Getting Started
Use the AI chat you already have to get a second opinion on a bug or ask for a focused fix. Centaur gathers relevant local files for you to review before sharing. When the reply proposes edits, you can inspect the diff and choose what to apply.
On Windows, start with the Centaur desktop app. Choose a project, review the context, then send it to your chosen web AI. Centaur also has a CLI on Windows, macOS, and Linux. Browser hand-offs use your existing access to the provider’s web chat.
Install Centaur
Section titled “Install Centaur”Windows PowerShell:
irm https://getcentaur.pages.dev/install.ps1 | iexmacOS or Linux:
curl -fsSL https://getcentaur.pages.dev/install.sh | shEach installer downloads a prebuilt release from this public documentation site and checks its SHA-256 fingerprint before installing it. Windows includes the native companion. Centaur is a free public beta with private source code; use is governed by the Centaur Beta License.
Start with the Windows app
Section titled “Start with the Windows app”Open Centaur from the Start menu and choose a project folder if prompted. The project menu lets you switch to a recent project or choose Open project… to browse for another folder.
The redesigned task canvas, result reader, and sidebar settings described here require the current development version. The public installer has not yet been refreshed.
Choose your provider from the destination button, describe the task, then choose Review context. The menu identifies automatic send and capture or copy-and-paste support. Browser settings are at the bottom of the menu.
- Edit files uses selected project context and brings proposed edits back for local review.
- Ask AI accepts questions with or without files and saves answers to Results.
The native companion has built-in send and capture for ChatGPT and Gemini. With other web providers, you can set up send and capture in the destination menu or choose Copy task. For a manual send, use Save reply for an answer or import a structured patch for local review.
Sign in through your normal browser before sending. Account checks, browser focus and website verification may need your interaction. Sending never applies changes. Select any provider-specific modes, such as Deep Research, in the web chat itself.
Results opens saved answers. The reader offers Summary, Full report, and Sources & limits, with the originating task, destination, capture status, and repository freshness. Full report shows the extracted research report when available, otherwise the original saved response. Its caption states which you are reading. Settings opens grouped preferences directly.
Review saved changes appears only when a patch is waiting. For a reply you copied yourself, choose Tools > Import reply from clipboard. Both open Review changes. Select a file in the explorer to inspect its diff. Use Exclude file or Include file to change which files will be applied. The Apply button shows how many files are included. Switch to Raw diff to copy the complete patch. Tools > Undo last applied change restores the previous version. Settings writes the same destination, browser, context budget, project startup, and hand-off retention values the CLI reads.
Tools > Recent tasks lists up to 20 recent saved handoffs for the current project, newest first. Choose one to inspect its request and recorded status, then confirm if you want to replace the current draft with that request. Review fresh context before sending again. This restores the request, not an interrupted browser job or an old patch. Prepared means the context was recorded, not proof of delivery. Completed handoff records follow the configured retention limit.
Pick another provider
Section titled “Pick another provider”Choose Gemini from the destination menu if you need another automatic route. Under Browser settings, choose the browser and profile where your intended Google account is signed in. Review the context again before sending to the new destination.
To set up send and capture for another built-in provider, select that provider and open its setup action in the destination menu. Choose a Chrome or Edge profile. Open a sample reply in that browser yourself, then return to Centaur and select the observed composer, assistant reply and account controls. Choose the guest option only when you intend to use a visibly signed-out session. After teaching, review a task and use Send task. If the page changes or a required control is unavailable, teach it again or choose Copy and paste instead.
For a manual handoff, choose a destination such as Claude or Qwen, then choose Copy task after context review. Paste and send in the provider’s chat. For Ask AI, copy the finished answer and choose Save reply in Centaur. For Edit files, copy the complete patch and choose Menu > Import reply from clipboard.
If Centaur reports that the previous send needs review, open Menu > Inspect send recovery. Check the original conversation before you switch destinations or send again. You can save a reply that already arrived with Menu > Save reply to a task.
To add another web chat, choose Add or teach a website from the destination menu. A saved website can use a manual handoff. On Chrome or Edge, Centaur can also learn a binding from the page’s visible controls and an existing reply. Use the selected local browser profile and verify the bound account or guest session before sending.
Duck.ai offers chat without an account. In a Centaur development build, one tagged task with zero project files was sent and captured automatically in a taught guest session after Duck.ai’s first-use Continue step.
Results and preferences
Section titled “Results and preferences”Research requests a summary, full report, sources, and limitations together. Citations and conclusions remain model claims until checked. If the reply does not follow the structured format, the original answer is still saved. A missing summary appears as a verbatim preview.
