Nuanu Flow

Working with Nuanu Flow

Nuanu Flow is a work-management platform (a Plane fork) with an AI layer: BPMN processes orchestrate humans and AI agent employees, decisions gate approvals, and an artifacts registry stores versioned files. You talk to it through the bundled nuanu-flow MCP server.

Session activation

On a startup or resume turn, treat Nuanu Flow as the session's task tracker. When onboarding status is not already established in the thread, call the read-only onboarding_next tool at most once. Continue only its returned step when incomplete. If onboarding is complete, do not interrupt unrelated work. If the check is unavailable, continue the user's request without a retry loop.

The SessionStart hook may also provide a repository binding loaded from .nuanu-flow.json. Discovery is deliberately local and bounded: it walks only to the Git root, reads at most 4 KiB, performs no network call, and fails open. Validate the selected workspace/project lazily on the first real Flow operation.

Calling convention (read this first)

The nuanu-flow MCP server runs in compact mode by default and publishes only two tools:

• search_tools(query) — keyword search over the full catalog; returns matching tools' names + schemas.
• execute_tool(name, arguments) — run any catalog tool by name.

Every skill in this plugin names canonical catalog tool names (e.g. create_issue). Call them as execute_tool("create_issue", {...}) — you do not need search_tools when a skill already gives you the name and shape. If the server was switched to full mode (?optimize_context=full on the URL), the same names are directly callable as regular MCP tools.

Auth & connection — three modes

1. Proxy agent (default, interactive) — no env needed. On first contact the hosted MCP replies with an OAuth challenge; the browser opens Nuanu Flow, where the user can sign in or create an account and approve. A new account may authorize before it has a workspace; use onboarding next. You then act as that user, and your actions are attributed "via " (junction avatar in the app). Claude Code re-auth: /mcp -> mcp -> authenticate. In the Codex desktop app or IDE extension, use Authenticate for nuanu-flow. In Codex CLI, run the selected environment's codex:auth helper, which lets Codex open and wait for browser OAuth. If Codex reports Auth Unsupported, load codex-setup and verify OAuth discovery before using an advanced environment/Keychain fallback.
2. Ambient agent (headless) — NUANU_AGENT_KEY (nuanu_flow_…) is set; automatic inside worker-run task sessions. You act as the agent employee itself.
3. Manual token (CI/scripts) — NUANU_TOKEN (plane_api_… from Workspace Settings → API tokens) acts as the user without a browser.

Optional for all modes: NUANU_WORKSPACE (default workspace slug; overrides the consent-time choice), NUANU_MCP_URL (Claude endpoint override). Run /nuanu-flow:setup in Claude Code or load codex-setup in Codex for a guided check.

For every host, start with the universal installer prompt published in https://flow.nuanu.com/install.md. In Codex App, mention the installed plugin only after installation when the hosted guide requires a same-chat attachment attempt. After the host has attached or resumed, prove tool attachment by calling onboarding_next once before claiming setup is ready.

Object model

• Workspace (addressed by slug) → Projects (short identifier like ENG) → Flow items ("issues", addressed ENG-42).
• Flow items have: states (grouped backlog / unstarted / started / completed / cancelled), priority (urgent / high / medium / low / none), assignees, labels, estimates, sub-items (parent), relations, comments, attachments.
• Cycles = time-boxed sprints; Modules = feature buckets. Both contain Flow items.
• Teams group members and projects across the workspace. Objectives are portfolios that roll up projects. Views are saved filters. Automations are event → action rules.
• Processes = BPMN workflow templates; a run executes the graph step by step through human tasks, agent tasks, decisions, and gateways. Agent employees are configured AI agents (local runtime or remote workers). Decisions are human approve/deny/option gates inside runs.
• Artifacts = versioned files/documents in a registry, bound to entities (projects, runs, Flow items, …) and organized in logical folders.

Conventions that apply everywhere

• workspace_slug is optional in almost every tool. Resolution order is an explicit user/tool argument, the most specific matching repository scope, the root repository binding, connection default, NUANU_WORKSPACE, then the only accessible workspace. Multiple unresolved workspaces require an explicit choice.
• .nuanu-flow.json contains version, workspace_slug, a root project_identifier, and optional path scopes. The nearest matching scope wins. .nuanu-flow.local.json may partially override it for local development only and must remain gitignored. Neither file may contain secrets, endpoints, callback URLs, or identity data.
• Create a repository binding only after the workspace and project are confirmed, normally through project-setup. Authentication and account onboarding alone are not enough to choose a project.
• Human aliases work alongside UUIDs: project_identifier ("ENG"), issue_identifier ("ENG-42"), state_name, assignee_emails, assignee_names, label_names, parent_ref. Alias matching is exact, not fuzzy.
• Every scoped create_* tool returns the created entity's id in structured output. Account-scoped create_workspace returns {workspace: {...}}.
• Description/content fields named *_html take HTML, not markdown (<p>…</p>, <ul><li>…). Markdown pasted there renders as literal text.
• Lists paginate with cursor + per_page.

Which skill to load

For Codex remote enrollment, use the versionless instructions at https://flow.nuanu.com/connect/remote-agent.md. The copied prompt contains a short-lived nuanu_join_… enrollment token and installs the plugin first when needed.

Tools Used

search_tools, execute_tool, list_workspaces, get_workspace, list_projects, search_issues