Developer Packages

Most users should start with create-snipara to connect shared Project Intelligence. Hosted MCP remains the agent surface; snipara-companion is the Git-style workflow continuity surface for local agent work; the public REST API is the supported software-integration surface.

Primary path

For LLM agents: use Hosted MCP

Do not start package selection by asking whether to install snipara-mcp. Start by connecting the agent to Hosted MCP.

Primary path

For product integration: use API / SDK

Use the REST contracts when Snipara is part of your product, backend, or data pipeline.

Default Recommendation

For LLM agents: use Hosted MCP.
For product integration: use the REST API.
Use create-snipara for Project Intelligence onboarding, auth, local config, and doctor checks.
Use snipara-companion for day-two agent workflows: status, briefs, timelines, spec-driven feature artifacts, hosted or local plans, phase commits, handoffs, resume, and reviewed local or hosted Context Control.
Use snipara-orchestrator only when production validation, htasks, or explicit multi-agent coordination are required.
Only install specialized packages when you already know you need them.

Harness capabilities and verification

The harness is the workflow around the customer’s coding model, not another model. Project Brain supplies project sources and reviewed memory. Companion and Orchestrator are optional and independent: either package can be used without the other. Companion retains plans, phases, tasks, attempts and evidence in the project folder. Orchestrator dispatches configured native tasks and binds validation and independent review to a candidate before controlled integration.

Define the allowed files, required checks and execution limits before dispatch. The native host keeps its own permission model. An independent reviewer may refuse a candidate even when its tests pass. Integration requires an explicit policy and an unchanged reviewed tree and patch; bounded retries are opt-in, and stopping for human attention is a valid result.

Context access and native execution are different capabilities.
Client	Observed local evidence
Claude Code / Claude CLI	Local full journey accepted with a test-local launcher for CLI controls.
Codex	Published-package journey exercised with proof, review and controlled integration.
LM Studio / Qwen Coder	External proof passed; independent review refused a defect and blocked integration.
Cursor, VS Code and other MCP clients	Project context and memory access; native controller execution remains adapter-dependent.

These are configuration-specific development observations, not universal compatibility, reliability or authenticated human approval. Saved-checkpoint resumption requires the project folder and its artifacts; it is not live-process migration or MCP-only recovery. See the published validation scope, Companion resume requirements and Orchestrator configuration.

Native plugin modules, setup and distribution

The plugin is an access and installation channel, not an additional execution engine. Snipara, Companion and Orchestrator are enabled by default in the native package and have independent on/off settings. Use the settings skill, then refresh or restart the coding client to load the generated profile.

Snipara supplies hosted sources, project context and reviewed memory. Connect through the client’s OAuth flow and choose an accessible project.
Companion supplies local plans, checkpoints, evidence and resumption. It needs explicit local engine setup and project alignment.
Orchestrator supplies configured execution, coordination, proof and review gates. It needs explicit engine, model, policy and permission configuration.

Enabling a module does not install its engine, dispatch workers or deploy a repository. Native OAuth credentials stay with the coding host; local engines use their own credentials. With Snipara off, the plugin restricts its local modules to status, help and version. Direct independent package use is a separate path from a disabled plugin profile.

The plugin is free; a Snipara account and project access are required. Hosted service subscriptions are purchased separately on snipara.com. The publisher’s GitHub marketplace is separate from official Anthropic and OpenAI directory approval; directory submissions being prepared do not imply approval.

Publisher: Starbox Group Gmbh. Plugin code and documentation: MIT. Snipara branding, hosted service, backend and local engine licenses are separate. The 2.0.0 prerelease notes describe that release; marketplace profiles may pin later local-engine versions. Follow the current client installation instructions.

