Skip to content
Browse docsRoadmap

Documentation Roadmap

What we’re working on.

The state of Tilden across Workspace, CLI and MCP: what is live, what is built and waiting, and what comes next. Priorities follow what we learn with invited teams. This is direction, not a delivery calendar.

Updated November 7, 2026

What is live

  • The CLI, free and open source. aibill 0.9.17 is on npm (published November 7, 2026) and runs on your machine: OpenAI and Anthropic usage exports, Cursor and Copilot, Claude Code and Codex session logs; a spend report, limits, context health, a dead-context scan, and workspace connect and push. CLI guide →
  • What 0.9.17 fixed. On 0.9.15 a second push could count a folder’s day twice, compressed Codex sessions were skipped, several OpenAI price rows were wrong, and the Anthropic sync crashed on one documented date form. All four are fixed in 0.9.17, which also carries the one-push subscription tier and aibill guard. Move each machine to aibill@latest before its next push. 0.9.16 was tagged and never published.
  • Workspace cost view across providers. Hosted, invite-only beta. OpenAI and Anthropic cost reports, read from provider admin keys at daily grain and grouped by provider project, workspace or key, with up to two years of history (one year for OpenAI), refreshed on a schedule. The key Tilden holds for this never writes. Overview, Briefing and the hosted MCP read one ledger, so they cannot disagree. Cursor and Copilot connectors exist for allowlisted tenants; their figures never enter the Overview number. Workspace guide →
  • Receipts per project. A month’s AI cost per project, never per person, on the receipts page, in the Briefing and through the workspace_roi MCP tool, in sentences with the gaps named. Discover proposes projects from your own cost data and applies the confident ones; each is labelled “Proposed by Tilden” and can be undone. Seat plans come from a machine’s report, not a form; they fill in once a machine on 0.9.17 pushes its subscription tier.
  • Hosted MCP. A supported AI client signs in with OAuth 2.1 and reads spend, projects, briefing, budget settings and the month’s receipts. Read-only: it cannot refresh providers, change budgets or manage keys. Live on our own tenant; invited teams get it through assisted onboarding. MCP guide →

Built, waiting on a release or a connection

  • Seat tiers from your machines. 0.9.17 is out; the receipts page fills in each coding agent’s seat cost once a machine on it pushes. aibill guard, coding-agent policy with CLI hooks and a model allowlist, is on npm and waits to be installed in a repository.
  • Results from accepted code changes. The read-only GitHub reader and the verdict rules are merged. The GitHub App is being set up and has no installations yet, so no project shows a result, a return or a verdict, and no page pretends otherwise.
  • Caps and pauses. Per-project caps, an OpenAI rate-limit pause and an Anthropic key-status pause with Resume are built behind a separate, opt-in Enforce key. No live cap has run yet; the first one is next.
  • Invoice reconciliation. The monthly close is merged but not enabled on any tenant, so no month reads “billed” yet.
  • The Monday briefing. The Briefing renders on demand in Workspace today. The Monday email and its schedule are not built and wait for the first design partners; we will not promise them in copy until they ship.
  • Fixes from the November 7 audit. Refresh scheduling (empty days re-read forever, a provider row stuck on its first sync) and Briefing edge cases (a silently substituted window, a blank next step). Being fixed this week.

Results and return, per project and per agent

  • Results per project and per agent. Once the GitHub App is connected: accepted code changes set against the month’s receipt, first per project, then per agent. Rates, return and verdict get exercised on our own tenant before any partner sees them.
  • The first live cap. On our own tenant with a test Enforce key; then Codex policy; then enforcement on a worker cadence instead of at render time.
  • Local git and other code hosts as proof sources. GitHub is the first connector, not the rule. An accepted code change is the unit, wherever it lives.
  • One key, several folders. Splitting one API key’s cost across folders at key grain. It needs Anthropic’s analytics API; today it runs on fixtures only.
  • The Monday briefing email and its schedule, after the first design partners are on board.

What Tilden will not do

No productivity scores. No company-wide ROI percentage that Tilden asserts. No public pricing page; plans are discussed in sales conversations. Receipts, seats, controls and results stay per project and per agent, never per person.

Bring a real spending question

Request access if you want to help shape Tilden. Selected partners receive assisted onboarding. Tell us which sources you use and what you cannot explain about your AI spend today; requesting access does not grant it immediately or promise a delivery date.