Kilocode CLI
7 days since its last release · 7.8.27 ·
Kilocode is an AI coding tool from Kilo Org. Its latest release is 7.8.27, released on 5 Nov 2026. It shipped 16 releases in the last 30 days, and we have tracked 261 since 13 Oct 2025.
- Latest
- 7.8.27
- 5 Nov 2026
- Last 30 days
- 16
- releases
- Usual gap
- ~2d
- between releases
- Tracked
- 261
- since 13 Oct 2025
- Homepage
- kilo.ai
Releases per month
last 12 months- 14 Dec
- 20 Jan
- 26 Feb
- 31 Mar
- 36 Apr
- 30 May
- 24 Jun
- 14 Jul
- 10 Aug
- 23 Sep
- 18 Oct
- 2 Nov
Release history
261 releasesReleases 69
-
Hand off a local worktree to Cloud from Agent Manager
- Agent Manager can continue a local worktree in Cloud, copy its open conversations, and open the Cloud workspace in a browser or copy its link
- New OpenShell-ready Docker image keeps Kilo's own sandbox off and skips the network calls that OpenShell blocks
- Images from image models such as Google Nano Banana and from the image generation tool now show inline in the chat
- Kilo can be installed through the Nix overlay again, and the existing opencode attribute still works
- Many dependency updates fix critical, high, moderate and low severity security advisories
-
Local API can now manage hosted Cloud workspaces and sessions
- Local session-host API can list, create, read, check status of and delete hosted workspaces
- List workspace sessions and profiles, and start another session in a workspace via the local Kilo API
- New local system feed route streams Cloud session preparation and status events
- Stored credentials are exchanged for short-lived Cloud Agent control tokens and never sent to the Worker
- Hosted attachments stop working once their session or workspace is deleted
-
CLI now installable as a conda package on Anaconda.org
- The CLI is now published as the kilo-cli conda package for Linux, macOS and Windows, installable with pixi or conda
- Agents can choose the base branch when starting Agent Manager worktree sessions
- Linked custom auto router entries reuse platform benchmarks, with model suggestions and per-variant link status
- Fixed auto-compaction starting too early and looping on models with unknown context size or after being stopped
- Patched zod in the VS Code extension to fix a memory-exhaustion advisory
-
Hosted cloud sessions now available through the local Kilo API
- Editors can find, read and work with hosted cloud sessions through the local Kilo API without ever handling the session's provider credentials
- Remote MCP servers that redirect to another origin, or answer POST with a 301, 302 or 303 redirect, no longer connect; use the server's final URL instead
- MCP TypeScript SDK updated to 1.32.1 to fix security issues, including MCP HTTP and OAuth requests following cross-origin redirects
- VS Code: copying selected chat text with Cmd/Ctrl+C or right-click Copy works again in VS Code 1.143, without clearing the selection or moving focus
- VS Code: DOMPurify and vite updated to fix sanitization issues and an editor CLI flag vulnerability (CVE-2026-112944)
-
Custom auto routers can now pick and run models on your machine
- kilo-auto/custom routers now route on the client, using your own provider models and running the chosen model locally
- New kilo auto router add, remove and list commands manage which provider models a router can use
- kilo auto eval runs missing client evals in Docker and uploads each result as it finishes
- The browseropen permission prompt can now remember always-allow rules for the current origin or all origins, and it appears in Auto-Approve settings
- Fixed subagents losing edit and write when delegated by an agent that only allows edits on specific paths
-
Agent Manager sessions can run custom agents, plus a redesigned worktree dialog
- Agents starting Agent Manager sessions can now pick which agent runs them, so custom agents and their permission rules apply instead of the workspace default
- Redesigned New Worktree dialog in VS Code: base branch, branch name and versions sit inline, and the Advanced options section is gone
- Kilo mobile app can now see and use the agents a remote-controlled CLI can run, including project agents from .kilo/agents
- Custom agents' own edit rules no longer limit the subagents they start; use a session-level edit restriction to confine delegated writes
- Model pickers now show why a provider failed to load and offer to sign in again when credentials were rejected
-
Kilo Gateway traffic now goes through a new default endpoint
- Chat, models, embeddings, autocomplete and transcription traffic now goes through ai-gateway.kilo.ai by default
- New KILOAIGATEWAYURL environment variable overrides the gateway URL, and setting only KILOAPIURL still routes traffic through that server
- Experimental Search Ignored Paths setting lets the glob tool search chosen ignored folders or files, such as build output
- The mobile app can fork a session from a specific message onto a remote CLI, restoring only the file edits made before it
- Fixes for grep on non-UTF-8 files, MCP sign-in resets and Windows background jobs that stayed active
-
Setting to let file search reach ignored folders, plus fixes
- New experimental Search Ignored Paths setting lists ignored folders or files, such as build output, that the agent's glob tool should still search
