Moonshot
Kimi Code CLI
2 days since its last release · 2.5.3 ·
Kimi Code is an AI coding tool from Moonshot. Its latest release is 2.5.3, released on 10 Nov 2026. It shipped 11 releases in the last 30 days, and we have tracked 92 since 26 May 2026.
- Latest
- 2.5.3
- 10 Nov 2026
- Last 30 days
- 11
- releases
- Usual gap
- ~1d
- between releases
- Tracked
- 92
- since 26 May 2026
- Homepage
- kimi.com/code/docs
Releases per month
last 12 months- Dec
- Jan
- Feb
- Mar
- Apr
- 5 May
- 26 Jun
- 27 Jul
- 12 Aug
- 11 Sep
- 7 Oct
- 4 Nov
Release history
92 releasesReleases 77
-
Steering now interrupts active model requests mid-turn
- Steering now interrupts an in-progress model request and continues in the same turn
- New POST /api/v1/sessions/{sessionid}/steer endpoint steers a session with a single request
- Web UI interaction improvements and bug fixes
- Fixed stdio MCP servers launched with script commands like npx failing to start on Windows
-
Sessions can run in SSH or Docker environments and be forked
- New sessions can run in SSH, Docker or custom-command environments using --environment and --environment-cwd
- Fork a running session, or fork from any earlier turn
- Use /model and /effort while the agent is working; the change applies from the next turn
- Projects with multiple directories now merge skills from all of them, and the primary directory wins on name conflicts
- Breaking server API changes: several REST task and status endpoints are removed in favour of WebSocket frames, and global search indexing is now off by default
-
File reads now capped at 100,000 characters by default
- A single file read now returns at most 100,000 characters or about 25,000 tokens, down from 500,000 characters
- Set maxchars under [read] in config.toml to raise the file read limits again
- Built-in tool and subagent prompts are shorter, as are repeated goal and plan mode reminders
- Web: improved interactions and fixes for known bugs
-
Import Claude Code and Codex sessions, and Plan mode now needs your approval
- New /import-sessions and kimi import-sessions bring in Claude Code and Codex sessions from the last 15 or 30 days
- Plan mode starts only when you turn it on, and every plan now waits for your review, even with Never Ask or kimi -p
- Tower raises the default agent concurrency limit from 16 to 256 and adds per-spawn model selection
- Ctrl-O expands thinking while it streams, and /provider refresh re-fetches provider model lists
- Workspace MCP servers reconnect automatically, and server approval and question APIs are replaced by a single interactions endpoint
-
Swarm mode now warns about high usage when turned on
- Turning on swarm mode with /swarm now shows a high-usage warning
- AgentSwarm calls rejected before any subagent starts no longer show as a failed swarm, and the agent points you to /swarm if you asked for one
- Updated model switch confirmation prompts
- Swarm and goal start confirmations now give the permission-mode notice in one sentence instead of two repeated lines
- Web interface interaction improvements and bug fixes
-
Agent no longer launches subagents or swarms on its own
- Tasks are handled directly by default instead of proactively launching subagents
- To delegate, ask in your message, AGENTS.md, a skill, or an enabled multi-agent mode
- Agent swarms now start only when swarm mode is turned on; the agent can no longer launch AgentSwarm itself
- Swarms are no longer available in kimi -p or ACP clients such as Zed
-
Session lock fixes and takeover of stuck sessions
- After a retry, Kimi Code offers to take over a session that is still in use, so sessions left locked by an exited process can be used again
- New KIMICODESESSIONLOCK=0 switch turns off the session lock
- Session-in-use errors now show the name of the client holding the session
- Fixed sessions occasionally staying occupied after an application crash on macOS and Linux
- Fixed workspace trust failures when Git Bash is unavailable on Windows
-
Model-switch prompt, plugin skill renames and auto session titles
- Confirmation dialog when switching models mid-session: new session, compact and continue, continue, or cancel (only when enabled by cloud config)
- Plugin skills now use qualified pluginId:skillName names; invoking by bare name no longer works
- Session title is generated automatically after the first turn and shown as the terminal window title
- Remote control transfers raised from 10 MiB to 100 MiB via streaming
- dangerouscommandguard = false now turns off all built-in confirmation guards; /compact queues while the agent is busy
-
Opt-in progress updates, kimi app command and Tower mode changes
- New opt-in NotifyUser feature shows the agent's mid-turn progress with an optional one-line title, enabled in /experiments or settings
- New kimi app [path] command opens a new desktop chat in the current or given directory
- Background-task waits in a turn are capped at 90 seconds, and a new message ends the wait immediately
- Tower mode: a message sent mid-run aborts the current turn and runs on its own; orchestration is now deterministic
- Fullscreen mode gets sticky user message summaries you can click to jump back, plus Ctrl-S for skill prompts
-
Filesystem watchers back on by default; 2.1.0 defensive changes rolled back
- Filesystem watchers for config and workspace files are on by default again
- Turn them off with [watch] enabled = false or KIMICODEWATCH=0
- Rolls back some of the overly defensive changes made in 2.1.0
-
Experimental fullscreen TUI mode with clickable blocks, plus symlink hardening
- New TUI mode setting in /settings with an experimental fullscreen layout; takes effect after restart, or set tuimode in tui.toml
