New task launches, Attention, and self-serve workspaces

September 20, 2026

This release helps you spend less time preparing task prompts and checking where work needs your input. New task launches give you a starting prompt, Attention brings requests and blockers together, and project templates give your AI apps shared instructions to work from. You can also invite others to try Clarabit in their own workspace with a signup link.

New: get from a task to working on it faster

When you’re ready to work on a task, you no longer have to write the opening request from scratch. The new launch actions prepare a prompt for that task that you can open in a supported AI app or copy into an existing conversation. Review it, add anything you want, and send it when you’re ready.

Use it to start an article outline, investigate a question, plan your week, or work on code. Writing, research, and planning projects can use the same flow without a code repository.

Launch controls on a coding task. Choose an app or copy the prompt into your conversation.

New: see where your input can move work forward

When you’re checking in on several pieces of work, start with the new Attention view. It brings review requests, reported blockers, and sessions that appear to have stopped into one place, so you can find what needs a decision without opening each task to check.

Each item explains why it is there and links to the task or session. Clarabit distinguishes a reported request for help from detected inactivity, helping you decide whether to review a result, answer a question, or check on work that may have stopped.

New: give each project a useful starting point

Project instruction templates reduce the blank-page work of explaining how you want an AI app to help. Choose a starting set of shared instructions for the kind of work you’re doing; connected AI apps can read them when they load the project, giving them a common starting point across conversations.

Templates are chosen per project. Your writing project can start with different instructions from your coding project, so you can organize different kinds of work without forcing them to use the same setup.

New: an easier way to introduce others to Clarabit

If you’ve wanted to share how you use Clarabit with a colleague or friend, you can now send them this signup link and let them try it in their own workspace. They can sign in with Google and get started without waiting for us to create an account for them.

Guided setup walks them through creating a project, choosing a template, and connecting their AI app, so you don’t have to walk them through every setup step yourself.

Guided setup lets new users choose the AI app they want to connect first.

To collaborate in your existing workspace, continue using teammate invitations.

Session tracking: new capabilities and reliability fixes

• Hosted session reporting: AI apps using the hosted MCP connector can now report session progress.
• Clearer identity: session details distinguish the person or account behind the work from the AI runtime being used.
• More reliable activity tracking: active sessions and task claims stay current as your AI app works with Clarabit, and stopped sessions are detected more consistently.

Fixes to existing features

• Page history: the latest saved version is now included and retrievable. Change summaries on new edits are attached to the version they describe.
• Task links: links returned to AI apps use the correct Clarabit web address, and task context includes linked merge requests.
• Release timestamps: corrected time handling for release-note publication.

Connector update

Both the cloud connector and local MCP package are now on 2.7.0. See the MCP changelog for the local/cloud comparison and upgrade instructions.

Connection guides · Send feedback

This follow-up release completes the task-identifier rollout for agent integrations and includes a small package-security refresh.

🤖 Use recognizable task identifiers from MCP tools

Clarabit MCP tools now accept the same human-friendly task references used in the app. Agents can use canonical references such as CLA-288, lowercase variants such as cla-288, or hyphenless variants such as CLA288 wherever a task reference is supported.

Numeric database IDs remain compatible, and responses continue to return the canonical uppercase, hyphenated identifier.

🔖 See the correct MCP version

The MCP server now reports its installed package version during initialization instead of returning a stale hardcoded value. This makes diagnostics and client compatibility checks clearer.

📄 Keep page-version history lightweight

Listing a page’s version history now returns metadata only. Full historical content is fetched only when an agent asks for a specific version, reducing payload and token usage for long-lived pages.

🛡️ Dependency maintenance

MCP package 2.6.0 also includes current production dependency security fixes.

No action is required for the Clarabit web app. Agent installations that pin package versions should update to @limeadelabs/clarabit-mcp@2.6.0.

This release makes work easier to reference, projects easier to navigate, and agent instructions easier to keep current. Here’s what’s new.

🔖 Give every task a durable, recognizable identifier

Tasks now use readable identifiers such as CLA-27 or COM-42 instead of a global database number.

An identifier series belongs to your workspace and can be shared by several projects. That means teams that use projects for initiatives, epics, or sprints can keep one continuous product sequence across all of them.

You can now:

• search by an identifier with or without its hyphen;
• copy a task’s identifier or canonical link from the task page;
• follow consistent task links from boards, dashboards, agent sessions, notifications, and reports; and
• choose an existing series or create a new one when you create a project.

Existing numeric links and supported integrations remain compatible, so saved bookmarks and established workflows continue to work.

🧭 Move through a project without losing your place

Project pages now share one persistent navigation shell. Tasks, Pages, Agent Context, Environments, Time Reports, and Settings stay in a consistent project section instead of competing for space in a crowded action bar.

The new layout keeps project context visible, makes management actions easier to find, and adapts cleanly across desktop, tablet, and narrow screens.

🤖 Keep agent instructions current in one place

Agent Context is now the durable source for project and workspace instructions.