- Mobile app can fork a session from a specific message onto a remote CLI, keeping only the transcript and file edits up to that point
- .gitignored files and folders such as nodemodules are now kept out of glob results and grep include searches
- MCP sign-in no longer resets, and signing in to a remote MCP server updates the auth warning in every open session
- Windows background jobs use less CPU and stop properly, and grep no longer fails on files that aren't valid UTF-8
-
Combined agent, model and reasoning picker, plus many CLI fixes
- VS Code prompt toolbar now shows agent, model and reasoning selectors as one split control, with less used actions moving into a More actions menu
- Projects in the Agent Manager sidebar can be reordered with drag and drop, and the order is kept after a restart
- Integrated Browser pages can open in the external default browser, and Shift-click opens chat links in the system browser
- A plugin or config command named goal no longer hides every slash command and is now available as /goal:command
- Permission prompts now persist across TUI session switches, and compaction failures are reported instead of silently stopping
-
MCP OAuth sign-in, subagent steering and multi-project Agent Manager
- MCP servers now support OAuth sign-in, with recoverable auth failures flagged as needing sign-in and credentials kept when sign-in is cancelled
- In the TUI you can steer a running subagent from its view, or press Esc twice to stop it while the parent session keeps working
- New memorymodel setting in kilo.jsonc picks a separate model for automatic project memory saves
- VS Code: multi-project Agent Manager is now on by default, the Integrated Browser works in any session and Settings now has search
- Fixes a TUI that could start to a blank screen, write and edit failing in the Linux sandbox, and security issues in simple-git
-
Clickable next-step suggestions and auto-recovery from context limits
- Suggested next steps now appear in the CLI as clickable action buttons, showing whether each was accepted or dismissed
- Context-limit errors from providers are now handled by compacting the conversation and retrying
- Claude Sonnet, Opus and Claude-family aliases use the full output limit, so long reasoning responses no longer end early
- Failed reverts or redos now show an error instead of failing silently
- Sessions waiting on a wakeup or cron task now show as scheduled, along with their wake time
-
Clickable next-step suggestions and automatic recovery from context limits
- CLI shows suggested next steps as clickable action buttons and marks whether each was accepted or dismissed
- Context-limit errors from a provider now trigger automatic compaction and a retry
- Claude Sonnet, Opus and Claude-family aliases use the full output limit, so long reasoning no longer ends early
- VS Code: @ mentions suggest files from every folder in a multi-root workspace, and single sub-agents can be stopped from their task card
- Failed reverts and redos now show an error instead of failing silently
-
Agent Manager browser previews can now load public HTTPS pages
- Localhost previews now load public HTTPS pages and CDN resources in a high-resolution streamed Agent Manager browser
- The TUI shows network errors inline, waits for the connection to return and then resumes on its own instead of appearing to hang
- A running session cleanup can be stopped from settings, and the session database now shrinks on disk after cleanup
- Azure Entra ID sign-in accepts a resource name or a full endpoint URL
- Adds Cloudflare AI Gateway routing, GPT-5.6 none reasoning effort and more reliable Bedrock reasoning handling
-
MCP servers with companion skills arrive in the marketplace
- Install MCP servers bundled with companion skills from the marketplace, preview the included skills, and remove them along with the server
- Choosing a model, provider or reasoning effort for each subagent task is now on by default, with no experimental flag needed
- MCP OAuth now supports client metadata documents alongside dynamic registration and configured client credentials
- Smoother tool-call animations in VS Code chat, with live file paths and change counts while edits are being written
- Agent Manager fixes keep tabs, review context and browser pages when switching worktrees, and stop slow re-renders
-
Agents can schedule cron tasks, set goals and install Marketplace plugins
- Session cron scheduling with croncreate, cronlist and crondelete fires recurring or one-shot tasks between turns, and the tasks survive --resume
- New monitor action for backgroundprocess streams a long command's output as it arrives, so the agent no longer has to poll logs
- The agent can start or resume a session goal itself with the new goal tool, controlled by the goal permission
- Plugins can now be installed from the Kilo Marketplace, including from git repositories pinned to a ref
- Fixes kilo mcp auth failures, links sessions to pull requests more reliably, and stops empty compactions from replacing the conversation
-
Conversation-aware compaction and wider reasoning effort support
- Pulls in OpenCode v1.18.16 to v1.18.18 changes, including conversation-aware compaction and support for more reasoning effort settings