- In fullscreen, click a folded block to open or close it, and use a clickable Jump to bottom indicator
- File tools can no longer reach files outside the working directory through symlinks
- Project-local configuration is applied only after the workspace is trusted
- Faster startup and lower memory use; fixes for large sessions, forked session titles, MCP OAuth re-authorization and Remote Control login
-
Faster session loading, API key env vars and approval fixes
- Sessions open, resume and restore faster, including long sessions and workspaces with many sessions
- Providers can read their API key from a named environment variable via apikeyenv
- Bash commands that cannot be statically analyzed no longer prompt for approval in Ask When Needed mode
- Fixed kimi -p exiting early and cancelling the active turn when a cron task fires
- kimi install-app is renamed kimi install-desktop; the old name still works as a hidden alias
-
Desktop app launcher and mermaid diagrams in the terminal
- New /desktop command (alias /install-desktop) and kimi install-app open the Kimi Code desktop app page in the browser
- Mermaid code blocks now render as diagrams in the terminal; turn off in /settings or with mermaid = "off" in tui.toml
- Diff code blocks are now syntax highlighted
- Steering a running turn no longer duplicates messages, keeps the original prompt visible, and fixes undo and attachment issues
- Images are uploaded as file references, and the oldest media are dropped with a warning when requests get too large
-
Delete sessions from the picker; web session titles always on
- Delete sessions from the session picker: press Ctrl+X on a session, then y to confirm
- Web: AI session titles are always on, generated after the first turn and regenerable from the rename field
- Goal time budgets no longer count time the session was closed, and the 24-hour limit is removed
- New kimi upgrade -y flag skips the confirmation prompt; rm -rf on only /tmp paths no longer asks for confirmation
- Fixes for a crash on retried LLM requests, Remote Control uploads over 3.5MB, and web UI stuttering and loading hangs
-
Read-only tools for /btw, experimental Updates panel, web media rail
- The /btw side agent now has read-only tools
- Experimental Updates panel shows paginated progress messages from the main agent and subagents (set KIMICODEEXPERIMENTALNOTIFYUSER=1)
- Subagent model pool, Remote Control and the session-index database are now always on
- Web: delete sessions permanently from the context menu, and preview images and videos in a reorderable media rail
- Finished tool calls collapse to a header and one outcome row; Ctrl+O reveals hidden output
-
Context budget reminders and an experimental Tower mode in the web UI
- Model is reminded of its context budget before automatic compaction and pointed to the session event log afterwards
- Experimental Tower multi-agent collaboration mode in the web UI via /tower, with optional base branch
- Turn-level file history is now always on; the experimental flag is removed
- Auto permission mode no longer blocks dangerous or unanalyzable commands, and a warning shows after switching to Ask When Needed or Never Ask
- Web UI adds selection comments and quote-to-chat, and renames permission modes to Always Ask, Ask When Needed and Never Ask
-
Subagent model pool on by default, dangerous-command guard, kimi session list
- The subagent model pool ([secondarymodel]) is now on by default in every launch mode; opt out with KIMICODEEXPERIMENTALSECONDARYMODEL=0
- Dangerous shell commands such as shutdown, reboot or rm -rf are blocked in Auto mode and always ask first in Manual and YOLO modes; the guard can be turned off
- New kimi session list command lists sessions from the command line
- Web: new Plugins panel in Settings to browse the marketplace and install, enable, disable or remove plugins
- Session forks are much faster, config.toml keeps its comments and formatting when updated, and the Bash tool's cwd is no longer limited to the workspace
-
Experimental Remote Control, tower mode and subagent fork
- Remote Control (experimental) lets you reach a local web session remotely via kimi rc, kimi web --remote-control or /remote-control
- Experimental tower mode for multi-agent orchestration, started with /tower on and /tower <objective>
- Optional fork parameter starts a subagent with a snapshot of the caller's conversation history (experimental)
- Web UI: multi-tab right sidebar, Move to background button on running Bash commands and subagents, better mobile bottom sheets
- New [swarm] timeoutms setting for AgentSwarm subagents; many fixes for session resume, ACP and Windows paths
-
WaitFor tool and kimi.ai / kimi.com OAuth login options
- Two OAuth login methods are supported: kimi.ai and kimi.com
- New WaitFor tool lets the agent wait for a background task within the current turn instead of ending the turn
- Edit and Write now require reading an existing file first, and reject writes if the file changed on disk since it was read
- Sub-agents no longer spawn their own sub-agents by default; custom agent profiles can still allow it
- Official Kimi Datasource plugin adds 13 data sources, including WHO, FAO, OECD, FRED, Eurostat and Chinese government data
-
Web UI: subagent panel stays expanded, new Lab tab with sidebar toggle
- Settings gains a Lab tab with an optional multi-tab sidebar toggle, off by default
- When enabled, the sidebar shows Open, Done and Workspaces tabs
- Subagent detail panel now keeps the working process fully expanded
- Subagent detail panel no longer shows the end-of-turn timestamp footer
-
Use several skills in one prompt; Windows binary auto-updates