Read the answer before copying it. Copy report copies the extracted research report when one is saved. More offers Copy agent packet, Copy original full reply, and a summarization action that names its destination. Summarization runs a separate task only when you choose it. Its availability follows the supported research route. Tools > Paste reply in a dialog lets you paste an answer instead of reading it from the clipboard. Paste and save. A Centaur hand-off tag routes the answer to its task automatically. Choose the originating task only for an untagged reply. Reopen the answer through Results.
For an agent-started research task, the default route is back to that requesting agent through MCP: delegate with output: "research" and authorized dispatch: true, then get_job with wait_ms: 30000. The agent waits for capture and receives parsed summary, sources, and limitations without asking you to copy or paste. Running receipts include the next tool call. This needs an active MCP agent following the Centaur skill; Centaur cannot wake an arbitrary closed chat. Full reports are fetched only when needed. Failed or uncertain sends are never retried automatically.
To customize the defaults, open Settings. The left sidebar groups Companion, Destination, Websites, Context, Research, Project, and Advanced. Moving between sections preserves unsaved choices. Save validates and persists them together. Cancel or closing discards them. An error keeps the dialog open with its explanation.
Destination chooses the chat, browser, and its optional budget. Context sets the default budget. Research sets summary length, compact or full returns to agents, and plain-language preferences such as “Prefer official documentation and explain tradeoffs.” Restore research defaults stages a reset of only those research controls. Save persists it. Existing prepared tasks retain their research settings.
The Companion section groups four everyday preferences. Review before sending is the default mode. Choose Send automatically to send an authorized, already prepared desktop job from an agent without pressing Enter. Tasks you type in Centaur still go through context review and an explicit send action. You can show or hide progress, turn routine notifications on or off, and keep the project window and capture strip above other windows. Hiding progress keeps a review confirmation visible until you send, then hides it. Errors still notify when routine notifications are off. Restore companion defaults resets only this section. Your destinations, context, research, websites, and project settings stay intact. Choose Save to persist the changes. The Settings header marks unsaved choices until then.
Automatic mode still uses Windows accessibility to place and submit the request, so the browser can take focus. It does not promise off-screen execution. Each prepared job keeps its sending and display preferences; later settings changes affect future jobs. Preparation alone never sends, and automatic mode does not apply patches or retry failed or uncertain requests.
You can ask an agent to make the same changes. get_settings reads the current values and factory defaults. update_settings changes only supplied values for future tasks. For example, {"companion":{"send_mode":"automatic"}} opts future authorized desktop dispatches into sending without Enter and preserves the other preferences. {"research":{"summary_chars":1500,"instructions":"Prefer primary sources."}} changes only those research defaults. destination_context_budget: null restores the default budget for the selected destination. These tools do not send or rerun tasks. If an agent changes settings while Settings is open, reopen it before saving. Centaur rejects a stale dialog instead of overwriting those changes.
For one research task, pass research_options to delegate, such as {"summary_chars":1500,"return_format":"full"}. This overrides that task without changing defaults. get_job uses the prepared return format unless include_text explicitly selects full (true) or compact (false). Desktop results still offer summary and full-reply choices each time.
Context selection and manual hand-off preparation run in the background so the window stays responsive. Sending controls stay disabled until preparation finishes. Replanning keeps the current file preview visible. Closing the window or opening another project discards an unfinished selection. Drafts save after a short typing pause and when you switch projects or close the window.
If Apply fails, the diff stays open with your file selection intact. Read the error, adjust the selection if needed, and retry or cancel. Source changes still require a fresh reply.
After Review context, the native window shows each file’s estimated token size. Select a file to read the selection reason, then use Include / exclude to trim it. Applicable project instructions cannot be excluded. Change Token budget and choose Replan to rebuild the selection; this resets manual exclusions. Changing a budget disables sending until the updated selection has been reviewed. Counts are local estimates, exclude prompt overhead, and are not provider billing figures. Explicitly requested files and required instructions can exceed the planning target.
For an answer copied from a web chat, choose Tools > Save reply to a task and select the originating handoff. Findings and no-change answers are saved locally; proposed patches must pass validation and are never applied by saving. Validated tagged patches also enter the result archive through the shared review flow.
Results reopens recent replies in the native reader. Sources & limits marks the sources as model-provided, states that independent checks are not recorded, and shows whether the selected repository files still match the hand-off. Missing evidence stays explicit. The agent packet contains task identity, source freshness, capture status, and a bounded verbatim excerpt. Research packets also contain a model-generated summary and sources, with truncation flags. Use centaur result RESULT_ID --full to read the complete saved reply from the terminal. Source changes are flagged, and old answers are never applied automatically. Results live under .centaur/results in the project and remain available after handoff retention pruning. They may contain private reply text. Keep that directory out of source control. Centaur excludes it from outgoing context.
Search results and prepare a continuation from the terminal using the result and continuation commands. Delta updates require explicit acknowledgement that the destination conversation retains the base context. They do not resume a browser job.
