crate caches. Focused checks use the dev/test profile; release builds are not needed for local verification. Developers with spare disk space can opt in with CARGO_INCREMENTAL=1.

Before committing CLI changes, run the local pre-commit check:

bun run --cwd cli precommit

That uses the pinned Rust toolchain in cli/rust-toolchain.toml, runs cargo fmt, applies machine-fixable clippy suggestions, runs cargo fmt again, then runs the same bun run --cwd cli ci command used by GitHub Actions. CI keeps rustfmt as the formatting authority; clippy still denies warnings, but allows format-argument style churn such as inlined format args.

Notes

The CLI is still early and may have bugs.