- Activate multiple skills in a single prompt by typing / after whitespace to insert a skill token
- The Windows native single-binary CLI now supports automatic updates
- Web UI sidebar gains Open, Done and Workspaces tabs, and sessions can be marked done or reopened
- New web session management page to filter by workspace, status and time, with batch mark done or reopen
- Web UI: mentions render as icon pills, Subagent panel renamed Background Agent, new kimi web --web-title flag
-
Web UI overhaul: plan/goal pills, new panels, fuzzy slash search
- Web composer: plan and goal modes are now removable pills armed via /plan or /goal, and Swarm moves to a toolbar toggle
- New plan viewer panel, reworked subagent card grid with status filters, and a background Bash panel with status filtering and output view
- Slash commands support fuzzy search by description, pinyin or pinyin initials; slash and @ menus are restyled with scroll fade
- Experimental automatic session titles in the web UI; the 64 MiB limit on web session exports is removed
- Fixes for tool call ids on self-hosted OpenAI-compatible endpoints, CJK text absorbed into bare URLs, MCP OAuth cancellation and Windows file-watcher errors
-
Experimental fullscreen TUI, subagent model pool and LaTeX rendering
- Experimental fullscreen TUI mode, enabled with KIMICODETUIFULLSCREEN=1
- Experimental secondary-model pool: choose default or named models the main agent picks from when spawning subagents
- LaTeX math formulas in messages now render as Unicode
- Workspace trust prompt shows project MCP launch targets and defaults to declining trust
- Fixes for Ctrl+C during API retries, 400 errors after interrupted thinking, and plugin Markdown files misread as skills
-
Live progress for background subagents in /tasks, plus web previews
- The /tasks panel now shows live work progress of background subagents
- Web: image and video tool results open in a fullscreen preview, with zoom for images
- MCP servers that require OAuth are detected without setting auth: "oauth" in the config
- Model request retries now show attempt count and provider error detail in the loading indicator
- Windows security fixes for binary planting before the workspace trust prompt, and a clear error when Git for Windows is missing
-
Cache-expiry reminders, Windows Computer Use and web UI fixes
- Shows a cache-expiry reminder when you resume a long-idle session or submit after a long idle stretch
- Kimi Computer Use now works on Windows x64, installed from /plugins, with the underlying error shown if setup fails
- Web: a persistent failure card with one-click resume stays in the session when a model request fails
- Web: new flat view for the sidebar session list, and subagent tasks show their model and thinking level
- kimi -p now waits for background tasks and subagents to finish instead of exiting after the main turn
-
Computer Use and WebBridge plugins; v2 engine now default
- Kimi Computer Use and Kimi WebBridge are now built-in official marketplace entries, installable from /plugins with retry for interrupted setup
- CLI surfaces (TUI, kimi -p, acp, export, provider) now run on the agent-core-v2 engine by default; KIMICODELEGACYFLAG=1 restores the legacy engine
- /fork no longer switches to the forked session; the current session and its background tasks keep running, and the fork appears in /sessions
- The CLI now asks whether to trust the current folder on startup
- MCP tool results now pass structuredContent and meta to the model, and MCP OAuth re-authorization no longer fails
-
New hook events for turns, prompts, tasks and session heartbeats
- Adds TurnStarted, UserPromptQueued, TaskStarted and SessionHeartbeat hook events, configured under [[hooks]] in config.toml
- Hook payloads now include the session title and client type; SessionStart also includes the model and profile
- New [tokencounting] config section picks how context token counts are derived: measured+estimated, measured or estimated
- Renamed loopcontrol keys to maxattemptsperstep and maxstepsperturn; old keys no longer take effect on the v2 engine
- Fixes answering interactive question prompts when a provider returns tool call IDs containing colons
-
Fixes for kimi web startup errors and Esc interrupts, plus TUI redraws
- Fixed sporadic "model is not configured" errors when starting kimi web
- Pressing Esc to interrupt a turn now keeps the assistant's partial output, and the model is told it was interrupted
- TUI no longer redraws the full screen as often
- Web: @ file mentions now work in a new-session draft, and code blocks get Monaco-based highlighting
- Web: fixed code block fonts and line numbers, permission mode order and colours, and a wrong thinking level on new sessions
-
Plugins can add custom agents and system-prompt instructions
- Plugins can ship custom agents in an agents/ directory or via manifest paths, found automatically and usable for sub-agent delegation
- Enabled plugins can add agent system-prompt instructions with systemPrompt or systemPromptPath in kimi.plugin.json
- Custom agents can now be defined in Markdown on agent-core
- New /secondarymodel command sets the secondary model used by subagents
- TaskOutput no longer waits on background tasks and returns an immediate snapshot; also fixed request headers and sessions missing from the picker
-
Customizable footer status line, plus quota fail-fast
- Add a customizable footer status line, configured via [statusline] in tui.toml
- Fail fast when account quota or balance is exhausted instead of silently retrying for about 3 minutes
- Stop the turn after repeated invalid tool calls instead of retrying indefinitely
