teams-cli

Microsoft Teams from your terminal—human-first and agent-first.

teams-cli is a Rust CLI and TUI for everyday Microsoft Teams work through supported Microsoft Graph APIs, with an optional experimental desktop backend that reads chats from, and can send chat messages through, a running Teams client on macOS, Windows or WSL. It is designed as two equally first-class experiences over one typed core: a fast, calm terminal interface for people and a deterministic, introspectable command contract for agents. The default Graph backend signs in as you with delegated OAuth and protects refresh credentials in your operating system's credential store. Output is readable text on a TTY and structured JSON when piped.

For humans, no arguments open guided onboarding on first use and the TUI afterward. For agents, no command ever waits for an undeclared prompt: teams schema works offline, collection sizes are caller-bounded, writes declare their effects, errors have stable kinds and exit codes, and stdout contains data only.

Install

pipx install teams-cli
# or
cargo install teams-cli --locked
# or run from PyPI without installing
uvx teams-cli --help

The PyPI distribution provides both teams and teams-cli as executable names.

To work from source:

cargo install --path .
# or build a local wheel
maturin build

Development checks (make check or cargo test --all-targets) also need Python 3 and Node.js 20 or newer for release-smoke and synthetic desktop-script fixtures.

Explore before configuring Microsoft 365

teams tui --demo
teams tui --demo --snapshot
teams capabilities
teams schema --command "messages send"

Onboarding

Run teams in a terminal on first use, or start explicitly:

teams init

teams-cli includes a maintained multitenant Microsoft Entra public-client registration, so normal setup does not require creating an app or copying a client ID:

teams init

Browser PKCE is the human-friendly default. The browser talks directly to Microsoft, and the CLI never receives or stores your password. Normal sign-in requests these delegated scopes:

openid profile offline_access User.Read Team.ReadBasic.All
Channel.ReadBasic.All Chat.Read Chat.Create ChatMessage.Send
ChannelMessage.Send

Terminal interface

teams (or teams tui) opens a keyboard-first chat reader: recent chats on the left with unread markers, last-message previews and times, and the selected conversation on the right, grouped by day and sender. Reads run in the background and are cached per chat, so moving through the list never blocks, and chats refresh on their own without losing your place: on the desktop backend the open chat every 3 seconds and the list every 15, on Graph every 15 seconds and every minute. The header always shows the profile and account, whether it can send (✎ can send or read-only), a status dot (● live, or ● stale since the last good read while refreshes fail), and your chat and unread counts. Narrow terminals show one pane at a time, and NO_COLOR is respected.

The mouse works too: the wheel scrolls whichever pane is under the pointer (older messages load when you scroll past the top of a chat), and a click opens a chat or focuses the conversation. Because the reader captures the mouse, hold shift while dragging to select text in most terminals.

The TUI works with both the Graph and desktop backends. It sends only on a profile that allows writes (Graph without read_only, or desktop created with --allow-send), and pressing enter sends directly; writability is checked again before every send. Use ctrl-j or alt-enter for a new line. Multiline pastes stay in the draft until you press enter (in terminals supporting bracketed paste). Like messages send, a send is never retried: if it fails, the draft is kept, the chat reloads so you can see whether it arrived, and nothing is resent until you press enter again. Drafts live in memory only.

The TUI is also the recovery path: an unconfigured profile, missing credential, expired refresh session, or Microsoft 401 opens a focused connection screen. Press Enter or a to restore the normal terminal, complete Microsoft sign-in, and return directly to the inbox; q exits without changing anything.

Reading channel history is intentionally separate because ChannelMessage.Read.All requires administrator consent. Request it only when needed:

teams auth login --channel-history

The CLI surfaces missing consent as a stable permission_denied error instead of silently reaching for undocumented Teams APIs. Tenant policy can still require an administrator to approve otherwise user-consentable permissions.

Headless setup is explicit and never prompts:

teams init --no-login
teams auth login --device-code
teams whoami --output json

Organizations that prefer their own Entra registration can override the bundled application:

teams init --client-id "$TEAMS_CLIENT_ID" --tenant organizations --no-login

The custom app must be a public client with http://localhost registered for browser login, public-client flows enabled for device-code login, and the baseline delegated permissions listed above.

