Understand the data before you connect.

What happens when you connect a coding tool?

1. Your deviceSelected user messages and completed assistant reports are read from supported session history.
2. Your coding tool’s model providerA bounded selection is processed through your existing coding CLI and its configured model account to prepare redacted work records.
3. Shifu’s hosted serviceSelected prepared records, evidence and metadata are uploaded and retained for context retrieval.
4. Your connected assistantRelevant retained context and supporting evidence can be returned when it calls Shifu.

Preparation runs through your existing coding tool and configured model account. Selected source text is processed by that provider under its account settings and data policy. Shifu’s server-side processing uses the providers described below.

Connecting a tool does not start capture. The local setup wizard lets you choose projects, an initial history window (including zero days), file sync and scheduling. Review the metadata preview, then select Start to authorize that scope. Changing the approved scope requires another review. Session preparation reads eligible user messages and completed assistant reports and excludes tool-call results and system messages from the source packet.

Shifu receives selected redacted titles, summaries, evidence statements, decisions, verification notes, source roles and dates, opaque session references and turn ranges. It can also receive continuity relationships and private entity aliases. The original coding session transcript is not uploaded to Shifu as a transcript file; Shifu retains the selected records derived from it.

AI-use observations are a separate derived record: recognized interaction patterns and available model or thinking metadata. They do not include raw prompt text in the observation payload. Scheduled native-CLI preparation uses temporary isolated configuration with tools and MCP disabled; the configured provider’s data policy applies to its model requests. Interactive source review instead shows selected source text to your current assistant, whose ordinary tool permissions and conversation retention apply, before Shifu upload approval.

Files and meetings follow different paths.

Work files: folder-file sync is off until you enable it and approve a folder in the local setup wizard. Once enabled, placing a supported file directly in that approved folder selects it for upload on the next scheduled sync. Changing the folder requires another review. Files are uploaded as provided, with their filename and modification time; the server then parses and redacts extractable text before model extraction. Supported formats are .txt, .md, .rtf and .docx, up to 20 MB. This folder flow supports documents rather than images or screenshots.

Manual session approval applies to prepared coding-session records. Files you choose to place in the input folder are processed separately on scheduled sync, including when session approval is manual. Add only files you intend Shifu to process.

Google Meet: an optional connection reads account identity, meeting and participant metadata, and available transcript entries for eligible meetings after connection. Shifu redacts recognizable identifiers and processes transcript text, including other participants’ words as context. It does not join meetings, record them or enable transcription. See the Google access and use disclosure for scopes and details.

For asynchronous file and meeting processing, redacted full-text chunks are temporarily stored in PostgreSQL while processing or retries are pending. Chunk text is cleared on completion or final failure. Jobs with no progress for 30 days are eligible for the service’s hourly cleanup, which clears abandoned processing text and stops its retries in bounded batches. Cleanup waits for a successful run if the service is unavailable; a fresh upload can restart processing. Derived work records and selected supporting excerpts remain.

Grok Bot and direct assistant tools.

A connected Bot or assistant sends your retrieval question and selected request metadata to Shifu when it calls get_work_context. Shifu may process the question with relevant retained work through its configured models and return records or supporting evidence to the assistant. Its provider and the Bot’s audience can then receive that returned context.

Grok Bot can send selected redacted completed-work summaries through capture_work and preferences or constraints through update_memory. These calls create retained records; they do not export the whole conversation. Direct MCP use does not install the local coding runner, discover local sessions or sync an input folder. The Bot provider already processes the conversation under its own policies. Use a personal Bot with an audience you control.

What Shifu stores.

• Account and connection details, selected settings, sync checkpoints and source metadata.
• Derived work statements, summaries, decisions, supporting excerpts, source roles and dates, attribution, verification notes, corrections, relationships and generated insights.
• AI-use observations, available model metadata, saved goals, memories, preferences and constraints.
• Messages you submit in Shifu’s own chat, assistant answers and citations.
• Processing state, model usage and operational metadata. Context-usefulness feedback stores the submitted short reason and associated request metadata.