- Show a quota note after installing official plugins that bill against plan quota, and a notice when a used official plugin has an update
- Remove the 50 MB limit on file uploads to the built-in server; fix garbled line numbers in web code blocks
-
Goal mode fixes for step limits and mid-goal messages, plus /undo fix
- Goals no longer pause when a turn hits the per-turn step limit; work continues in more turns instead
- Messages sent while a goal is running are steered into the active goal turn instead of being rejected
- /undo now restores conversation history, todo lists, plan mode and task notifications consistently
- Web: copying selected chat text over plain HTTP no longer replaces the clipboard with an event placeholder
- Hosts can defer registered user-tool schemas until needed by setting disclosure to deferred
-
Default MCP timeouts, subagent model overrides, thinking fix
- Set global default MCP server timeouts in config.toml or with environment variables
- Experimental secondary-model bindings for new subagents, with per-agent model preferences and subagent-only overrides
- Fixes lost thinking content on OpenAI-compatible endpoints that name the reasoning field differently, such as newer vLLM
- Web search and web fetch services can be configured with environment variables, no OAuth login needed
- Removed the toolbar tip suggesting the superpowers plugin
-
Pick thinking effort levels in ACP clients; custom Markdown agents
- ACP clients such as Zed can now choose a thinking effort level (off, low, medium, high) instead of only on or off
- Custom agents can be defined as Markdown files with frontmatter and used as the main agent or a sub-agent (v2 engine only)
- Custom agents can be limited to certain sub-agent types, and global tool gating restricts which tools agents may use (v2 engine only)
- Videos attached to a prompt now reach the model with the prompt and stay playable in the chat after a reload
- Many more providers import from the models.dev catalog, and thinking levels now follow each model's real capabilities
-
Web servers run in the foreground only; ACP works without login
- Web servers now run foreground-only: /web always starts a new server and Ctrl+C stops it
- kimi web kill and kimi web ps subcommands are removed
- kimi server kill stays as a deprecated fallback for servers started before 0.28.0
- ACP sessions can start with configured non-OAuth model credentials, no terminal login needed
- Fixed running subagents not picking up permission mode switches made after they started
-
New kimi web command replaces kimi server, with multi-server support
- kimi web replaces the kimi server command tree and runs the server in the foreground; the background daemon and OS-service commands are removed
- Several servers can share one home directory, each taking the next free port
- Manage instances with kimi web kill, kimi web ps and kimi web rotate-token; kimi server commands now print a deprecation notice and exit 1
- Thinking effort only persists for levels below the model's top tier (max)
- YOLO and Auto permission mode descriptions corrected: YOLO may still ask questions, Auto never asks
-
New /copy command and automatic model list refresh with API key
- New /copy slash command copies the last assistant message to the clipboard
- Using an API key for Kimi coding models now also fetches the latest model list automatically
- Fixed Esc and Ctrl+C cancelling compaction instead of closing an open /btw panel
- Fixed repeated request rejections after an interrupted model response
- Fixed the URL fetch tool letting crafted domains and redirects reach loopback or internal network services
-
Coder subagents get the main agent's full tool set
- Coder subagents can now use background tasks, todo lists, plan mode, skill invocation and nested agents, like the main agent
- A subagent run waits for its background tasks to settle before reporting completion
- Model and effort pickers now warn that switching invalidates the prompt cache and suggest /new to avoid extra token costs
- Explicit thinking off is now honored on OpenAI-compatible providers instead of being treated as never configured
- Fixes for wrong reasoning effort options on Kimi models via Anthropic protocol, context size under-reporting and resumed sessions jumping up the list
-
Web chat accepts any file type, plus security fixes for the web server
- Web chat can now attach any file type; files the model can't read inline are uploaded and passed as a file path it can read on demand
- Fixed a bearer-token bypass using percent-encoded API paths that allowed unauthenticated access to every API route
- Fixed the session filesystem API following symlinks outside the workspace, which could expose host files
- Anthropic-compatible models now honor adaptivethinking = false and apply official effort profiles, with a 128k output fallback for unknown models
- Fixed the CLI exiting when clipboard image reading fails (falls back to pasting text); web UI shows detailed model request errors and all attachments as chips
-
Print mode stays alive for background tasks; goal and web fixes
- kimi -p now stays alive while background tasks run and feeds each completion back as a new turn; printbackgroundmode = exit restores the old behaviour
- In print mode, background Bash tasks and subagents have no timeout by default; interactive defaults are unchanged
- kimi -p "/goal ..." stays alive until the goal finishes; printbackgroundmode and printmaxturns now work on the experimental engine
- Subagent timeout is a fixed 2 hours by default, set via [subagent] timeoutms; LLM retries per step rise from 3 to 10
- New built-in check-kimi-code-docs skill answers product questions from the official docs, plus goal-budget and web UI fixes
