Kombai

Figma connects through Figma's own OAuth, and the agent reads one frame at a time from a link, while copying designs back to Figma goes through the clipboard. Worktree chats open a GitHub pull request through the GitHub CLI when it is signed in, or push the branch and open GitHub's compare page.

The agent calls external MCP servers over stdio, SSE or streamable HTTP with custom headers, static OAuth or dynamic client registration, with each tool switchable, indexes npm packages and Storybook from a Git address using the workspace's credentials, and runs terminal commands on its own. No native Jira or Slack connector exists.

Sourcedocs.kombai.com/features/mcp/external-mcp-servers.mdread 2026-11-08

Several agents run at once, in one canvas or many, and in the desktop app each chat can take its own Git worktree on a separate branch so parallel tasks do not collide. The default agent keeps a todo list as it works, Plan mode writes an editable plan file before any code, and Debug mode proposes a plan and applies it only after approval. Context Graph builds split their analysis across parallel subagents, a background tasks panel tracks work in progress, and a canvas node can be locked while an agent edits it.

Sourcedocs.kombai.com/features/worktrees.mdread 2026-11-08

Context Graphs index the components, hooks, stores, design tokens, types and constants in the workspace, in npm packages and, on the Team plan, in Storybook, built by dedicated subagents and kept between sessions until someone rebuilds them from Settings. A ContextRules file steers what counts as reusable, and graphs sync to a stack file so teammates pick them up.

Design systems persist as files Kombai builds from a codebase, a Figma file, a live website or a canvas design, a Component Library draws indexed components in up to eight states, and curated retrieval covers more than 400 frontend libraries. Where the index is stored is not stated.

Sourcedocs.kombai.com/context/context-graphs.mdread 2026-11-08

Plan mode stops for Approve and Code or Approve and Design before anything is written, and checkpoints restore the code, the conversation or both. Terminal commands run under one of three settings: Ask Every Time, Auto approve for me, where a classifier decides and which is the default, or Bypass permissions. Every change appears as a diff in the Changes panel, edits made by terminal commands can be undone, merging goes through a pull request, and closing a worktree asks whether to keep or remove it. The agent may use the browser unless a setting turns that off.

Sourcedocs.kombai.com/features/terminal.mdread 2026-11-08

Kombai's trust center lists SOC 2 as compliant without naming Type I or Type II or a report date, alongside encryption at rest and in transit and six named subprocessors including AWS. Single sign on with SAML, audit logs, custom retention and a DPA sit in the Enterprise column of the pricing table with no setup guide, while Team plans get three roles: Team Admin, Team Member and Billing Admin. Every plan promises no AI training on customer data, though the terms give Kombai a perpetual license to aggregated data drawn from customer data.

Sourcekombai.trust.siteread 2026-11-08

Credit logs record daily, rollover, purchased and expired credits, each chat shows the credits it used, and an Outputs card counts the files added and removed. Team admins see usage per member per day in near real time, but that view has no export and credit usage alerts are not yet available. Audit logs appear only as an Enterprise line on the pricing table, and no run trace, tool call log export or telemetry feed is offered for what the agent did.

Sourcedocs.kombai.com/kombai-for-teams/credit-usage/team-usage.mdread 2026-11-08

Agent Memory records what the agent learns about a project automatically and keeps it as Markdown files in the user's Kombai folder with an index file, turned on by default. Users can view, edit and delete each memory, which gives a purge path, though no expiry is set and no team or global scope is described. Rules at project, user and team scope, AGENTS files, skills and commands carry standing instructions alongside memory, and rules survive when long chats are compacted.

Sourcedocs.kombai.com/context/memory.mdread 2026-11-08

The desktop app runs on macOS, Windows and Linux, and its agents work on the developer's own machine against the local repository, with Design mode also working over Remote SSH. Local agents can run on the developer's own Claude Code or Codex subscription, though Kombai credits still apply to MCP calls. Kombai's own models, credits and hosted canvases run through its cloud service on AWS, and no hosting region, private cloud, VPC or self hosted edition is offered.

Sourcedocs.kombai.com/features/local-agents.mdread 2026-11-08

Kombai ships ready modes for design, planning, read only questions and debugging next to its default agent, plus frontend skills for improving a UI, accessibility and SEO audits, wireframes, themes, cloning a page from its address and generating images and video. A gallery of more than 20,000 designs, an Inspiration Library of templates and public design systems modeled on brands such as Discord, Stripe and Nike give teams starting points they can build from in their own stack.

Sourcedocs.kombai.com/features/modes.mdread 2026-11-08

Work starts from a chat in the desktop app or the editor extension, from slash commands and skills, or from Send to agent in the in app browser and the browser extension. Other agents can start Kombai work too, through the Kombai MCP server set up for VS Code, Cursor, Claude Code, Codex, Antigravity and Gemini CLI and through plugins for Codex, Claude Code, ChatGPT Desktop and Claude Desktop. Nothing starts a run on a schedule, from a webhook, from a repository or tracker event, or from a standalone command line or public API.

Sourcedocs.kombai.com/features/mcp/kombai-mcp-server.mdread 2026-11-08

Each chat picks one of 32 language models from 12 providers or the Auto router, which can change model on every turn, and models draw credits at multipliers from 0.03 to 5 times, with Auto between 0.1 and 1 times. Eight image models and 15 video models handle asset generation. Local agents can hand work to the developer's own Claude Code or Codex, but there are no bring your own API keys, and admins have no control over which models a team may use.

Sourcedocs.kombai.com/features/models.mdread 2026-11-08

The Kombai MCP server runs locally while the desktop app is open, showing a running badge and port, and writes its entry into the configuration of each client it supports. Kombai names no endpoint, no auth scheme and no list of MCP tools, giving only the skills it installs, such as designing, coding a design and improving or reviewing UI, and a recent release installs a lightweight skill in place of a full server entry. Plugins for Codex, Claude Code, ChatGPT Desktop and Claude Desktop sign in through the browser, and no public REST API or SDK exists.

Sourcedocs.kombai.com/features/mcp/kombai-mcp-server.mdread 2026-11-08

The agent runs the project's own lint and build checks after it writes code, and its browser tools add accessibility and SEO audits and Core Web Vitals readings for the pages it changed. Debug mode confirms a fix through the browser or the user's feedback, and the Improve UI skill closes with a verify stage. Kombai writes and runs no test suite of its own, no quality gate blocks a merge, and its Design Creativity Bench rates Kombai itself rather than the customer's output.

Sourcekombai.com/features/coderead 2026-11-08

In the desktop app the agent drives an in app browser by default or a Chrome window, and in the editor extension it always uses Chrome. It navigates, clicks, types, fills forms and scrolls through pages, with a separate browser session for each chat, and on macOS it can import cookies from a Chrome profile. The editable browser lets people change a page visually while Kombai turns those edits into code in the repository, so a visual tweak ends as a reviewable change.

Sourcedocs.kombai.com/browser/in-app-browser.mdread 2026-11-08