Source: André Lindenberg's LinkedIn post (2026-11-11) on the new Inbox feature of Plannotator, an open-source (MIT / Apache-2.0, about 10k stars) local review tool for AI coding agents.
TL;DR
Most coding agents ask for human input by blocking: the session sits open, holding its context, until someone answers. Plannotator Inbox flips that. Agents post questions, plan reviews, files and guided reviews to one local window and keep working or end their turn. The human clears the queue in one pass, and each answer is delivered back to its session as the agent's next turn. The human becomes an asynchronous reviewer across many agents, not a synchronous gate on each one.
What Plannotator is
A browser-based review surface that sits between a developer and their agents. It wires into each agent through that agent's own extension points (hooks, plugins, skills, slash commands); one installer detects installed agents and configures them. Supported: Claude Code, Codex, Copilot CLI, Gemini CLI, OpenCode, Kiro, Droid, Amp, Pi, T3 Code and others; anything else can connect over MCP. Built with Bun and TypeScript; state lives locally under the user's home directory, with no telemetry beyond a release check.
Four surfaces:
Plan review. When an agent leaves plan mode, a permission hook fires. A local server reads the plan from the hook payload and opens a review page. The reviewer annotates and approves or denies; a denial returns structured feedback, and the revised plan is shown as a diff against the previous one.
Code review. Reviews uncommitted diffs and GitHub PRs or GitLab MRs (also GitButler, jj, Perforce). Line comments go back to the agent; AI reviewers can also comment on the diff.
Annotate. Mark up markdown, folders, fetched URLs, rendered HTML, or the agent's last message, and send the annotations back.
Inbox (new). One window where many agents leave messages and requests while they keep working. A reply comes back as the agent's next turn. On recent Claude Code versions the agent ends its turn instead of waiting, and the decision arrives later as a message.
Architecture
Capture at the agent's own seams. Plan review hooks the plan-mode exit permission request; there is no wrapper process or proxy around the model. Each agent integration is thin and native to that agent.
Local server plus browser UI. A small local HTTP server owns review state and serves single-file HTML review pages. Exports go to Obsidian, Bear or Notion; there is a VS Code extension and a terminal UI.
Inbox as a mailbox, not a lock. Requests are queued items with a return address (the session). The session is released immediately; delivery of the answer resumes it. This is the same move durable-execution runtimes make for long waits: persist the pending decision, free the worker, resume on signal.
Supply chain. Releases ship SHA-256 checksums, SLSA provenance and a CycloneDX SBOM, gated on a vulnerability scan; installer attestation checking is opt-in.
Why it matters
Human attention is the scarce resource in multi-agent work. With several agents running, blocking prompts become a queue the human services by tab-hopping. A single inbox batches that queue and makes its size visible.
Blocking holds cost. A session parked on a question keeps context, compute slots and sometimes locks. Ending the turn and resuming on reply avoids paying for idle waits and survives the human being away for hours.
Review on structured artifacts. Plans and diffs are reviewed as documents with inline annotations, and the agent gets structured feedback rather than a free-text chat reply. Plan diffs make each revision auditable.
Tradeoffs
Async answers arrive into changed state. If the agent keeps working after asking, the answer may refer to a plan or tree that has moved. Either the agent must stop on questions that gate the work, or each reply must carry the version it answers.
What may proceed without an answer is the agent's call. Non-blocking is safe for advisory questions; for approvals of risky actions the agent must still not act until the decision arrives. The tool moves the wait out of the session; it does not decide which waits are mandatory.
Batching delays urgent questions. One pass over a queue is efficient, but an agent stuck on a blocking question waits until the next pass. Priority or staleness signals matter as volume grows.
Integration surface is wide. Many agents, each wired through different hooks and plugins, means behavior can drift when any one agent changes its extension API.
Failure modes
An answer delivered to a session that has since been compacted, restarted or closed, landing without the context it refers to.
An agent treating "no answer yet" as permission and proceeding on a risky step.
A stale plan approved because the reviewer saw an earlier revision; mitigated by plan diffs, but only if the approval is bound to a specific revision.
Remote mode exposes the review server on a fixed port with no login; the project warns to keep it closed. Shared short links are encrypted and expire after 7 days, and are being replaced by a hosted product.
Review fatigue: an inbox that makes asking cheap invites agents to ask more, shifting load back onto the human.
Review checklist
Which agent requests are blocking (approval of an action) and which are advisory, and is that enforced, not just conventional?
Is every answer bound to the plan or diff revision it was given against?
What happens to an answer whose session no longer exists?
Can the inbox show age and priority, so urgent blocking questions are not stuck behind advisory ones?
Is remote or shared access disabled or authenticated?
Are approvals and denials logged with who decided and when, so the audit trail outlives the session?
Related
R034 AG-UI: interrupts as a protocol primitive between agent and user-facing app; Inbox is a local, cross-agent implementation of the same idea.
R049 A2UI 1.0: approval and question cards rendered as native UI components.
R035 Long-horizon agents and durable execution: persist the pending decision, release the worker, resume on signal.
R016 Do humans still need to check AI's facts?: route only disputes to a human, the scarce-attention argument behind batching review.