-
Slow Bash commands move to background instead of being killed
- Foreground Bash commands that hit their timeout keep running in the background and report back; set bashautobackgroundontimeout = false to kill instead
- Local server (kimi server run / kimi web) now runs on the agent-core-v2 engine by default, no opt-in flag needed
- New session diagnostic export: /export or Export session in the web menu downloads a ZIP of the session and troubleshooting logs
- Session fork now carries over media attachments, plan files, background task output and cron tasks
- Optional progressive tool disclosure for MCP tools, enabled with KIMICODEEXPERIMENTALTOOLSELECT=1
-
Print mode stays alive for background tasks, goals and cron; subagent timeout now 2 hours
- kimi -p can stay alive across background-task completions to steer follow-up turns, via [background].printbackgroundmode = "steer"
- kimi -p runs now stay alive while a goal is active or a cron task is pending, so continuations and cron fires complete
- New subagent.timeoutms setting (or KIMISUBAGENTTIMEOUTMS); default subagent timeout raised from 30 minutes to 2 hours
- Dismissing a question prompt now counts as choosing not to answer, rather than picking the recommended option
- Web UI: workspaces can be added by typing a path, the access token lasts 7 days, plus fixes for stuck sending state, login bounce and scrolling
-
Extra Usage balance in /usage and /status, plus safer image handling
- /usage and /status now show your Extra Usage (fuel pack) balance
- Image-heavy sessions stay within request-size limits: per-image 256 KB budget and 2000px cap, configurable under [image] in config.toml
- A request-too-large (HTTP 413) error now recovers automatically by retrying with older media replaced by text markers
- Fixed resuming sessions whose original working directory no longer exists
- Fixed prompt-mode goals so they run to completion and report invalid goal commands up front
-
Vercel plugin added to marketplace; goal-mode and exit-code fixes
- Vercel plugin is now in the bundled plugin marketplace: run /plugins and select Vercel Plugin to install it
- Fixed kimi -p runs exiting with code 0 when a turn fails
- Autonomous goals can no longer be paused by model-reported status updates, and the turn that starts a goal now counts toward its usage
- Image downscale cap raised from 2000px to 3000px, and width/height fixed for EXIF-rotated portrait photos
- Web UI: Enter confirms dialogs, the composer model switcher also sets the default model, and session skills show as /skill:<name>
-
Goal-mode fixes and thinking kept across turns on Anthropic
- Anthropic provider now keeps prior turns' thinking by default; turn off with [thinking] keep = "off" or KIMIMODELTHINKINGKEEP=off
- Goal mode fixes: one final outcome summary, failed starts restore permission mode, queued goals wait behind new messages
- Goal token budgets now count completion tokens and stop cleanly when exhausted
- kimi -p no longer abandons late-starting or long-running background subagents
- Web: chat streaming recovers after a stale tab connection; /goal, /btw and slash skills work on the new-session screen
-
Preserved Thinking on by default, plus archived session restore
- Preserved Thinking is now on by default for Kimi models when Thinking is on, keeping earlier reasoning across turns; set [thinking] keep = "off" to disable
- Web: new Archived sessions page in Settings to browse and restore archived sessions, backed by new server APIs
- Experimental progressive tool disclosure (selecttools) loads MCP tools on demand; off by default and inert without model support
- Web: swarm footer replaced by one inline tool card showing live subagent progress and the aggregated result
- Fixes for Gemini tool calling, missing sessions in the list, @ file completion in large projects, and many web UI glitches
-
Web chat image zoom, thinking-level picker and server flags
- Web chat: click an uploaded image to enlarge it, and uploaded videos now play correctly
- Web model picker gets a segmented thinking-level control for models with multiple reasoning efforts
- kimi server run adds --dangerous-bypass-auth (no token on trusted networks) and --keep-alive (stay up past the idle timeout)
- kimi -p now waits for background subagents to finish and responds to their results before exiting
- Recent TUI transcript rendering changes reverted to the original upstream behavior while rendering issues are reworked
-
Fixes for interrupted sessions, duplicate tool ids and Windows upgrades
- Sessions no longer drop later messages after a turn is interrupted between a tool call and its result
- Requests no longer get rejected by strict providers when the model emits duplicate tool call ids
- kimi upgrade no longer fails on Windows, and background updates no longer flash a console window
- Shell commands now find user-installed tools like Homebrew's gh, as PATH is enriched from the login shell at startup
- Context-compaction notes now capture a forward plan, so the agent continues more coherently after auto-compaction
-
Large images keep original detail; TUI and web UI fixes
- Compressed large images now tell the model the original and delivered details, and the original stays available for cropped or full-resolution reads
- Fixes TUI bugs that blanked the screen or hid the input box, and a crash when resizing narrow with CJK or emoji text
- Shell commands are saved to input history; press Up on an empty ! prompt to browse them in bash mode
- Compaction handoff summaries now keep the latest intent, key tool results, decisions, open questions and items to re-check
