An open-source multi-model AI agent workbench on your machine. Map a codebase, cross-check PDFs and spreadsheets, or finish a draft — the lead session dispatches subagents on the models you choose, while sessions stay in one local store shared by desktop, terminal, and browser.
An open-source multi-model AI agent workbench on your machine. Map a codebase, cross-check PDFs and spreadsheets, or finish a draft — the lead session dispatches subagents on the models you choose, while sessions stay in one local store shared by desktop, terminal, and browser.
Choose your preferred form factor — grab the standalone desktop app or launch the CLI in your terminal.
Ready-to-use desktop application with built-in server and visual sessions.
https://github.com/X-T-E-R/kiki/releasesReleases Place the binary in your PATH and launch an agent session anywhere.
$ kikiEnter from the task you have at hand: each page says what Kiki does with it, then links to the full steps and field reference.
Point Kiki at a codebase, a folder of PDFs and spreadsheets, or a draft. It reads what is there, plans, and works in place — and a session you start in the desktop app, kiki in the terminal, or kiki web in the browser is the same session everywhere.
When a job divides, the main agent dispatches isolated subagents — a fast model to explore, another to write or implement, a third vendor to review. Open any subagent's own transcript, while background tasks notify the lead when they finish.
See what it doesEach agent profile is one Markdown file you own: frontmatter sets its model, thinking effort, allowed tools, and which child agents it may dispatch; the body is its system prompt. Start from built-in roles like explore, implementer, and reviewer, or write your own.
Keep a target in view across turns with /goal, queue messages while the agent is busy with per-message send timing, run cron prompts while a Kiki process holds the session, and keep undoable memory across global, workspace, and persona scopes.
Give a recurring collaborator a name, standing rules, isolated memory, and a fixed daily conversation — or pull two to six personas and existing threads into a room to work through one topic under a host and a message budget.
See what it doesWhen work goes beyond plain text, connect a browser (Ego Lite, an extension, or an isolated profile) or install plugins for what you need: extract local PDFs and Office files (Kiki Extract), edit Word, Excel, and PowerPoint (Kiki Office Suite), audit CSVs (Kiki Work), query data (Kimi Datasource), or generate media (Kiki Media).
Replace any named part of the built-in prompt globally, per model, or per agent — down to rewriting one tool description a model misuses — inspect the effective prompt with kiki prompt-fields, and choose between Manual, Auto, Approve for me, and YOLO.
The Usage page names the concurrency rule holding each waiting request back, marks calls where a provider returned no usage as unknown instead of a fake zero, and syncs content-free stats only after you preview and approve the payload.
See what it doesWork files stay in your folder while sessions, history, and credentials stay in your local Kiki home. Isolate work and personal spaces, pair a WeChat or QQ chat to reply and approve from your phone, import earlier history as a session or archive, or connect editors over kiki acp and kiki seat.