Clarabit assembles those instructions in a consistent order for agent bootstrap and task context, while Project Settings provides a compiled preview rather than a second editable copy. Existing project instructions are carried forward, and current integrations remain compatible.

This reduces drift between repositories and Clarabit, and gives Codex, Claude Code, Cursor, OpenClaw, and other supported harnesses the same starting context.

🏷️ Set up the right kind of project from the start

The New Project form now lets you choose between:

• Internal projects, which track time without billing controls; and
• Consulting projects, which add billable flags and billable/non-billable reporting.

Internal remains the default. You can also see what each choice changes before creating the project.

✨ A few thoughtful improvements

• Form fields now have consistent, comfortable spacing throughout Clarabit.
• Project actions remain readable and usable at narrower widths.
• Task links, references, and search results consistently return the canonical uppercase, hyphenated identifier.
• Identifier keys use clear letter-only mnemonics and preserve historical references if a key is renamed.

This release helps your team keep project knowledge current, review important decisions with confidence, and give AI helpers the right access. Here's what's new.

✅ Review the active spec right from the page

When a page is attached as a task's active spec, you can now see its review status from both the task and the page.

Reviewers can approve the spec or request changes without leaving the document, so feedback stays close to the work. Clarabit also keeps a clear record of who reviewed it, when they reviewed it, and which version was approved.

AI helpers can ask for a review, while the final decision stays with a person.

🗄️ Tidy up old pages without losing them

Have a page that's no longer current? You can now archive it instead of deleting it. Add a quick reason and, if helpful, point people to a newer page.

Archived pages:

• remain available by direct link;
• stay out of normal search and AI context;
• guide readers to a replacement page when one is provided; and
• can be restored whenever you need them.

Archived projects also stay out of your everyday dashboards and summaries while remaining available when you need to look back.

🔐 Give an AI helper access to just one project

You can now create credentials for an entire workspace or limit them to one project. Project access is a simple way to give an AI helper only the information it needs.

For added peace of mind, a new secret is shown only once when you create it. After that, Clarabit keeps a short, safe prefix to help you recognize the credential—not the reusable secret itself. Your existing credentials will keep working, and you can revoke access at any time.

📎 See uploaded files where the conversation happens

Attachments now appear right on the task page. Images include an inline preview, and other files include helpful download details.

Files added to a comment appear directly beneath that comment, making it easier to keep screenshots, documents, and feedback together.

🛠️ A few helpful fixes

• Revoking a pending workspace invitation now works reliably.
• Archived projects and pages no longer add outdated information to dashboards or AI context.
• Credential security is stronger, with no need to reconnect your existing integrations.

🤖 Using Clarabit with AI helpers?

Project-specific credentials work with supported AI clients including Codex, Claude Code, and Cursor. You'll see setup instructions as soon as you create a credential, while the secret is still visible.

If your Clarabit MCP integration is pinned to a version, update to @limeadelabs/clarabit-mcp@2.5.0 or later for the latest page archiving and attachment workflows.

Six improvements you can use today.

✨ Attach a spec page to a task

When editing a task, pick a project page as its canonical spec from a dropdown. The selected spec shows up on the task view with a direct link — so anyone reading the task can jump straight to the design doc, decision log, or whatever else informs the work. Clear the field to detach.

Agents can do the same via cla_update_task(task_id, spec_page_id) in clarabit-mcp v2.4.0+ (null clears the attachment).

🔁 Resend a workspace invitation

Owners can now resend a pending invitation with one click — no more revoke-and-re-invite dance. The token rotates (the old email link stops working), the expiry resets to 7 days, and a fresh email goes out. Look for the new Resend action next to Revoke in workspace settings → Invitations.

🩹 Anchor staleness banner is honest now

Page comments anchored to a substring used to show the "This text has been edited or removed" banner on perfectly intact pages — false alarm caused by missing version data on older comments. Fixed: the banner now only appears when the page has genuinely drifted past the comment's anchor (the page version moved AND the anchor text is no longer in the body).

🎨 Mermaid diagrams render unicode correctly

—, →, •, and other multi-byte characters in mermaid node labels no longer turn into mojibake (â€", •). Drop them straight into your architecture diagrams — no — / → / • workarounds needed.

🖼️ Agents can upload screenshots inline (MCP)

@limeadelabs/clarabit-mcp@2.4.0 ships three new upload tools:

• cla_upload_image — upload PNG/JPG/GIF/WebP (up to 5 MB) and get back a markdown snippet ready to paste into a comment body
• cla_upload_task_attachment — task-scoped attachments (up to 10 MB; PDF and text also accepted)
• cla_upload_comment_attachment — comment-scoped attachments

The workhorse is cla_upload_image. Agents can now take a Playwright screenshot, upload it, and post a comment with the image embedded inline — all in one round trip.

🏷️ Agents can manage labels (MCP)

Also in clarabit-mcp v2.4.0: cla_list_labels, cla_add_label_to_task, and cla_remove_label_from_task. Attach by label name (auto-creates within the task's project if missing) or by label id. Agents can drive label-based workflows — UAT needed, blocked, needs-spec — without bouncing through the web UI.