- Removes the experimental micro compaction feature; fixes a sluggish web UI after opening many sessions and refreshes its icons
-
Automatic image compression and a redesigned web UI
- Oversized images are downsampled and re-encoded before reaching the model, cutting vision-token cost and avoiding provider image-size errors
- Web UI refreshed with a new design system, grouped tool calls, a Cmd/Ctrl+K session search palette and inline queued prompts
- Model alias overrides under [models."<alias>".overrides] keep manual thinking effort levels and metadata through provider catalog refreshes
- Each web session now keeps its own plan, swarm and goal mode toggles, and the Thinking-by-default setting now takes effect
- Fixes for transcript jumping while scrolling, terminal left in raw mode after a crash, and memory growth in long sessions
-
Plugin slash commands and reworked conversation compaction
- Plugins can provide slash commands via a commands field in their manifest, invoked as plugin:command with $ARGUMENTS expansion
- Compaction reworked: keeps recent user prompts plus one summary, and fixes HTTP 400 errors on strict providers like Gemini, Vertex and Anthropic
- Micro-compaction now defaults to off
- Press Esc twice while idle to open the undo selector
- Web chat renders Mermaid diagrams; headless runs (kimi -p) now force-exit instead of hanging
-
Background model list refresh, ripgrep-based Glob, and fixes
- Provider model lists now refresh automatically in the background, so new models appear without restarting
- Glob now uses ripgrep: respects .gitignore, supports brace patterns, returns only files, keeps partial results with a warning
- Fixed provider error messages showing as blank lines in the TUI when the server returns an HTML error page
- Fixed the web composer hiding behind the mobile Safari toolbar and the page auto-zooming on focus
- Malformed tool call arguments now fall back to schema validation handling
-
Anthropic-compatible protocol support, custom LLM headers, web UI polish
- Managed Kimi Code now supports the Anthropic-compatible protocol, including video input
- New KIMICODECUSTOMHEADERS setting adds custom headers to outbound LLM requests
- Recovers from provider 413 context overflows by compacting before retrying
- Compaction output is capped at 128k tokens by default, and the configured max output size is now respected
- Web UI: sessions are created on first message, with completion sound, question notifications and a step-by-step question wizard
-
Plugin hooks, faster long sessions and a new kimi update command
- Plugins can declare hooks in their manifest to run scripts on lifecycle events
- Long conversations stay responsive: rendered lines are cached and only recent turns are kept in the transcript
- kimi update works as an alias for kimi upgrade
- Completion tokens are capped to the remaining context window for chat-completions providers, avoiding overflow errors
- Web chat input grows with its content and has an expandable editor; the /web command now signs in automatically
-
Shell mode with ! commands and a tabbed plugin panel
- Shell mode: type ! to run shell commands whose output the AI can see; Ctrl+B moves long-running ones to the background
- Redesigned /plugins panel with Installed, Official, Third-party and Custom tabs, plus update badges on installed plugins
- Confirmation prompt before installing third-party plugins
- Local server now requires a bearer token, and --host exposure adds TLS, password and rate-limit hardening
- Web UI shows line-by-line diffs for file edits and the plan body when exiting plan mode
-
Shortcuts for continue, todo lists and model switching, plus web fixes
- Add -c as a shorthand for --continue
- Alt+S in the model picker switches the model for the current session only, without saving it as the default
- Ctrl+T expands or collapses a truncated todo list; Ctrl+O expands a running Bash command before its result arrives
- Tab-completing a slash command now shows subcommand suggestions, and Tab no longer opens the file completion list
- Fix web yolo mode auto-approving plan reviews and sensitive file access
-
Add extra workspace directories with /add-dir and --add-dir
- New /add-dir command and kimi --add-dir flag add extra working directories to a session, and can be remembered per project
- Project-level local config now lives in .kimi-code/local.toml; adding it to .gitignore is recommended
- Ctrl+B moves long-running foreground commands and subagents into background tasks, viewable in the /tasks panel
- Faster session snapshot loading, and provider safety-policy blocks are now surfaced instead of treated as completed turns
- Fixes for empty text blocks breaking requests, mismatched image extensions, and empty console windows on Windows
-
Web session search, plus a cap on AgentSwarm concurrency
- Instant session search in the web sidebar, filtering by title and the last user prompt
- New environment variable caps AgentSwarm concurrency during the initial ramp, helping avoid provider rate limits
- Redesigned web login dialog leads with a single Authorize in browser button, with manual code entry as a fallback
- Web chat lazy-loads older messages on scroll up, and the app now loads all sessions instead of only the 20 most recent
- Slash menu fixes: highlighted command stays visible, long names wrap, and skill selection no longer sends immediately
-
Server-hosted web UI with new kimi server and kimi web commands
- New kimi server command to start, stop and manage a local server
- New kimi web command opens the server-hosted web UI in a browser
- Web chat gets a session list, auto-scroll and server REST and WebSocket APIs