- Adds provider compatibility fixes and better retry handling from upstream OpenCode
- Pulls in OpenCode v1.18.14 to v1.18.15 changes: message ordering fixes, compaction serialization, more locale coverage and TUI improvements
- VS Code: approving permissions now retries after brief connection drops, so agents no longer get stuck waiting after auto or manual approval
-
Auto-compaction fix and "Sent by Kilo" labels on automatic prompts
- Auto-compaction no longer triggers far below the configured context threshold and now uses provider-reported token usage
- Prompts Kilo sends for you, like expanded slash commands and code actions, get a "Sent by Kilo" header and collapse when long
- Supports encrypted reasoning from OpenAI models accessed through Amazon Bedrock
- semanticsearch now names the root it searched and says if an empty result came from an incomplete, disabled or failed index
- Fixes MCP requests repeating every second, and project files with ${env:...} or ${file:...} placeholders now load
-
Sessions now link to GitLab and Bitbucket pull requests too
- CLI sessions can link to GitLab merge requests and Bitbucket pull requests, and kilo pr status keeps the link after the session exits
- New experimental Programmatic Tool Calling setting lets the agent call MCP tools from a confined JavaScript program, sending fewer tool definitions to the model
- Read-only gh commands run without a prompt in Plan and Ask modes and the Explore agent, and read-only git commands no longer ask to leave the sandbox
- Session titles are generated once there is enough context, instead of from a short first message
- kilo remote now logs when each remote session starts and ends, with model, working directory, exit reason and duration
-
KiloClaw removed, plus ChatGPT Codex quota in usage view
- KiloClaw is removed, including the /kiloclaw TUI command and chat view, the Kilo Chat client and its gateway endpoints
- ChatGPT Codex quota now appears alongside other provider usage in the CLI and VS Code
- Deleting an Agent Manager worktree is now near-instant in large repositories, with checkpoint cleanup finishing in the background
- Deprecated /local-review and /local-review-uncommitted aliases are removed; use /review with its scopes instead
- Skills listed in skills.paths with a leading slash, such as /.github/skills, now load from the project root
-
Marketplace moves into the CLI and malformed tool calls stop looping
- Marketplace catalog, installs, removals and installed-item detection now run in the CLI backend, so editor clients share one marketplace API
- A turn now aborts with an error after three identical invalid tool calls instead of retrying the same broken call forever
- Kilo Swarm setting moves from Experimental to Agent Behaviour and stays on by default
- The Kilo Swarm setting now uses the top-level sharedagentboard key; experimental.sharedagentboard is ignored and logs a warning if still set
-
Faster, smarter session search across long conversation histories
- Local session recall is faster for common search terms across large conversation histories
- Whole-word matches now rank above matches inside longer words
- When no session contains every term, flagged partial matches are returned instead
- Small typos are now matched against session titles
-
Scheduled session wakeups, permission feedback and faster worktrees
- The agent can schedule a session to resume later, list its pending wakeups and cancel them before they fire
- Permission denials can include feedback so the agent can adjust before retrying
- Kilo Swarm shared board is now on by default and can be turned off in Agent Behaviour settings
- Agent Manager creates worktrees faster by pre-warming reusable checkouts, and first snapshots start several times faster
- In VS Code, inline review comments can be posted to a checked-out GitHub pull request or sent to Kilo
-
Experimental Kilo Swarm agent board can now be enabled via env vars
- The experimental shared agent board (Kilo Swarm) can be turned on with the KILOEXPERIMENTALSHAREDAGENTBOARD environment variable
- The umbrella KILOEXPERIMENTAL environment variable also enables it
- The existing experimental.sharedagentboard config key still works
-
Mention a model inline with @ in VS Code, plus faster chat rendering
- VS Code: type @ in the prompt to open a model picker and insert an @provider/model mention for Agent Manager or subagent instructions
- VS Code: streamed text and reasoning now appear smoothly as they arrive, and bash, diff and to-do tool cards render faster
- Agent Manager notifications now open the right session, and duplicate permission prompts are fixed
- The working timer now pauses while a turn waits for your permission or answer
- CLI: Enter submits and Escape cancels a custom answer in the TUI question dialog when autocomplete is active
-
New /goal command keeps agents working toward a session goal
- New /goal command keeps working toward a session goal, with pause, resume and clear controls and completion or blocker reports
- Breaking: the interactiveterminal tool and in-session terminal controls are removed, so run commands that need keyboard input in your own terminal
- New /caffeinate command, plus notifications that explain when Kilo keeps the computer awake while agents work
