Under the hood.

Helen links changes from people and agents to your company's plan, and checks the plan every day.

The plan is her reference point, and you ratify every piece of it.

Start from your agent. One command, one prompt.

Your agent signs you in, creates the workspace, and drafts the first context chain from the repository, issues, PR history, and documents it can already reach.

Your team ratifies the decisions that matter. After that, each agent can load the ticket, its context chain, open findings, and specifications for its task instead of reconstructing them.

  1. 1Add Helen's Foundry to Claude Code.

    claude mcp add --transport http helens-foundry https://hf.dev/mcp

    In Codex, Cursor, or another MCP client, add https://hf.dev/mcp as a remote server. Your browser opens to sign in with Google, GitHub, or your email.

  2. 2Paste this prompt.

    Use the Helen's Foundry MCP server. If I have no workspace there, create one for our company. Then follow its getting-started guide to draft our strategy, goals, and customers from this repository and what you know of our plans. Tell me when something is ready for me to ratify.

    Each decision your agent drafts counts only after you ratify it in Helen's Foundry.

  • activity_feed_list
  • artifact_types
  • artifacts_list
  • artifacts_search
  • members_list
  • workspace_audit
  • workspace_summary
  • workspaces_list
  • artifacts_get
  • chain_get
  • decisions_list
  • docs_get
  • external_references_get
  • external_references_list
  • pr_report_get
  • artifacts_create
  • artifacts_delete
  • artifacts_update
  • decisions_archive
  • decisions_record
  • links_add
  • links_list
  • links_remove
  • person_decisions_request
  • proposals_list
  • reconcile
  • signals_list
  • send_feedback
  • workspaces_create

The plan as a graph. Four tiers, each informing the next.

A ticket links up through its project to the goals and strategy above it, so any piece of work leads back to the decisions it serves.

  1. Orient Mission, vision, and domain model
  2. Directional Strategy and goals
  3. Structural Personas and problems
  4. Execution Projects and tickets, with requirements on each ticket

Decisions stay current.

Revise a decision and the new version replaces it. Every earlier version stays on record, and an agent can ask for the chain as it stood on any date.

Work is checked against the current version. Work still under a decision you closed gets flagged.

Honest onboarding.

Helen drafts your foundation from pasted text, a link, or a few sentences. Your coding agent can also draft it over MCP from the repository, issues, and docs it already reaches.

Every field is tagged: extractedinferredabsent

Only the mission and vision can be inferred. Everything else comes from your sources or stays blank for you to fill.

What shows up on the pull request.

Link the PR to its execution artifact in the PR body with Implements <artifact link>. Closes, Fixes, and Resolves also work. Helen never approves or requests changes; your reviewers keep those calls. The GitHub App posts two things.

hf.dev / linked check run
ResultWhen
SuccessThe PR links to an execution artifact (project or ticket), and every authority up to three levels above it is ratified.
NeutralThe PR has no artifact link but is link-exempt, the workspace has nothing to link yet, or an authority in the chain is unratified and the workspace is in advisory mode.
FailureThe PR has no execution artifact link and is not link-exempt, or an authority in the chain is unratified and the workspace is in enforcing mode.

The check blocks a merge only when you make it a required check in branch protection.

helens-foundry bot commented · updated on each push

Linked to SAML sign-in for enterprise workspaces · ticket

SpecificationSource
Admins can require SAML sign-in for every member.SAML sign-in for enterprise workspaces · ticket
A seat belongs to one member in one workspace.Workspaces and seats · domain model

Ratify Enterprise SSO to clear the neutral check.

The sticky PR comment. One comment, updated in place on each push. It names the artifact the PR is linked to, lists the specifications that apply and where each comes from, and names any premise that still needs ratification.

Findings.

Each flag Helen raises is a finding, and each one waits in Reconcile for a person to acknowledge, fix, or dismiss it. Helen keeps a dismissed finding closed while the evidence behind it stays the same.

Reconcile keeps conflicts, two decisions in tension, apart from gaps, something the plan should say and doesn't.

Integrations.

Bring in context from the tools your team already uses, and connect any MCP client.

  • GitHub
  • Linear
  • Jira
  • Notion
  • Confluence
  • Slack
  • Sentry
  • Figma
  • Trello
  • Productboard
  • Aha!
  • Canny
  • Dovetail
  • Intercom
  • Zendesk
  • Freshdesk
  • Guru
  • Attio
  • incident.io
  • Any MCP client

24 checks, grouped by how they work.

Each check raises a finding that a person reconciles. Detector pages explain the method, its limits, and the research it cites.

Structural 13

Rules over the context chain: owners, links, priorities, and lifecycle fields.

  • Building on a discontinued foundationExplainer Active work still builds on an artifact above it that was abandoned, marked obsolete, or marked a duplicate.
  • Stranded by a completed parent Open work sits under a parent that completed or migrated, so the artifact needs to close or move to an active parent.
  • Two people working on the same thing Two in-flight artifacts under the same parent have different owners working on them at the same time.
  • Lifecycle data inconsistency An artifact's status fields disagree with each other, such as a closed artifact with no closure reason or an open artifact with a closure reason.
  • Duplicate singleton More than one open, unarchived artifact exists for a type that allows only one in the workspace, such as mission, vision, or strategy.
  • No work addresses this yet An active persona or problem definition has no active project linked to it.
  • Circular dependency A chain of depends_on links leads from an artifact back to itself, so the work has no valid order.
  • In progress, but unassigned Work that is in progress or in review has no owner assigned.
  • No shared top-level goal The workspace has several top-level goals or strategies with no shared parent above them, so work below them can point in competing directions.
  • Not broken into tasks yet A mission, vision, strategy, goal, or domain model has children, but no projects or tickets beneath it.
  • In progress with no priority set Work that is in progress or in review has no priority set, so you cannot rank it against other in-flight work.
  • No owner A committed strategy or goal has no owner accountable for its outcome.
  • Urgent, no owner A goal or strategy has urgent work beneath it, but no work under it has an owner assigned.

Recency 3

Time-aware rules: staleness, validation windows, and plans frozen while their work changes.

  • Stale project with no active work An active project has only closed work and has not been updated recently.
  • Needs ratification or validation A committed artifact lacks current human ratification, or its validation is missing or past its expected window.
  • Plan hasn't kept up with the work Work below a strategy or goal keeps changing while the plan itself has not been updated.

Semantic 4

An LLM judges a pair of artifacts for contradiction or duplication.

  • Contradicts a related artifact Two related artifacts appear to state conflicting goals or directions.
  • Possible duplicate intent Two artifacts in different places appear to state the same intent, creating duplicate records.
  • Contradicts its parentExplainer A child artifact's intent appears to conflict with the parent artifact above it.
  • Copied from parent instead of decomposed A child artifact appears to repeat its parent instead of breaking the work into a specific step.

Vector 1

Embedding geometry over sibling work.

  • Sub-items don't fit together The children of a strategic artifact pull in unrelated directions instead of supporting one shared purpose.

Property 2

An LLM judges one artifact against its context.

  • Metric no one is working toward A goal or strategy states a measurable target, but no work below it addresses that target.
  • Doesn't meet its quality rubric The artifact fails its type, writing, or specification rubric. Grades and suggested revisions explain what needs attention.

Portfolio 1

A workspace-wide signal instead of a per-artifact finding.

  • Distribution observation Work under a parent is concentrated by priority or branch, or has no priority set. Check whether this distribution matches your priorities.

Start with one repository this week.