- OAuth token refresh failures now show the underlying connection error instead of a login prompt
- Resumed sessions no longer reuse turn ids already in history; web login page and no-workspace startup fixed
-
New kimi vis command opens the session visualizer in your browser
- New built-in kimi vis command launches the session visualizer in your browser, pointed at your local sessions
- kimi vis supports --port and --host, --no-open, and kimi vis <sessionId> deep-links to a single session
- Anthropic-compatible providers no longer read ambient Anthropic shell credentials or custom headers
- Banners can now be shown at configurable frequencies, with display state kept locally
- Fixes repeated compaction when context stays over the blocking threshold; the /btw side panel is now at most one-third of the terminal
-
Browse sessions across all folders; legacy SSE MCP servers supported
- New all-sessions picker with name search, paginated browsing and copyable resume commands for sessions in other working directories
- Legacy SSE MCP servers now work alongside stdio and streamable HTTP
- Resumed sessions recover when an interrupted tool call result was never recorded
- Loaded skills now include their directory so the agent can find bundled scripts and templates
- Pressing Ctrl-C clears draft editor text before cancelling an active stream
-
Mode flags now work when resuming sessions, plus streaming Bash output
- --auto, --yolo and --plan can now be combined with --session or --continue, applying the mode to the resumed session
- Foreground Bash stdout and stderr now stream while commands are still running
- Sub-skills get their parent's prefix in their names and show up as dotted slash commands in the TUI
- Fixed endless desktop notifications in iTerm2 by only sending progress sequences to terminals that support them
- Invalid config.toml sections are dropped with a warning instead of stopping startup, and resuming a plan-mode session no longer fails
-
Swarm tasks get a YOLO option, nicer autocomplete, and shell fixes
- Starting swarm tasks from Manual mode now offers a YOLO choice, and AgentSwarm tool calls must run alone in a model response
- Autocomplete wraps long descriptions onto a second line and finds slash commands by alias, so typing /clear suggests new (clear)
- Background tasks now stop by default when a session closes, and active turns are cancelled on shutdown so shell commands don't outlive prompt-mode exits
- A tips banner now appears below the welcome panel on startup
- Fixes for premature stream close errors on shell timeouts, Windows ACP file reads and edits, and unsupported audio/video being silently dropped
-
New Interrupt hook event and image output fix for OpenAI-compatible APIs
- New Interrupt hook event fires when the user interrupts a turn, such as pressing Esc, so hooks can see the turn stop instead of staying on a working state
- Image outputs from tools are now preserved when using OpenAI-compatible chat completions
-
Claude Fable 5 support and an interactive undo selector
- Adds Claude Fable 5 support to the Anthropic provider
- New interactive undo selector, with clearer messages when the undo limit is reached
- YOLO mode no longer asks before writing or editing files outside the working directory
- Kimi Datasource now uses the matching OAuth credentials and service endpoint for the active environment
- Fixes goal marker text overflowing the terminal width, and forking sessions is blocked during active turns
-
Custom color themes and import from Claude Code and Codex
- Custom color themes: define a palette as a JSON file in ~/.kimi-code/themes/ or generate one with the /custom-theme skill command
- New /import-from-cc-codex command imports selected Claude Code and Codex instructions, Skills and MCP settings
- Plugin marketplace now shows available updates with an update badge, and Enter updates the plugin in place
- Grouped subagent progress shows active status breakdowns and elapsed time
- Fixes for device login URL visibility, Windows builds and long queued message display
-
Old experimental config no longer blocks startup; xhigh effort passes through
- Obsolete experimental config entries can stay in the config file without blocking startup
- The xhigh reasoning effort is now passed through for OpenAI-compatible chat completions requests
-
Agent swarms via /swarm, plus goals and micro compaction on by default
- New /swarm command runs agent swarms with live progress and rate-limit-aware retries
- Goals, background questions and sub-skill discovery no longer need experimental opt-ins
- Micro compaction is on by default, with its opt-out flag kept
- Standard HTTPPROXY, HTTPSPROXY, ALLPROXY and NOPROXY variables are honored, including SOCKS proxies
- Homebrew installs now update with brew upgrade kimi-code instead of npm
-
Built-in skills as slash commands, new env vars, 30-minute subagent timeout
- Built-in skills now appear as direct slash commands, listed ahead of external skill commands
- Subagents get a fixed 30-minute timeout, with short resume instructions when they time out
- Env vars migrated from kimi-cli: sampling temperature and top-p, preserved thinking, and KIMICODENOAUTOUPDATE to disable updates
- Experimental sub-skill discovery behind KIMICODEEXPERIMENTALSUBSKILL, with review and consolidate skills
- Goal queue fixes, and YOLO mode now asks before starting a goal so you can switch to Auto
-
Queue goals for the agent, reload config live, and a config doctor
- Queue several goals with /goal next <objective>; the agent picks up the next one when the current goal is done, and /goal next manage edits the queue
- New /reload applies updated config files to the current session, and /reload-tui reloads only TUI preferences