- Opt-in one-time import of global Claude Code instructions, simple skills and MCP definitions
- View and reset the shared agent board without stopping agents or clearing conversations
-
PR review threads, CI fixes and Keep Awake come to Agent Manager
- GitHub PR review threads now show inline with avatars, with jump-to-diff, copy, send-to-chat and emoji reactions
- Failed CI checks can be sent to the agent from the PR panel, and PR checks are grouped with failures kept visible
- New default-off Keep Awake toggle prevents system sleep while agents work, without keeping the display on
- Agent Manager sessions can start in a chosen existing worktree and reply to the session that prompted them
- Large sessions load faster and live subagent transcripts perform better with many parallel agents
-
Experimental Kilo Swarm board and richer Agent Manager PR tools
- Kilo Swarm, an opt-in experimental shared board, lets parallel agents in a session post and read messages with boardpost and boardread
- Agent Manager PR view now shows discussion comments, review submissions, unresolved thread counts and a refresh button
- New /update-from-base command updates a worktree from its saved base branch without discarding uncommitted work
- The @ menu now searches files, past chats and entries with one ranked query, including names with spaces
- Agents can set default answers for single-select questions, confirmed with Enter in the CLI and VS Code
-
Kilo Swarm shared board for parallel agents, richer PR tools
- New experimental Kilo Swarm lets agents in a session post to and read from a shared board, opt-in via Experimental settings
- Agent Manager PR view now shows discussion comments and reviews, unresolved thread counts and a refresh button
- Update a worktree from its base branch with /update-from-base without stashing or discarding uncommitted work
- The @ menu now searches files, past chats and entries together with one query, including names with spaces
- Agents can set default answers for single-select questions, confirmed with Enter in the CLI and VS Code
-
Preview local apps in Agent Manager with built-in developer tools
- Preview local applications in Agent Manager with embedded developer tools, grouped diagnostics and element-level feedback for frontend changes
- Experimental setting lets the orchestrating agent pick the model, provider and reasoning effort for each subagent task
- Agent Manager terminal tabs, worktrees and subagent tabs now show live status: working, waiting for input, retrying, errors and completed
- Edit queued messages in VS Code before they are sent, keeping their text and attachments
- Stop the main agent without cancelling background agents, plus a Stop all button and faster forking of long sessions
-
Preview local apps in Agent Manager with built-in dev tools
- Preview local applications in Agent Manager with embedded developer tools, grouped diagnostics and element-level feedback for frontend changes
- Experimental setting lets the orchestrating agent pick the model, provider and reasoning effort for each subagent task
- Agent Manager terminal tabs, worktrees and subagent tabs now show live status such as working, waiting for input, retrying, errors and completed
- Edit queued messages in VS Code before they are sent, keeping their text and attachments
- Stop the main agent without cancelling background agents, plus a Stop all button and faster forking of long sessions
-
Agent Manager worktrees gain an @ picker and /review worktree
- Reference other Agent Manager worktrees with a searchable @ picker to share paths, branches and session IDs
- Review all committed and uncommitted worktree changes with /review worktree
- Choose a separate model for conversation compaction in Context settings
- Follow-up prompts in Agent Manager are queued when a session is busy instead of rejected
- Faster change loading, consistent status indicators and many fixes to background agents and chat
-
Background subagents on by default, plus Agent Manager upgrades
- Background subagents are now on by default, with completion notifications and a collapsible strip in the chat header showing running agents
- Edit, write and patch changes now open in the Agent Manager side panel, and multi-file patches show every changed file
- New Agent Manager document inspector previews Markdown with inline review comments you can send to the agent
- See your provider plan usage and quota windows in the CLI and VS Code profile
- Orchestrating agents can now answer a managed session's pending question instead of only stopping it
-
Reworked PR review comments in Agent Manager, now sent to the agent
- Agent Manager PR panel reworks review comments: resolved threads collapse into a group, replies show, and one button sends all unresolved comments to the agent
- Delegated subagent sessions open in Agent Manager inspector tabs alongside terminals
- New Display settings for the default open or collapsed state of MCP and tool blocks, and token throughput now shows by default
- Agent Manager shortcuts now follow the focused area, and terminals and nested sessions stay alive across config reloads
- NVIDIA Parakeet TDT 0.6B v3 is the new default speech-to-text model, and experimental tool-output pruning has been removed
-
Clickable file references in agent responses and PR comment actions