Quick Matrix

Package	Use it for	Audience	Status
`snipara-mcp`	Local stdio bridge, client package development, advanced MCP testing	Developers and compatibility users	Developer tooling
`create-snipara`	Project Intelligence onboarding, free signup, local config generation, doctor, repair, companion workflow setup	Users who want a simple agent connection path	Onboarding
`snipara-memory`	Standalone memory engine when memory itself is the product you want	Specialized teams and OSS adopters	Specialized
`snipara-orchestrator`	Production-first validation, proofs, drift detection, htasks, advanced agent workflows	Advanced teams	Advanced
`snipara-sandbox`	Sandboxed code execution layer that complements Snipara	Users who need execution, not just context	Adjacent
`snipara-openclaw-hooks`	OpenClaw-specific hook automation	OpenClaw users	Specialized
`snipara-companion`	Git-style local continuity CLI: status, brief, timeline, phase commits, handoffs, resume, context-control plan/apply/drift, and hook-oriented workflows	Users who want managed agent work state across sessions and tools	Optional helper
`snipara-openclaw-install`	OpenClaw onboarding wrapper	OpenClaw users who prefer a guided installer	Published

Choose by Use Case

MCP for Codex, Claude Code, Cursor, ChatGPT, Gemini, VS Code

Start with create-snipara when the user wants guided onboarding and auth. Start with Hosted MCP directly when they already have the project slug and API key. Use local snipara-mcp only when the client needs stdio, you want browser-login compatibility testing, or you are developing MCP client behavior.

Continuity for agent sessions

Use snipara-companion status, brief, timeline, workflow phase-commit, handoff, and workflow resume when the agent needs a durable work log rather than just a one-time MCP connection.

Spec-driven feature delivery

Use snipara-companion feature init to create a reviewable specification, then choose a Hosted Snipara plan with feature plan or author local numbered phases in plan.md and import them with feature tasks --from-plan. Both paths feed the same managed workflow chunks and keep .snipara/workflow/current.json as the only runtime state.

Context control for reviewed mutations

Use snipara-companion context-control plan, apply, drift, and validate when local Project Intelligence state needs the same trust pattern as infrastructure changes: preview the exact diff, apply the reviewed plan, and report drift. Context Control adds hosted-diff and hosted-apply: the backend scopes the project from authentication, rechecks remote hashes atomically, requires an approved Decision Request and EDITOR API-key access, and returns a per-source receipt. Hosted reconciliation is add/update-only. It never deletes unmanaged hosted documents and blocks managed authority promotions.

npx -y snipara-companion@latest context-control plan --summary "record reviewed context state" --output .snipara/context-control/plans/demo.json

npx -y snipara-companion@latest context-control hosted-diff --manifest snipara.project-context.json --output .snipara/context-control/plans/hosted.json --emit-decision-request

npx -y snipara-companion@latest context-control hosted-apply --plan .snipara/context-control/plans/hosted.json --approval .snipara/decisions/resolved/<request-id>.json

Codex specifically

Use Hosted MCP config plus AGENTS.md. The default hosted-onlyprofile does not install snipara-companion. Select --profile hosted-companion or run npx create-snipara@latest repair --with-companion when local, compaction-safe workflow commands are useful. Local snipara-mcp is only needed for stdio compatibility, development, or tests.

Business workflows

Use Helvabase for RFP, proposal, client dossier, and reviewed business-source workflows. Snipara package docs stay focused on developer and agent integration.

Python application integration

Use the REST API. A snipara Python SDK exists in the monorepo as preview source, but no snipara distribution is currently published on PyPI. Do not route product backend work through a local agent helper flow.

Production validation and orchestration

Use snipara-orchestrator only if you need proof-based validation and production drift detection, or if you are intentionally coordinating hierarchical tasks across agents. It is an advanced package, not the default install path, and it is not bundled into npx create-snipara --profile full-stack.

Companion vs orchestrator

Do not merge snipara-companion and snipara-orchestrator. They solve different problems: companion is a local helper CLI for hook-oriented workflows, while orchestrator is an advanced Python package for production validation, htasks, and orchestration.

OpenClaw

Connect through Hosted MCP first. Use create-snipara only if you want scaffolding, then add snipara-openclaw-hooks for OpenClaw-specific automation. The native OpenClaw plugin is the preferred path. The published snipara-openclaw-install wrapper remains available and is not marked deprecated in npm; use snipara-openclaw-hooks when you specifically need hook automation.

Recommended Paths

Agent Connection

npx create-snipara

Guided free-account, auth, Hosted MCP, context, and memory onboarding.

Agent Continuity

snipara-companion status

Git-style status, briefs, timelines, handoffs, and resume for long-running agent work.

Product Integration

curl https://api.snipara.com/...

Use the documented REST contracts for backend-to-backend and application workflows.

Related Guides

Last updated: September 24, 2026