- New doctor command validates Kimi Code configuration files
- Built-in update-config skill lets Kimi edit its own config files
- Experimental feature toggles are now persistent, with a TUI panel that reloads the session to apply changes
-
IDE support via new ACP command, plus /btw side conversations
- New kimi acp subcommand speaks Agent Client Protocol over stdio so IDEs like Zed and JetBrains AI Chat can drive sessions directly
- New /btw command opens a side-channel conversation without steering the active main turn
- /btw can open its side-channel panel before you type a question
- Fixes the Ctrl+G external editor on Windows and goal budget tool schemas for OpenAI-compatible providers
- Enabled experimental flags are now logged at startup, and TUI dialogs and selectors share a consistent look
-
Experimental /goal mode, /undo, and background auto-upgrades
- Experimental goal mode: /goal keeps Kimi working on one task across turns, enabled with KIMICODEEXPERIMENTALGOALCOMMAND=1
- New /undo command withdraws the last prompt from conversation history
- Automatic upgrades run in the background (can be disabled in tui.toml), plus a new kimi upgrade command
- New kimi provider CLI subcommand to add, remove and list providers from a custom registry or models.dev
- Background structured questions let agents keep working while waiting for answers; new approval lifecycle hook events
-
New /provider command, tabbed model selector and cron reminders
- New /provider command for managing AI providers, with support for custom registry imports
- Model selector is now tabbed
- Scheduled reminders render distinctly in the TUI, and cron fired events are exposed to SDK clients
- KIMIMODELADAPTIVETHINKING forces adaptive thinking on or off for custom-named Anthropic-backed endpoints
- Fixes for glob backslash escaping, truncated compaction summaries and Kimi Platform API key login labels
-
Configure models via env vars, install plugins from GitHub URLs
- Run against a specific model with KIMIMODEL environment variables, no config.toml edits needed
- Install plugins directly from GitHub repository URLs, with origin and trust level shown in the plugin manager
- Default 1000-step per-turn limit removed; maxstepsperturn in config still sets a custom limit
- Sessions can be queried by sessionId or workDir, and resuming from another directory shows a cd command
- Background agents now show their real status (lost, failed, killed) and failed streams or context overflows are retried or recovered
-
Auto permission mode and scheduled tasks and reminders
- New /auto slash command and --auto CLI flag turn on auto permission mode
- Scheduled tasks: ask for a reminder, a recurring job on a 5-field cron schedule, or a self-timed follow-up
- Write and Edit approval prompts show file content and diff, with a full-screen viewer on ctrl+e
- Expanding a Bash tool card with ctrl+o now shows the full command; the todo panel caps at five rows with a +N more indicator
- Fixes for compaction edge cases, datasource tool responses and migration of the legacy defaultyolo setting
-
Plugin management, session export commands and looser read permissions
- Plugins can be installed user-globally and managed interactively, and can now provide skills and their own MCP servers
- Reads outside the working directory no longer prompt; session approvals match the exact call and path rules ignore case
- New /export-md and /export-debug-zip commands export the current session as Markdown or a debug ZIP
- Pasting again on a folded paste marker expands it back to the original content
- Fixes a TUI crash when pull request lookup fails at startup and a thinking spinner that lingered after a turn ended
-
/logout lets you pick which provider to log out of
- /logout now opens a picker to choose the provider, with the current one highlighted; also available as /disconnect
- OpenAI-compatible reasoning models (DeepSeek, Qwen, One API and other gateways) work without a hand-set reasoningkey
- Fixed user skills in ~/.agents/ not being loaded
- Fixed Plan mode so new sessions no longer fail after a plan review is rejected
- TUI exits promptly when the terminal goes away, avoiding leftover processes pinning a CPU core
-
New /connect command to set up any provider and model
- Add a /connect command that configures a provider and model from a model catalog
- Provider and model pickers support type-to-search and pagination; /model is paginated with many models
- Ctrl-J now inserts a new line in the prompt
- Model choices from the terminal UI are saved as the default, and the configured default thinking state applies to new sessions
- User skills migrate from ~/.kimi/skills/ to ~/.kimi-code/skills/ on first launch
Smaller updates 15
-
Fixes MCP tool declaration errors and Web UI message queuing
-
Fixes repeated HTTP 400 session errors and web UI glitches
-
Fixes for session resume, compaction and duplicate messages
-
Fixes for subagent memory, slow rendering and Ctrl+C exiting
-
Fix when the kimi-cli migration prompt is shown
-
Web app fixes: per-session permission mode, composer input and attachments
-
Fixes for pasted images and videos not reaching the model
-
Fixes for thinking effort, built-in tools and stuck sessions
-
Fixes for unsupported images, API retries and web notifications
-
Fixes misleading "OAuth login expired" error for unavailable models
-
Spinner stays visible while encrypted reasoning streams
-
Fixes Zed/ACP thread startup and web sidebar unread dots
-
Web login and settings fixes, plus background server fixes
-
Model picker now refreshes provider model metadata first
-
Fixes a crash when starting a goal in the TUI