- File paths in agent responses now become clickable links that open the file at the referenced line, with a warning for dead links
- Agent Manager can resolve and unresolve PR review threads and jump to the comments section
- Built-in experimental Morph WarpGrep codebase search tool removed
- Subagents can now run their own allowed bash commands without inheriting a parent agent's read-only limits
- Bedrock and Vertex AI connections in VS Code accept structured AWS keys and Google Cloud service-account JSON
-
Agent Manager gets terminal shortcuts and repo picker for worktrees
- Agent Manager adds keyboard shortcuts to focus, create, switch and hide embedded terminals, and Cmd+Shift+M now focuses its prompt
- New --auto flag starts the TUI in a run-scoped auto-approve mode that can be exited mid-session from the command palette
- The /review command adds staged, unpushed and quick modes, with nested suggestions in VS Code
- MCP servers can add their instructions to the model context, and MCP resources are available as tools
- Model search ranks results by provider and suggests models based on your usage; multi-project worktrees let you pick the repository
-
Speech-to-text models now discovered from the Kilo Gateway
- Available speech-to-text models are now fetched from the Kilo Gateway, with offline fallback and organization model restrictions still applied
- The first Agent Manager terminal prompt appears sooner because optional terminal enhancements now load after the initial paint
- The last file stays visible when you expand multi-file patch results
-
Choose a model and reasoning level for each workflow
- Set a model and reasoning variant per workflow from Agent Behaviour settings
- Voice input bound to Cmd/Ctrl+K with hold-to-talk and release-to-send, and starts faster on macOS
- Agent Manager sections and worktree drag-and-drop work again when multiple projects are shown
- Multi-file patch tool calls now show total added and removed line counts
- Agent Manager keeps prompt and terminal focus per session, and ended terminals restart when you type
-
Multi-tab side-panel terminals in Agent Manager and a prompt navigator rail
- Agent Manager terminals can open in the VS Code terminal or an embedded side panel that keeps running in the background
- Side panel supports multiple terminal tabs with reordering, closing and live titles from the running program
- New prompt navigator rail in the chat transcript lists your prompts and jumps to any turn
- Holding Cmd/Ctrl now shows 1-9 shortcut badges on Agent Manager sidebar cards
- kilo console URLs now include basic-auth credentials for headless hosts, plus many MCP and checkpoint fixes
-
Reference past chats with @ to reuse earlier conversations as context
- Typing @ in the prompt opens a searchable picker of past chats and attaches the chosen session's transcript as context, in the CLI TUI and VS Code
- Expanding a tool call now shows why it was auto-approved and which rule allowed it, from agent, project or global config, or YOLO mode
- Queued chat messages can be deleted in VS Code before they run, and remote clients can now show queued messages
- The /sessions history is now filtered to sessions in the current Agent Manager worktree, and the sidebar tab bar gets a searchable open-tabs switcher
- Request and Gateway generation IDs are now shown when a model response ends unexpectedly, and plan implementation sessions open immediately
-
Simpler project memory settings with new slash-command completion
- Project memory settings simplified, with folder inspection replacing direct editing and legacy memory audit logs removed
- Nested memory slash commands now autocomplete and have status views, and empty projects are handled better
- Clear warning when a running session blocks a reload, instead of a generic Reload failed error
- Linux sandbox setup no longer fails when a writable directory contains an unreadable subdirectory
- Scroll position now resets when switching between diff files in chat and the diff viewer
-
Per-model reasoning effort in Compare Models, plus Shift+Tab effort cycling
- Agent Manager Compare Models can now set a reasoning effort per model, so worktrees can run the same prompt at different effort levels
- Shift+Tab cycles reasoning effort variants in prompt inputs, and a setting can turn it off to restore focus navigation
- Agents can now stop and remove a targeted Agent Manager session
- Codebase indexing supports a custom file extension allowlist to limit scans and cover more text formats
- Chat now renders right-to-left and mixed-language text correctly, and JetBrains gains Context settings for compaction and ignore patterns
-
Project memory gets a VS Code menu, quick actions and verbose details
- Project-memory activity in VS Code now has a task-header menu, quick actions and optional verbose details
- New verbose project-memory setting shows recalled memory snippets in conversation markers
- Distinct icons for context, project memory and code indexing, with grouped controls that keep task titles visible in narrow sidebars
- Fixed sessions freezing after submitting or dismissing a question
- Newly submitted messages now appear immediately after reverting a conversation
-
Chat search opens from the Command Palette and expands matching blocks