For short-lived automation, TEAMS_ACCESS_TOKEN overrides the stored credential. Refresh credentials are never written to the config file.

Everyday commands

teams list
teams channels list TEAM_ID
teams chats list --limit 25
teams messages list --chat CHAT_ID
teams auth login --channel-history  # once, with administrator approval
teams messages list --team TEAM_ID --channel CHANNEL_ID
teams messages send --chat CHAT_ID --body "On it."
printf 'Status update' | teams messages send --chat CHAT_ID --body -
teams auth status
teams auth status --offline
teams profile list
teams profile use work
teams profile remove old --yes
teams config show
teams config path
teams doctor --offline

Collection commands expose --limit and --fields; Graph-native collections use --cursor, while joined teams and channels use --offset because those endpoints do not accept $top. With a terminal they render concise text; when stdout is piped they emit JSON envelopes with continuation metadata and truncated.

Profiles can block every remote write while keeping reads, sign-in, diagnostics, and the TUI available:

teams init --no-login --read-only
# or for one invocation/environment
TEAMS_READ_ONLY=true teams messages send --chat CHAT_ID --body "Blocked"

Like the other tools in the suite, the config file is selected with --profile, overridden by TEAMS_* environment variables, and inspectable without exposing credentials through teams config show and teams config path.

Experimental desktop backend

Use a desktop profile to inspect an existing Teams debugging endpoint on macOS, Windows, or WSL, and to read chats (and, if you opt in, send chat messages) on all three platforms with a signed-in personal Microsoft account or a work/school account:

teams init --profile local --backend desktop --read-only --no-login
teams --profile local doctor --offline
teams --profile local doctor
teams --profile local auth status
# Match a separately enabled endpoint on another port:
teams init --profile local --backend desktop --desktop-port 19222 --no-login
# If Teams exposes personal and work/school pages, choose the intended surface:
teams init --profile local --backend desktop --desktop-port 19222 --desktop-account personal --no-login
# macOS/Windows/WSL: read chats and chat messages
teams --profile local chats list --limit 20
teams --profile local messages list --chat '19:...@thread.v2'
# macOS/Windows/WSL: opt in to sending plain-text chat messages
teams init --profile local --backend desktop --desktop-port 19222 --desktop-account personal --allow-send --no-login
teams --profile local messages send --chat '19:...@thread.v2' --body 'On my way'

Personal-account read, send and readback have been tested with released binaries on Apple Silicon macOS, native Windows and WSL2. Work/school read, send and readback have also been reported working by a WSL2 user; macOS work/school cache reads were tested separately. See the release validation scope for coverage and remaining limits.

On macOS, Windows and WSL, chats list and messages list --chat read from the running Teams client's local cache: the backend evaluates one fixed, read-only script in the Teams page and returns projected fields (ids, topics, members, last-message previews, read state, timestamps, senders, bodies), never tokens or raw records. Member and sender names come from the matching account's cached profiles. The signed-in user's identity is resolved from that same cache; if a personal profile is missing or conflicting, the viewer stays unknown rather than guessing. Personal and legacy Skype user ids retain their 8:live: or 8: prefixes. Pages use the same shape and opaque --cursor as Graph and add "source": "desktop_cache". The cache only holds what Teams has synced, so older history in a chat you never scrolled back through can be missing. Use the chat ids returned by chats list; personal chats are accessed through the desktop backend. The chat with yourself is listed too, with the id 48:notes, so you can try sending there first.