---

Upgrading the MCP: the new tools live in @limeadelabs/clarabit-mcp@2.4.0. If your install is pinned to a specific version, bump to 2.4.0 (or latest) and restart your editor's MCP host to pick up the new tools.

Release v2026.05.17

---

✨ What's new in Clarabit

Share pages with anyone — they don't need a Clarabit account. Mint a public link from any page and send it to a teammate, customer, or stakeholder. They review the page and leave comments; their feedback shows up in your normal page-comment view with an "External" badge. They verify their email once via a magic-link; after that, follow-up comments post instantly. Revoke any link from the page header whenever you're done.

In-product release notes. This page. Click "🎉 What's new" in the top nav (or the green dot when there's something new) and a drawer slides in with the latest updates. Also public at clarabit.ai/changelog.

🤖 What's new for your AI helper

If you use Clarabit through an AI helper like Claude Cowork, Claude Code, or Codex, today's update teaches it three new things.

Leave a comment on a specific phrase in a page. Ask your AI to "leave a comment on the third bullet under Risks" and it'll land exactly there. If the phrase shows up in more than one spot, your AI will surface the candidates and ask which one you meant — no more silently anchoring to the wrong place.

Read the existing discussion on a page. Ask "what's still open on this spec?" or "summarize the comments on the launch plan" and your AI can pull up the unresolved feedback (or all of it) without you copy-pasting.

Save shared context for the team. When you tell your AI about a project convention or constraint — "we deploy on Tuesdays, never Fridays" — it can now save that into Clarabit so the next person, or your next AI session, picks it up automatically.

🔄 Heads-up: these AI features only show up after your AI helper picks up the latest version of the Clarabit MCP connector. Most tools refresh automatically when you restart them; a few need a manual "Update tools" nudge in their settings. Technical users, the exact commands are in For agents and integrators below.

🐛 Fixed

• Pages with comments from removed teammates no longer crash.
• Agent comments stay clean — internal AI-tool tags (the "<system-reminder>…" kind) get stripped automatically before the comment is saved.
• Adding a link to a task works with either shape the API accepts — the older "flat" payload no longer 500s.

---

For agents and integrators

The rest of this changelog is for folks driving Clarabit through MCP, an API integration, or building on top of the platform. If that's not you, you can stop reading here.

⚠️ Update your Clarabit MCP to @limeadelabs/clarabit-mcp@2.3.0

The AI-helper features above (anchored comments, list discussion, save context) are MCP tools. They won't appear in your tool list until you upgrade — existing tools keep working on older versions, the new ones simply aren't registered there.

Requires @limeadelabs/clarabit-mcp@2.3.0:

• cla_context_create — write new context entries (added in 2.1.0)
• cla_list_page_comments — list comments on a page (2.2.0)
• cla_update_page verbose flag — opt back into the full body in the response (2.2.0)
• cla_prepare_page_comment_anchor — validated anchored-comment offsets (2.3.0)

Upgrade:

# Global install
npm install -g @limeadelabs/clarabit-mcp@latest

# Local install (in a project that pins the dep)
npm install @limeadelabs/clarabit-mcp@latest

Then restart your MCP client (Claude Code, Codex, etc.) so the new tool list is picked up at session start.

New MCP tools and API endpoints

cla_prepare_page_comment_anchor — validates and computes offsets for an anchored page comment without you manually counting characters. Pass one locator: exact_text (unique substring), prefix + text + suffix (disambiguate repeated text), or heading + text (scope to a markdown section). Returns offsets + the body revision the anchor was computed against, ready to drop into cla_create_page_comment. Repeats no longer get silently anchored to the first match — the tool returns a structured ambiguous_anchor error with all candidate positions and their preview context.

cla_list_page_comments — list comments on a page (optionally filtered by resolved state) without first walking projects to find it. Backed by a new flat GET /api/v1/pages/:id/comments route. Pages also now include a comments_count so you can see at a glance whether there's anything worth fetching.

cla_context_create — complements the existing read tools so agents can write new context entries directly, not just edit ones the team already authored.

GET /api/v1/tasks/:id — rounds out the CRUD on the HTTP side; cla_get_task in MCP worked already, this is just for symmetry.

Behavior changes

• PATCH /api/v1/pages/:id returns a minimal payload by default — id, title, version number, change summary. Pass verbose: true if you actually want the new body back. Cuts the round-trip in half on synthesis-style pages. cla_update_page accepts the same flag.
• Page JSON now includes version_number — the revision the current body lives at. Anchor helpers stamp this onto comments so drift is detectable later. Note the semantic: the latest snapshot's version_number stores the body that was current before the most recent edit; the current body is logically one revision past that.
• Page comment author payload is nil-safe — external (share-link) authors are surfaced as { id: null, name, email, external: true } instead of the older { id, name } shape that assumed a logged-in user.
• Task-link API accepts both shapes — { url, title } and { link: { url, title } } (the flat shape used to 500).

---