- Chat search can now be toggled from the Command Palette, and focus returns to the chat input when it closes
- Search auto-expands the collapsed tool call or reasoning block that holds the current match
- File @mentions now work for filenames with spaces or non-ASCII characters instead of failing with "File not found"
-
Agent Manager can now inspect and prompt other idle sessions
- The native Agent Manager tool can inspect managed sessions and send a prompt to an idle one, with separate explicit approval required
- Hovering a bar in the task timeline highlights the matching tool call in the transcript
- Prompt enhancer added to the New Worktree dialog
- Closing a session tab or the Agent Manager tab now stops active sessions and their subagents
- File mentions now respect read and ignore permissions, and sandbox state carries over to new Agent Manager sessions
-
In-chat search and multi-session tabs come to VS Code
- Search the current conversation in VS Code, with match case, whole word and regex options, and step through matches
- Open several sessions from the same repository as tabs from the sidebar and editor-tab chats
- New Browse files option in the @ mention menu lets you mention files outside the workspace
- Sandboxed tools can now reach network destinations you configure while other network access stays restricted
- JetBrains sessions gain rollback and redo controls, and orphaned kilo serve processes now shut down automatically
-
In-chat search and multi-session tabs come to the VS Code extension
- Search the current conversation from the session header, with match case, whole word and regex options plus next/previous navigation
- Open multiple sessions from the same repository as tabs in the sidebar and editor-tab chats
- New Browse files option in the @ mention dropdown lets you mention files outside the current workspace
- Sandboxed tools can now be allowed to reach specific network destinations while other network access stays restricted
- JetBrains sessions gain rollback and redo controls, and headless kilo serve now shuts down when its editor exits
-
Experimental AI image generation, plus a reload command for config and skills
- Experimental image generation tool for text-to-image and image editing, enabled via experimental.imagegeneration in config
- New /reload action picks up config, skills, agents, commands and MCP prompts changed on disk without losing sessions
- Project memory comes to the VS Code extension with settings controls, a /memory command and command-palette entries
- Images the agent reads now render inline in chat, and autocomplete works in Jupyter notebooks
- Commit messages are generated in the selected UI language, and codebase indexing respects nested ignore files
-
Experimental AI image generation, plus a reload command for config and skills
- Experimental image generation tool for text-to-image and image editing via Kilo Gateway or your own OpenRouter key, enabled in config
- New reload action picks up changed config, skills, agents, commands and MCP prompts without losing sessions
- Images the agent reads now appear inline in chat with a click-to-open full-size preview
- Project memory comes to the VS Code extension with settings controls, a /memory command and command-palette entries
- Autocomplete in Jupyter notebooks, clickable task timeline navigation, and commit messages in your UI language
-
Bidirectional text in prompts, plus sandbox and worktree fixes
- Prompt input now supports bidirectional text
- Fixed sandboxed file operations failing with "Module not found" due to a missing worker file in the VS Code extension bundle
- Fixed the reasoning-variant and mode pickers in the New Worktree dialog so selections actually apply
-
Agent Manager now names branches automatically after the task
- Agent Manager branches get task-focused names once the conversation's workstream is clear, with optional prefixes and a setting to turn it off
- Plan mode shows an Implement / Keep refining panel instead of asking you to type a numbered choice
- New opt-in Session Cost Alert warns before you continue a session that has spent more than a set dollar amount
- Closed read-only bash escapes where allowed commands could run any program through flags such as rg --pre and sort --compress-program
- Fixed slow message loading when opening or switching sessions with large histories
-
Optional sandbox limits agent network and file access on macOS and Linux
- Optional macOS and Linux sandboxes block outbound network access and confine agent writes to project and Kilo state directories
- Toggle the sandbox per session with /sandbox, or per worktree in the Agent Manager New Worktree modal
- Read, edit, run and create Jupyter notebook cells from Kilo in VS Code, with autocomplete now working in notebooks
- Marketplace lists skills, agents and MCP servers together, filters them to your workspace and suggests matching items
- New Anaconda Desktop local model provider, NVIDIA Parakeet speech-to-text, and per-session model and reasoning choice in Agent Manager
-
Background agents push live updates, plus OpenAI WebSocket support
- Experimental background agents now push updates without polling
- Experimental WebSocket transport for OpenAI responses, enabled with KILOEXPERIMENTALWEBSOCKETS=true
