Skip to content

Set up Snipara for your repository

Choose Context to answer with project sources, or Local work to keep a task, its progress and its checks together. Both use the coding client and model you already have.

Primary path
Two normal setup paths
You choose the work you want to do. Local work prepares its compatible tools together; separate package selection is reserved for advanced integrations.

Context: connect project knowledge

For Codex and Claude Code, start with the native plugin. Open your actual repository, use the plugin's setup instructions, and follow its browser link to sign in or create an account. Create a workspace if needed, then choose or create the project for this repository.

For a manual Hosted MCP connection or another compatible client, run the guided Context profile from the project folder. Replace the client value with your supported client.

npx create-snipara@latest init --profile hosted-only --client codex

Hosted context requires a Snipara account and project access. Service subscriptions are separate from the free plugin. Authentication and project selection do not upload your repository. Confirm the destination and exact file list before an explicit upload or sync.

  1. Ask the connection to report its selected workspace and project.
  2. For native setup, read SNIPARA_USAGE.md after it has finished indexing. Ask one project question and compare the cited section with its source.
  3. Add the README or Markdown documents you select through Snipara's website or a supported upload/sync tool. Report missing or pending sources instead of inventing an answer.

Local work: one setup for the task

Run the command from the repository you want to work on. It prepares compatible Companion and Sandbox components in a private managed installation. You do not need to enable or maintain them separately for this path.

Codex
npx create-snipara@latest init --profile local-work --client codex
Claude Code
npx create-snipara@latest init --profile local-work --client claude-code

Local task records and deterministic repository validation need no Snipara account or LLM provider key. Fresh Local work setup starts locally, without signup or a Hosted project. Hosted context is optional: request it explicitly with --connect-context. If this repository already has a native Snipara connection, keep that source binding; local setup must not replace it with an unrelated global project connection.

  • Supply an objective, acceptance criteria and the allowed files. Keep the task brief, checkpoints and next step with the repository.
  • Use the coding agent you already have. Paid model dispatch requires separate explicit provider, permission and budget configuration.
  • Prepare and review the required checks. Managed validation derives a shared lot from the canonical request when no scope is supplied, so compatible CLI and MCP calls can share exclusion and verified reuse.
  • Read verification and review evidence before integrating. A recorded task closure or a successful installation does not prove tests, integration or deployment happened.

Start a new client session after setup. Saved task state can help you resume; it does not restore a running process. The normal task view distinguishes work, missing verification, required review and an integration recommendation without granting execution authority.

What preparation includes

The managed tools are installed as a compatible bundle. A replacement becomes active only after its preparation checks succeed; a failed preparation preserves the previous bundle. Existing task state and validation receipts remain separate from the tool installation.

Tool preparation itself does not install Node, Python, Docker, dependency images or a coding model. It does not start repository tests, authenticate Hosted Snipara, dispatch agents or deploy your application. The guided setup can separately connect Hosted context when you request it. A saved connection entry alone does not establish current authorization. Installation, project activation and runtime verification are separate steps.

Requirements for checks

  • Existing Node.js 18+, npm and Python 3.10+ for managed tool preparation.
  • macOS or Linux with a running local Docker daemon for repository validation.
  • A separately prepared immutable image containing the project's actual dependencies.
  • A reviewed validation request, bounded run and retained evidence for the exact sources.

A readiness check is not a passing repository test. If the required environment is missing, choose whether to configure it, use an explicitly authorized alternative, or continue without that test. Failure does not authorize execution on the host.

Check, repair and upgrade

npx create-snipara@latest doctor
npx create-snipara@latest toolchain status
npx create-snipara@latest repair --profile local-work
npx create-snipara@latest upgrade

Local-work diagnostics inspect the selected installation. They do not install missing tools, contact a model or run tests. Installed files, callable tools, authorization, environment readiness and verified results are separate states; an unchecked state stays unknown.

Repair rechecks local wiring; upgrade prepares a replacement bundle for an existing local-work installation. Advanced operators can prepare tools without project activation using toolchain prepare, or select the retained previous bundle using toolchain rollback. Reconnect the client after switching bundles. A live connection keeps its pinned components until it ends.

Changing repositories and adding sources

An open connection stays bound to its approved workspace. Open the new repository, repeat its setup when needed, and start a new chat/session unless project reload has been qualified for your client. A terminal directory change alone does not switch the chat's Hosted project or validation root. Confirm the selected project before retrieving or uploading.

Source selection remains explicit. Use a reviewed local file list, or connect the selected repository through Snipara's GitHub or GitLab integration for ongoing sync. Enabling push sync is separate from installing the plugin or local tools.

No project documentation yet?

The advanced Companion documentation bootstrap can prepare a reviewable Markdown brief from local evidence. Preview it, review the observed facts and unresolved questions, then explicitly apply and select it for sync. Existing guides stay under your control.

snipara-companion docs bootstrap --preview

Advanced and existing integrations

For a reproducible installer example, pin the package version. New versions are published by Snipara's release workflow; @latest selects the current published release.

npx create-snipara@2.2.18 init --profile local-work --client codex

Existing standalone APIs and profiles remain available. They serve explicit integrations; they are not prerequisites for the two normal journeys above. Existing flags keep their meaning and do not turn an older installation into a managed local-work bundle.

ProfilePurpose
hosted-onlyContext: Hosted MCP configuration without the managed local bundle.
local-workCompatible managed tools for repository task continuity, coordination and checks.
hosted-companionHosted MCP with an independent Companion setup.
full-stackHosted MCP, Companion and Sandbox, each set up independently.
runtime-onlyStandalone Sandbox setup without Hosted Snipara.

The manual default hosted-only profile writes the Hosted MCP connection and client instructions; it does not install the local-work bundle. Its default catalog exposes 13 tools with canonical public tool names; discovery exposes additional supported tools. Catalog visibility does not grant execution permission.

Advanced hooks, GitHub automation and provider-backed inference have their own explicit setup. Hook availability depends on the client and package; the remote Directory package contains no command hooks or packaged scripts and cannot select your local folder automatically. Its project must be chosen explicitly.

Standalone Sandbox 3 separates optional inference dependencies from deterministic validation. Provider-backed run and agent flows require the inference extra and a configured provider. None of these settings is needed just to read project context or perform account-free deterministic validation.

Existing AGENTS.md, CLAUDE.md, client connections and project artifacts are left unchanged. Generated merge-ready copies under .snipara/templatesare available for review. Review proposed instructions and generated configuration before merging them into your own setup. Never commit credentials or put a real key in a copied configuration example.

Last updated: September 24, 2026