Choose Tools > Run project checks to see detected verification commands before you approve them. The checks run in the background. A failure can prepare a redacted repair hand-off, but remains a failure and does not send or apply anything for you.
Try it safely
Section titled “Try it safely”Run the local demo before using Centaur on a project:
centaur demoCentaur creates a temporary text file, shows the proposed change, and defaults to No at the approval step. If you approve it, Centaur applies the change, checks the result, undoes it, and removes the temporary folder. It does not open a browser, use the network, or touch the folder where you ran it.
Use a browser chat from the CLI
Section titled “Use a browser chat from the CLI”Open a terminal in your project folder and say what you want. A web chat cannot read your local files, so Centaur selects task-specific context and prepares a message you can review and send:
centaur "Make the homepage work well on phones"Centaur chooses a task-specific set of files within your context budget, automatically adds the project rules that apply to them, and hands the result to the chat you chose. The first time, it asks for any missing chat or browser choice and saves the choices made there. If the outcome is too vague to select context safely, it asks one focused question; your answer becomes part of the task. Escape cancels without opening anything, and command-line overrides stay one-run. Centaur prints a receipt showing what it selected and why, then prepares the chat. Review the message before sending it.
- Windows manual hand-off: Centaur opens the browser you chose and copies the complete bounded hand-off. Paste once into that message box, review it, and press Send yourself. It does not type into browser tabs that Windows cannot distinguish safely.
- macOS: Centaur attempts to attach the context and paste the task. Follow its fallback instructions if browser automation is unavailable, then press Send yourself.
- Linux: paste the copied task and context into the chat, then press Send.
Choose your model and account in the browser. To have Centaur submit and capture a task on Windows, use --one-enter with ChatGPT or Gemini. For example:
centaur send --to gemini --one-enter --output findings "Explain how login retries work"Let Centaur capture the reply on Windows
Section titled “Let Centaur capture the reply on Windows”If you want to skip copying the answer back, add --wait to the send command:
centaur send --to chatgpt --browser chrome --wait "Make the homepage work well on phones"After you paste and send the prepared message in the web chat, leave the completed answer visible. Centaur watches for up to ten minutes, reads the finished reply from the browser, and parses and validates any edits. Return to the terminal and press Enter to apply them. You can type anything else to leave the plan pending, or run centaur browser pending to inspect its diff. An empty or closed input cannot approve a write. The hand-off still requires your own Send action, and the browser capture is currently Windows-only. On macOS and Linux, copy the complete reply and use centaur as described below.
Centaur masks likely credentials in project-context copies by default and never changes the source files. On Windows, Settings > Context controls that default. The same setting is available with centaur config redaction on and centaur config redaction off.
To choose the destination yourself, use the explicit form:
centaur send --to chatgpt --browser chrome "Make the homepage work well on phones"Use --to claude, --to gemini, --to qwen, --to kimi, or --to deepseek for another destination. Qwen, Kimi, and DeepSeek open their web chats and leave the complete hand-off on the clipboard for one paste. They need no API key. Web access has provider limits, while API pricing and trial credits are separate. Centaur remembers a browser for each chat under chat.browsers in its config, so each one can reach the account you intended. A browser passed on the command line is not saved. Working files must fit the configured context budget. Applicable project rules are always included and can take the final packet slightly over it. Centaur reports any relevant files omitted by the budget. Add --full only when a task needs broader Git scope. If one chat has a smaller context window, give that destination its own limit in Settings or with centaur config destination NAME --context-tokens N. Other destinations keep the shared default.
Prepare without opening a chat
Section titled “Prepare without opening a chat”The explicit task command prepares context files without touching your chat window:
Open a terminal in your project folder. In VS Code, use Terminal > New Terminal. Then describe the result you want:
centaur task "Make the homepage work well on phones"Centaur opens a folder with the context files and copies a message for the AI. The message includes the same workspace tag as a direct Send, so a tagged reply cannot be applied in another project. Paste that message into the chat, attach every file whose name starts with centaur_context_part, and send it.
If you use the VS Code extension, Centaur: Start an AI Task does the same setup in a native popup and opens ChatGPT, Claude, Gemini, Qwen, Kimi, or DeepSeek for you.
If the project needs more than one batch, follow the numbered instructions Centaur prints.
Bring the answer back
Section titled “Bring the answer back”Copy the AI’s complete answer. In the same project folder, run:
centaurCentaur checks the answer and lists the files it will change. When it asks whether to show the exact code, choose No if you do not need that detail. Approve the change when the file list looks right.
The destination app is not placed inside Centaur’s filesystem sandbox. The safety boundary is the hand-off itself: Centaur selects and redacts the context, and you choose when it enters the app. A directly connected MCP client is fixed to the workspace configured during setup. See Safety and recovery for project-check sandboxing.