- Sessions can store custom metadata through the API and SDK, and config now loads from the opened location upward
- TUI adds a workspace management dialog and a prompt that resizes with terminal width
- Fixes for Google tool calling, Anthropic Opus 4.7+ adaptive reasoning, MCP server removal and resumed sessions
-
BYOK badge for Kilo Gateway models, plus chat and Agent Manager fixes
- Kilo Gateway models that can use your enabled personal or organization provider key now show a BYOK badge
- The chat reading area is wider, up from 88 to 98 characters, so conversations, tools and diffs use more of the editor
- Fixed chat output that stopped updating after you reverted and resubmitted a prompt, and Redo controls that disappeared after rewinding a session
- Agent Manager no longer goes blank during concurrent subagent updates or fails to start when another extension has registered panel commands
- Completion sounds now play only for parent agent sessions, and worktree hover cards appear and dismiss faster
-
Import past conversations from Roo Code into Kilo sessions
- New Import Sessions from Roo Code button in the About settings tab
- Finds conversation history from an existing Roo Code installation and imports it
- Reuses the existing migration wizard to bring sessions over
-
Richer custom providers, a TUI diff viewer and pinned sessions
- Custom providers now support multiple provider APIs, adaptive thinking, split reasoning and output effort variants
- The terminal UI gets a diff viewer with a file tree for reviewing changes, on by default
- Pinned sessions with quick-switch slots, an expandable thinking view, and --replay to show recent history when resuming runs
- Grok OAuth sign-in with device-code login, plus PDF attachments for Grok
- Agent Manager adds session and worktree search, opt-in attention sounds, a shortcut to open a worktree's pull request, and remote control
-
Per-model reasoning settings for subagents, plus several fixes
- Task subagents, including custom ones with their own model and variant, can now use model-specific reasoning overrides
- Reasoning can now be turned off again on custom provider models after being enabled
- Clearer permission prompts and auto-approval rules for potential loops, with localized tool names
- Redo now restores reverted sessions on the first click, and question responses go to the worktree where the question was asked
- Removing a skill no longer recursively deletes working directories
-
Pick an agent autonomy preset during VS Code onboarding
- VS Code onboarding now lets you choose an agent autonomy preset
- VS Code deep links can now select Kilo Gateway catalog models
- Escape now stops Agent Manager prompts even while their sessions are still starting
- Submitted review comments now appear as interactive message cards instead of raw markdown
- Turn outcome warnings no longer appear when the only cause is unfinished to-dos
-
Semantic search indexing now works without a separate Qdrant server
- Embedded LanceDB is now the default vector store for semantic search; Qdrant and Intel Mac users can choose qdrant with indexing.vectorStore
- Code indexing can be configured separately for global and project settings in Kilo Console, the CLI TUI and VS Code
- Large chat transcripts, session forks, subagent transcripts and Agent Manager reviews load and scroll faster
- Agent Manager terminals now use the integrated terminal font settings, including Nerd Font glyphs
- Fixes for duplicate tool call execution, stale permission prompts and permission approvals in worktree sessions
-
Fork completed Agent Manager sessions, plus smoother chat scrolling
- New Fork Session button on completed Agent Manager sessions
- Task timelines stay in place while you review earlier activity, and resume following from the latest bar
- Chat auto-scroll pauses when you scroll up over nested tool output
- JetBrains reasoning blocks open while streaming, collapse when done, and hide when empty
- JetBrains chat hides separator lines and shows search tool paths relative to the repository
-
Terminal Bench scores in model details and cloud session file restore
- Model details now show Terminal Bench completion scores and per-attempt costs for supported models
- Importing cloud sessions restores filesystem changes from synced session diffs, including changes inherited across forks
- Long streamed responses are faster because only the active part of the Markdown is re-rendered
- Agent Manager stays responsive with multiple concurrent sessions and sends permission approvals to the correct worktree
- Removed the obsolete speech-to-text enabled setting and renamed Mercury autocomplete labels to separate FIM and Next Edit modes
-
JetBrains chat gets smoother streaming and lower memory use
- Hidden JetBrains session UIs are now disposed after a configurable timeout to reduce memory usage
- JetBrains chat streaming is faster and code fences stay intact without showing raw markers
- Markdown code blocks in JetBrains show as multiline boxed editors that expand to full height
- Selection and copy now work across fragments of a JetBrains session transcript
- Kilo HTTP API clients keep worktree routing, and inherited task-subagent restrictions stay active
-
Clearer free-model training notices and faster session switching
