Skip to main content
Most people should use the hosted remote server instead — nothing to clone or run. Run the servers locally when you’re developing against them, or where a hosted HTTP connection isn’t an option for your client. The repository is public and separate from the main Quiva application.

Requirements

  • Node.js 18 or later (the servers use the global fetch)
  • A Quiva API key — see API keys

Clone and install

Each server lives in its own package (quiva-flows-mcp, quiva-records-mcp, quiva-documents-mcp, quiva-workspaces-mcp, quiva-agents-mcp, quiva-distribution-mcp, quiva-coworker-mcp) and runs as an independent stdio process — connect to only the ones you need.

Configure credentials

The recommended credential is a profile: set QUIVA_PROFILE=<name> and the server reads the environment, account and API key from ~/.config/quiva/profiles.json. See Local AI agents to create one. Otherwise each server reads its own .env file:
Fill in one of the following. Credentials are looked up in this order: QUIVA_PROFILE, then QUIVA_API_KEY_<PROFILE>, then QUIVA_API_KEY, then the bearer token and email fallbacks: QUIVA_API_URL defaults to production (https://api.quiva.ai); leave it unset unless you’re pointed at a different environment.
An API integration credential is not an MCP credential. It only reaches the task-sync routes, so the servers cannot use it. Use a user API key.
The Abbie coworker server needs an admin or root user for its account-wide writes (organisation profile, task space, organisation environment). An email/password login or bearer token for such a user works; an API key does too, since the gateway resolves it to the same token.

Register with your client

Claude Code

Add each server you want:
Repeat for the other packages, changing the server name (quiva-records, quiva-documents and so on) and the path to its bin/run.sh. The -e variables stay the same. To avoid writing the key into your Claude Code config, use -e QUIVA_PROFILE=<name> instead of QUIVA_API_KEY. If the repository ships a .mcp.json at its root, opening a session inside that checkout registers everything in it automatically — approve the servers when Claude Code prompts you, and restart the session after adding credentials (MCP servers connect at session start).

Other clients

Any MCP client that launches a local command over stdio can run these the same way: point it at sh <package>/bin/run.sh with the environment variables above set, either in the client’s own config or via each package’s .env file.

Tool names

Running locally, each server’s tools are unprefixed — list_workflows, not flows_list_workflows — since only one server’s tools are loaded into that client connection at a time. The tool reference lists both the bare name and the prefix it carries on the hosted remote server.

Tool reference

Every tool, by server, with what it does
Using Claude Code, Codex or another local AI agent? See Local AI agents.