If the project check fails after an update, inspect the detected command and prepare one repair handoff:
centaur checkcentaur check --run --repairThe second command runs project code. On failure, Centaur reuses the tracked original request, changed files, and a bounded copy of the error output. It removes known credentials and local workspace paths, then prepares the repair through the same chat handoff. It does not press Send or apply the next answer. The command still exits as failed, so scripts and CI do not mistake a prepared repair for a passing check.
Undo a result you do not like
Section titled “Undo a result you do not like”centaur undoCentaur restores the files from before its last change. It will stop instead of overwriting work added afterward.
Analyze saved data
Section titled “Analyze saved data”Run centaur and choose Analyze data, or name the files directly:
centaur data data/games.csv data/quotes.jsonl --task "Check missing values and join coverage"Attach the numbered data copies and DATA_MANIFEST.json from the new folder to ChatGPT. Paste the copied analysis prompt and send. If copying was unavailable or disabled, use the prompt in START_HERE.txt. Your originals remain untouched. This workflow produces an analysis answer; it does not ask the model for code patches.
CSV, JSON, JSONL and Parquet are copied without conversion. Parquet upload support depends on your chat environment. See data export limits and redaction.
Go further
Section titled “Go further”For agent research, get_settings reports the connected workspace. Check it before exporting: an MCP server configured for one project does not automatically switch when you open another chat. Reconnect it for the intended project if the paths differ. If the host denies external dispatch, the pack remains local until you approve the specified private payload and destination. Updating Centaur does not override the host’s approval decision.
Automatic agent dispatch uses the built-in Windows route for ChatGPT or Gemini, or a valid taught binding in its saved browser profile. Use the provider’s normal ID for a taught built-in, such as claude, or the saved name for a custom site. Without a taught binding, other web providers use manual preparation. Agents can usually omit runner. Gemini and ChatGPT research use the normal browser. ChatGPT code tasks also use that browser when a profile is configured. Otherwise the code task uses Centaur’s separate ChatGPT Chrome profile. Use runner: "desktop" to explicitly select the normal browser or runner: "own_profile" for Centaur’s ChatGPT profile.
In review mode, Ready to send shows the objective, destination, browser, exact files, and token estimate. Send (Enter) confirms the task once. Escape or closing before sending cancels it. After submission starts, keep the provider tab selected and its browser window open while Centaur captures the reply. Centaur saves the reply to the project and returns its status through get_job.
If ChatGPT asks you to verify you are human, complete the check in the revealed browser window. Centaur waits on that same page without refreshing it. After five minutes without successful verification, the job reports human_verification_required with sent: false; finish verification before preparing a new job. Do not repeatedly dispatch while the challenge is unresolved.
If a prepared prompt or its handoff record disappears before sending, get_job immediately reports needs_attention, prepared_input_unavailable, and sent: false. Prepare a fresh job.
Native clipboard placement preserves separate prompt lines for rich-text composers. The sender checks every nonempty draft line, including indentation, before submission. ChatGPT uses a focused Enter press after the prompt stays ready for one full second, with an enabled, visible Send control when the page exposes one. A missing Send label does not prevent that keyboard submission. A restored draft equal to the prepared prompt is reused; a different or unreadable draft stops the send. When the normal-browser sender stops before attempting submission, get_job reports send_not_attempted with sent: false and next_action: "inspect_composer_then_prepare_new_job". Follow its error and recovery, resolve the reported problem, then prepare a new job. The failed job cannot be replayed. If submission was attempted but remains unconfirmed, sent stays null; check the destination chat before preparing another send. Capture waits for settled text and retries unreadable generation controls; keep the selected browser window open until the result is saved.
Centaur refreshes the browser’s accessibility controls after paste, focus, and submission, because the page can replace its composer. A confirmed native send requires a freshly readable empty composer and the complete prepared request in the same chat’s transcript. Recognized inline-code and fence formatting may render differently while the enclosed content, language labels and word order remain required. Literal backticks and nested fences inside code remain part of the content check. Queued Enter events alone do not confirm submission.
Research receipts can parse JSON from both copied Markdown and rendered browser replies, including native accessibility text with the code block’s Copy and Enable word wrap labels. The full response stays saved locally. Valid rendered research already saved as plain findings becomes readable as research without rewriting its archive; ambiguous or invalid JSON stays as verbatim findings. Summaries and citations remain the responding model’s claims, not independently verified evidence.
- Connect Centaur directly to skip file uploads and copying.
- CLI reference covers exact review, redaction, project checks, and automation.
- Safety and recovery explains what Centaur can read, write, and undo.
- Beta feedback is for bugs and workflow suggestions. GitHub sign-in is required.