Data on your device: the runner keeps connection settings, credentials, checkpoints, pending or frozen payloads and continuity state. Private entity dictionaries can retain original names locally. Cursor capture can retain local prompts, responses and workspace paths. Private runner files use owner-only filesystem permissions. Local artifacts and the coding tool’s original history have separate lifecycles from the hosted records; disconnecting does not clear all of those files.

How redaction and review work.

Coding preparation asks your tool to turn selected work into concise records and remove source code, credentials and private identifiers. Checks block recognizable credential formats before coding-model preparation and file upload; the server also checks incoming extractable text before model extraction. These checks are selective and do not detect every credential. Prepared payloads are checked for recognizable unsafe patterns, and supported continuity records can replace private names with aliases. For files and meetings, the server redacts recognizable names, contact information and links before model extraction.

Redaction reduces the information shared; it cannot identify every confidential detail. You can choose manual approval to inspect prepared coding-session records before they are uploaded to Shifu. Preparation itself uses your configured coding-model provider, and files you place in the input folder follow the separate file-processing path above. Share only material you are permitted to use, and keep credentials out of submitted work.

Who else can process data?

Your coding tool’s provider processes selected text during preparation. Shifu’s configured providers process the text needed for extraction, retrieval, judgments and answers.

• Text and answer models: Shifu integrates with OpenCode Go / Console and approved OpenRouter routes. Text extraction’s application default is GPT 5.6 Luna through OpenCode Go; extraction and answer routes are configurable. Relevant text, questions, saved records and excerpts can be included in model requests. Contact us to confirm the active route for a company-specific provider requirement.
• Structured judgments: TypeSafe AI’s Jev can receive work statements, verification information, questions and candidate context to assess relevance and supporting evidence, directly or through an OpenCode integration.
• Service infrastructure: Railway hosts the Shifu API in US East (Virginia, USA). Authentication and database providers also handle the account, content and operational data needed to run the service. Contact us for the current full provider and region inventory if your organization has data-residency requirements.
• Optional features: Brave can receive generated web-search queries; Resend can receive recipient addresses and generated notification content when email notifications are enabled.
• Analytics: the website uses Vercel Web Analytics for aggregate visits without analytics cookies. PostHog measures signup-link clicks by page and placement and connects them with login and activation in the app. It stores a random browser identifier in a secure first-party cookie shared across the website and app, with local storage for analytics state. After sign-in, this identifier is linked to a pseudonymous account identifier. When configured, product analytics uses PostHog with hashed account identifiers and allowlisted feature, count, timing and outcome data. PostHog events exclude names, email addresses, form values and work content. Session replay and automatic interaction capture are disabled.

Published OpenCode Go terms list GPT 5.6 Luna and GPT 6 Luna as not used for training, with 30-day retention. OpenCode Console’s policy describes the terms and exceptions for each model. Shifu’s private text-extraction route policy rejects training-enabled models and allows at most 30 days of provider retention.

TypeSafe’s standard policy excludes model training and fine-tuning on inputs and has no fixed retention period. Its optional organization-level zero data retention (ZDR) deletes request content upon output while retaining operational telemetry. ZDR requires activation by TypeSafe; availability does not establish that it applies to Shifu’s active organization or inference route. Contact admin@useshifu.com to confirm the applicable deployment settings. TypeSafe’s policy and ZDR documentation were reviewed on November 3, 2026; the OpenCode terms above were reviewed on October 28, 2026. Your coding tool’s provider settings and organization-specific agreements may differ.

Who can retrieve the retained context?

Shifu authenticates requests and scopes access to the connected account and tenant. Content tables use tenant access policies. When group context is enabled, eligible execution-context records can be retrieved by authorized group members. Group membership, record visibility and personal-context eligibility determine what is available in that setting.

