# Heedroom > A workspace for business teams running AI agents in operations. The decision inbox comes first. Heedroom is for operations and finance teams whose AI agents work on invoices, supplier onboarding, refunds, inbox triage and bookings, where the people answering agent questions are often not engineers. Categories: AI agent workspace, human-in-the-loop operations, agent oversight. Built by Ostap Kovalisko (https://ostapkvl.com), a Eucalipse project (https://eucalipse.com/heedroom). Updated: 2026-10-30 ## Status - Interactive prototype with synthetic data. No real agents are connected yet. - A first version for one team and three agents is planned. - Pricing: free and open source (MIT). Self-host it or use the live demo. - Source code: https://github.com/quietloop-labs/heedroom ## Demos - Live prototype: https://heedroom.com/app - Guesthouse demo: https://heedroom.com/demo/guesthouse. A fictional 8-room guesthouse in Istanbul gets guest messages on WhatsApp, email, a booking portal and Instagram, in Turkish, English, Russian, German and Arabic. For each message the agent shows its steps: understand, check the booking records, draft a reply in the guest's language, ask for a decision, and the outcome with an audit line. Low-risk questions (Wi-Fi, breakfast) are answered under a cited house rule. Late checkout, date changes, refunds outside policy, an urgent leak and a suspicious payment link wait for the permitted person (Owner, Front desk or On-call). Fictional data; nothing is sent and no model is called. ## Key features - Decision inbox: the home screen lists only the moments where an agent needs a person: decisions first, then work to review, then reporting problems. - Decisions with context: each question shows what it blocks, which sources disagree, the options and their consequences, the role allowed to answer and when it expires. - Evidence side by side: conflicting sources are shown next to each other with the exact document version the agent read. - Answers bound to a data version: an answer is tied to a request id and to the version of the case it was given on. If the case changed, the answer is recorded but not applied and checks run again. - Audit trail: who asked, who it was presented to, who answered, with which reason, and when the agent picked it up. - Agents grouped by team: Finance, Operations, Inbox and admin, instead of repositories or branches. - Runtime and model per agent: each agent shows where it runs (for example Claude Code or Codex) and which model it uses. - Cost as an estimate: usage and cost are shown per runtime and always labelled as estimates. Missing usage is shown as not reported, never as zero. - Unknown state: an agent that stops reporting becomes unknown after a threshold, with its last-seen time. Silence is not treated as success. - Cases and data versions: a case page shows every data version, the documents with versions and hashes, and checks tied to the version they ran on. - Invite-only access and roles: Viewer, Operator, Approver, Admin. Admins do not automatically get approval rights. Agents cannot approve their own requests. ## Operating modes - Solo approver: one permitted person answers each decision, with a reason. - Two-person rule: two different permitted people approve. In the prototype (pick it with the demo mode selector in /app; the second person is simulated). - Concilium (concept, simulated in the prototype, no real model calls): three independent model reviews look at the case first, each with an option, a reason and cited evidence. A person sees what they could not settle. Model agreement is never shown as a human answer. Escalation rules: - All reviews agree with evidence, the case is low-risk and moves no money: continue without a person (Concilium only, simulated), with an audit note. - Disagreement, abstention, missing evidence, an incomplete review or unknown risk: escalate to a permitted person, with the disagreement side by side. - High risk: a permitted person always decides. - Moves money: two different permitted people approve, in every mode. - The case changes: earlier votes and approvals no longer count. ## Not yet - Real agent connections (planned: a small SDK/CLI, Claude Code hooks and Codex JSON output). - Concilium with real models. Today the model reviews are simulated; no model is called. - External business actions such as payments or emails to customers (out of scope for the first version). - Subscription quota meters (token counts do not establish remaining allowance). ## Comparison summary Facts about other products come only from their public pages on 2026-10-27. - Orca (https://www.onorca.dev/): a free, open-source (MIT) desktop app for macOS, Windows and Linux, plus a mobile app, for running many coding agents in parallel, each in its own git worktree. Released product. - Agent View in Claude Code (https://claude.com/blog/agent-view-in-claude-code): a terminal list of Claude Code sessions with a "needs input" group you answer inline. Claude Code only. Research preview since May 2026. - Heedroom: a web app where business people with roles answer decisions from agents doing operations work, across runtimes. Prototype with synthetic data. Which one to use: - You write code only with Claude Code: use Agent View. - You run many coding agents across Claude Code, Codex and others: use Orca. - Your agents do business operations and non-engineers answer their questions: use Heedroom. ## Pages - [Heedroom home](https://heedroom.com/) - [Source code on GitHub (MIT)](https://github.com/quietloop-labs/heedroom) - [Live prototype (synthetic data)](https://heedroom.com/app) - [Account, onboarding, roles and audit log screens](https://heedroom.com/account) - [Heedroom vs Orca](https://heedroom.com/compare/orca) - [Heedroom vs Agent View in Claude Code](https://heedroom.com/compare/claude-agent-view) ## Data (JSON) - [facts.json](https://heedroom.com/data/facts.json): product facts, features, status, not-yet list - [compare.json](https://heedroom.com/data/compare.json): comparison with Orca and Agent View, with sources - [queries.json](https://heedroom.com/data/queries.json): common questions with short factual answers ## Articles - [Heedroom: a decision inbox for business agents](https://ostapkvl.com/writing/heedroom-a-decision-inbox-for-business-agents) - [The agent is waiting for you](https://ostapkvl.com/writing/the-agent-is-waiting-for-you) - [Every value needs a source](https://ostapkvl.com/writing/every-value-needs-a-source) - [Heedroom project page](https://ostapkvl.com/projects/heedroom)