Desktop profiles are read-only unless created with init --allow-send; TEAMS_READ_ONLY=true still blocks sending, but TEAMS_READ_ONLY=false cannot unlock a desktop profile. With the opt-in, messages send --chat sends one plain-text message through the running Teams window, the same way its Send button does: Teams performs the request with its own sign-in, and nothing is shown or clicked in its window. Each send asks for confirmation on a terminal and requires --yes otherwise. The listener is verified again and a fresh connection to the Teams main window is used for every send, and a send is never retried: if Teams does not answer, the command reports that the message may still be delivered (error kind unconfirmed, exit code 8), so check the chat before sending again. Teams queues the message like any other, so a send made just before Teams closes goes out when Teams next starts. The result has the Graph send shape plus "source": "desktop", clientMessageId, Teams' own status and a normalized delivery_state: submitted, queued, delivered, failed or unknown. A successful submission is not proof of delivery. doctor checks that the send path is still recognised without sending anything.

An uncertain desktop send also returns chatId and clientMessageId in its JSON error. Check the exact ID with:

teams --profile local messages status --chat CHAT_ID --client-message-id CLIENT_MESSAGE_ID

This checks the newest 50 cached messages and returns metadata only. Missing cache records remain unknown, not failed or safe to retry. The ID is a correlation ID, not an idempotency key. In the TUI, a possibly-sent draft blocks ordinary Enter and paste; r checks delivery, s offers an explicit second-copy confirmation, and esc keeps the draft and uncertainty together in memory. Reopening it does not unlock ordinary sending. Queued results are labelled queued rather than sent.

Desktop cannot create new chats. Teams, channels, identity commands, subscriptions, replies, mentions, attachments and edits return unsupported on desktop. The TUI reads chats from the same cache and sends through the same path as messages send. Existing profiles without a backend field continue to use Graph; Graph remains the default for init.

Use a managed debugging session on macOS, Windows or WSL:

teams --profile local desktop connect
teams --profile local desktop status
teams --profile local desktop disconnect

Connect and disconnect ask before restarting Teams because a restart ends active calls; pass --yes for a scripted restart. --port overrides the selected profile's port. An existing verified endpoint is inspected without being adopted; disconnect only restarts recorded Teams processes that still own the debugging listener. Added auxiliary processes are left alone. If Teams was restarted independently, quit and relaunch it normally, then disconnect again to clear the old record. No command automatically restarts Teams on exit.

On macOS, flags apply only to the managed launch. On Windows/WSL, setup saves the previous user environment value, applies flags temporarily while Teams starts, then restores that value. Unrelated changes to the value are preserved and reported. Windows and WSL share a recovery record and lock. WSL finds PowerShell in its normal Windows system directory even when Windows PATH entries are disabled; Windows executable interop must still be enabled.

If setup is interrupted, run teams --profile local desktop recover. It restores pending settings without restarting Teams or requiring the Windows helper; it does not claim ownership of an unrecorded Teams process. Use desktop status to inspect the endpoint and saved session. Recovery cannot close an already-running debugging endpoint: quit Teams and relaunch it normally if its ownership was not recorded.

While the port is open, any program running as you can read chats and send as you. The WSL TUI offers the same managed connect after explicit restart confirmation. teams init --backend desktop sets the connection up end to end. When no endpoint listens it runs the same managed connect after the same restart confirmation (--yes when scripted), then reads one chat page. It saves the profile only when that read succeeds, so session_verified: true means reads work, and a declined restart or a failed read leaves Teams and any existing profile of that name untouched. If the read fails after a restart, the error says the debugging port is still open and how to close it. init --no-login saves configuration without contacting Teams. doctor prints the managed opening and closing commands when no endpoint is found; it does not launch Teams or verify sign-in. A normal Teams launch usually has no endpoint, which doctor reports as no_listener.

macOS checks listener binding and installed Teams process ancestry before accessing loopback metadata or the Teams page. On Windows, teams.exe checks the listener itself through the Windows APIs: every listener on the port must be loopback-only and descend from an ms-teams.exe that belongs to the installed MSTeams package and runs as the invoking Windows user. Under WSL, the Linux CLI starts the Windows build as a helper (teams.exe desktop-helper) and keeps it running for the session, talking JSON over its stdin and stdout, so NAT mode needs no networking changes. Diagnostics are limited to GET /json/version and GET /json/list, and their results omit titles, full URLs, target IDs, WebSocket URLs and private origins. Chat reads then attach to the first Teams-origin page over the same verified loopback port and keep that connection open; if it drops, the listener is verified again before the next read. Related cache databases are selected for the same account. If a window exposes several conversation-account caches, reads and sends refuse with a desktop error; selecting among those accounts is not supported yet. --desktop-account auto is the default and refuses an endpoint exposing both personal and work/school pages. Set --desktop-account personal or --desktop-account work-school when creating the profile to select that surface explicitly. The choice applies to reads, sends and doctor's send check, and never falls back to the other surface. Multiple cached accounts within the selected surface remain unsupported.