- Free models now say when your data may be used for training and are marked with a brain circuit icon
- Shell command descriptions now appear in permission approval prompts
- Faster VS Code session switching for long transcripts
- Cloud Agent transcripts are back in VS Code session previews, and stalled cloud previews no longer load forever
- Fixes for Marketplace skill installs, message order after compaction, chat auto-scroll and unfinished inline review comments
-
Experimental Kilo Console now served from the local daemon
- Experimental: Kilo Console can now be served from the local daemon, and the kilo console command opens it automatically
- VS Code shows a retryable connection error and keeps unsent prompts when the background CLI process exits
- The Changes review keeps its scroll position while agents update files
- Background-process port discovery now runs only in the TUI and stops after startup, cutting unnecessary polling
-
DeepSeek joins Popular Providers and chat errors look clearer
- DeepSeek now appears in the Popular Providers list, replacing GitHub Copilot
- Improved styling of chat error messages in the VS Code extension
-
DeepSeek joins Popular Providers, plus clearer chat error styling
- DeepSeek now appears in the Popular Providers list, replacing GitHub Copilot
- Chat errors in the VS Code extension are styled more clearly
-
Mercury Next Edit autocomplete predicts multi-line edits
- New opt-in Mercury Next Edit autocomplete mode predicts multi-line and off-cursor edits, with Tab to jump and Tab to apply
- Mercury Next Edit also available through the Kilo Gateway using your Kilo account, so no separate Inception API key is needed
- Autocomplete model picker adds a "Not set (use server default)" option to follow the recommended default automatically
- Agent Manager tool is now available by default in VS Code
- Sessions can be renamed inline from the chat header and History, and a warning appears when a turn stops unexpectedly or leaves to-dos unfinished
-
Pick agent model and variant overrides from dropdowns
- Agent model and variant overrides can now be chosen from dropdowns
- Set a default model and reasoning effort for task subagents, falling back to the calling agent's model if unavailable
- New Mistral and Inception autocomplete options use your connected provider API keys
- Voice transcription now appears automatically when signed in with the Kilo provider
- Smoother streaming for inline subagents and Markdown code blocks, plus restored diff highlights in VS Code
-
Pick agent models and variant overrides from dropdowns
- Agent model and variant overrides can now be chosen from dropdowns
- Set the default task subagent model and reasoning effort, falling back to the calling agent's model if the override is unavailable
- New Mistral and Inception autocomplete options use your connected provider API keys
- Voice transcription now appears automatically when signed in with the Kilo provider
- Fixes for diff highlighting in VS Code themes, streamed code block stability and subagent streaming responsiveness
-
Agents can now run and manage long-running background processes
- Agents can start tracked background processes, such as dev servers, in the CLI and VS Code
- Background processes are cleaned up automatically when a session changes or ends
- The CLI adds process management UI with status and logs
- The TUI sidebar and process detail dialog show the ports each background process is using
Smaller updates 31
-
Maintenance release with no notable changes
-
Keeps the package manager's user-agent when reading npm config
-
Fixes for background job exit codes, compaction and multi-process use
-
Maintenance release with no notable changes
-
Maintenance release with no notable changes
-
Security hardening for URLs, MCP prompts and TLS certificates
-
Maintenance release with no notable changes
-
Security fix keeps MCP OAuth credentials tied to their issuing server
-
Every release artifact now ships with a signed software bill of materials
-
Fixes for agent switching and read-only bash commands in Ask mode
-
Maintenance release with no notable changes
-
Lower CPU use in large repos and fixes for codebase indexing
-
Shift+Tab works with slash menu open; clearer Agent Manager buttons
-
Fixes Windows startup blocked by PowerShell execution aliases
-
Fixes black-screen and Windows startup failures from v7.5.0
-
Security fix for markdown sanitization and less Agent Manager Git activity
-
Maintenance release with no notable changes
-
Maintenance release with no notable changes
-
Fixes file reads that could produce unusable tool output
-
Maintenance release with no notable changes
-
Maintenance release with no notable changes
-
Smoother chat auto-scroll and steadier layout
-
Fixes CLI failing to start on Apple Silicon Macs
-
Maintenance release with no notable changes
-
Maintenance release with no notable changes
-
Maintenance release with no notable changes
-
Maintenance release with no notable changes
-
Maintenance release with no notable changes
-
DeepSeek joins Popular Providers and chat errors get a cleaner look
-
DeepSeek replaces GitHub Copilot in Popular Providers list
-
Skills in .agents/skills stay on with Claude Code Compatibility off
Showing the newest 100 of 261.