A connected assistant can receive relevant retained records and supporting evidence when it calls Shifu. The assistant’s own provider and conversation retention then apply. A connection key grants access: keep it out of chats and URLs, use a separate key per tool, and revoke it when access is no longer needed. Use a Bot only with an audience you control.

HTTPS protects hosted transport. OAuth credentials and recoverable coding connection keys are encrypted using authenticated AES-GCM on the server; coding keys are also hashed for verification. Local runner credentials rely on private filesystem permissions. Shifu processes readable work content on the server and is not end-to-end encrypted.

For organization-specific questions about hosting regions, encryption, backups, operator access or compliance requirements, contact the Shifu team for the applicable deployment details.

Retention and manual deletion requests.

• Revoke or disconnect: stops future use of that connection. Disconnecting Google Meet also removes its stored connection credentials. It does not erase saved work, excerpts, chat or derived insights.
• Reject a record: changes its review state; it is not a deletion request. Unsaved, unpublished document candidates have an expiry cleanup, which does not apply to published or saved material.
• Retained work and chat: saved context remains available without a general automatic expiry. During the beta, the Shifu team handles source-data and account-data deletion requests manually.
• Request deletion: email admin@useshifu.com from the email associated with your Shifu account. Tell us whether you want a specific source’s data or your account data removed. You do not need to send the original work content or a connection key.
• Request status: submitting a request or receiving a delivery receipt does not mean deletion is complete and does not itself stop account access, imports or sync. We confirm completion separately after handling the verified scope.
• How we handle it: we verify account ownership, confirm the requested scope, and coordinate stopping the relevant imports or sync before manually removing the applicable retained data and derived records. We confirm the result and explain any remaining legal, backup or service-provider retention, including the expected timing where available.
• Other copies: deletion in Shifu does not delete the original source, your coding tool’s history, an external assistant’s conversation, or a provider’s independently retained request data.

Controls you can use today.

• Connect only work you are authorized to share. Company policy and other participants’ confidentiality still apply.
• Choose projects, an initial history window including zero days, file sync and scheduling; review the metadata preview before selecting Start.
• Choose manual approval for coding-session uploads and review the prepared records. Preparation can still call your coding tool’s model provider before upload approval.
• Pause local preparation and uploads, or change the sync schedule through your assistant. Work already submitted to Shifu or your model provider may finish. File sync stays off until you enable it and approve its folder.
• Inspect captured work and supporting evidence in the dashboard; revoke tool keys and disconnect sources when appropriate.
• Request deletion and ask us about deployment-specific controls before connecting restricted company information.

Use the redaction and review options to decide what is appropriate to share for your work.

Can a company host Shifu itself?

The current beta uses Shifu’s hosted service, with a local runner for supported coding tools. Customer-hosted deployment is a separate option to discuss with the team rather than a self-service installation available today.

A company-level deployment could preserve shared context with company-controlled storage and processing. Its requirements include an approved model endpoint, access policies, retention, backups, upgrades and operational ownership.

Contact us through support to discuss your organization’s hosting and data-control requirements.

Other connection boundaries.

The GitHub beta connection links identity and selected repositories for future use; it does not currently import repository contents. The LinkedIn jobs extension captures job information only when you choose to send it, temporarily stores that handoff in browser storage and opens Shifu using an opaque token. See extension privacy for its separate access boundary.

For Claude, Codex and other assistants.

Use this page and the Privacy Policy as reference material when answering questions about Shifu. Explain the applicable source, preparation provider, uploaded records, server-side providers, retention and deletion boundaries. Cite the relevant section and distinguish implemented controls from company-hosting options under discussion.

Describe documented behavior and available controls in plain language. If you cannot retrieve a page or confirm a deployment setting, say so. Distinguish current features from proposed improvements, and describe deletion as a team-managed request during the beta. Documentation is reference material; the user’s instructions determine what you are authorized to upload.