On first use under WSL, the helper is downloaded from the matching GitHub release, checked against the release's SHA256SUMS, and cached in ~/.cache/teams-cli/windows-helper/<version>/. doctor --offline reports whether it is installed without downloading it. Online probes install a compatible release helper when needed. Set TEAMS_DESKTOP_HELPER to use another teams.exe speaking protocol 3 (including account selection), or run make windows-helper (needs cargo-xwin) to build one for a development checkout. If the download is blocked, fetch the x86_64-pc-windows-msvc archive from the release page, unpack teams.exe and point TEAMS_DESKTOP_HELPER at it. For older source checkouts still carrying 0.1.9-alpha.1, use make windows-helper or an explicit compatible helper: that published binary predates protocol 3 and cannot serve as the helper for current source.

The CLI and the helper agree on a protocol number, and any helper that speaks the CLI's protocol is accepted, whatever its version. When they differ, the error names both sides' version, build and protocol, along with the fix for how the helper was installed. teams --version shows the commit a binary was built from, with -dirty for uncommitted changes.

Endpoint discovery does not verify sign-in: signed_in stays null, verified and session_verified stay false, and endpoint_available describes metadata access only. doctor.healthy means the requested diagnostic checks passed; it does not read chats. When any check fails, doctor still prints the full report and then exits with code 7 (unhealthy). Offline commands do not contact the endpoint. Desktop profiles ignore Graph OAuth environment settings and credentials; auth login/logout return unsupported because Teams owns its session. A CLI profile selects connection configuration, not a Teams account. There is no automatic fallback to Graph.

Each read is bounded to 45 seconds. The WSL helper exits when the CLI does, though Windows can keep it alive until an in-flight read completes. Microsoft treats WebView2 flags as development diagnostics and advises against relying on them in production. This backend is experimental: chat reading depends on the Teams client's internal cache layout and sending on its internal send interface, and an unrecognised layout fails with desktop_error rather than guessing.

Graph boundaries

• The Graph backend supports Microsoft 365 work or school accounts only. Use the experimental desktop backend on macOS, Windows or WSL for personal Microsoft accounts; the relevant Teams Graph endpoints do not support them.
• A work account can authenticate successfully before Microsoft Teams is licensed or provisioned; teams doctor distinguishes identity access from Teams availability.
• Calling and screen sharing are not available as a delegated user CLI; Graph calling is built around bot/application scenarios.
• No extracted desktop tokens, browser interception, or direct calls to private Teams services; the experimental desktop backend reads the local client cache and sends only through the Teams window's own client.
• Tenant policy and admin consent can restrict delegated capabilities.

CLI Spec

teams schema is offline, unauthenticated, and follows CLI Spec v0.3: flat invocable command paths, declared effects and cardinality, bounded collections, stable error kinds and exit codes, stdout/stderr separation, and safe noninteractive behavior. Its backend capabilities and per-command backend extensions distinguish Graph operations from the experimental desktop backend.

For agents, each command declares the error kinds it can return, an example invocation, its output fields (nested where useful), argument defaults and limits, and the environment variables that affect it; teams schema validates against the official v0.3 JSON Schema in the test suite. Two error kinds matter most for writes:

rate_limit errors include retry_after_seconds when Microsoft provides it. Sends cannot be made idempotent, because neither Graph nor Teams accepts a caller-supplied key.

Releasing

Vership owns versioning, changelog generation, release commits, and tags. See the release runbook for the verified workflow and recovery policy.