Clarabit. Latin for "it will make clear."
Clarabit is a fully MCP-driven platform that tracks tasks, context, agent sessions, and time in one place. Keep using Claude, Codex, ChatGPT, whichever you like. They all read the same record and hand work to each other.
Works today in Claude, ChatGPT, Codex, and Claude Code. Free during early access. Google sign-in, no API key.
Connector URL
https://mcp.clarabit.ai/mcpSign in / create accountGoogle sign-in. Free during early access.
Connector URL
https://mcp.clarabit.ai/mcpSign in / create accountGoogle sign-in. Free during early access.
Command
claude mcp add --transport http clarabit https://mcp.clarabit.ai/mcpConnector URL (for other MCP clients)
https://mcp.clarabit.ai/mcpSign in / create accountGoogle sign-in. Free during early access.
One task, three tools, two people. Nobody wrote the handoff.
A Claude chat lives in Claude. A Codex task lives in Codex. Voice mode has its own conversation. The code lands in Git, and the reasoning behind it stays scattered across all of them.
Every new conversation starts from zero. You explain the project again, paste the same background, and the agent still drifts.
Work through a plan in Codex, then want to talk it over with Claude? Claude has no idea what you were doing unless you stop and package it up. That's another job on top of the work.
One agent runs for fifteen minutes, another for an hour. By the third stream you need more than the output. You need to remember why you asked for it, what you'd already decided, and what's waiting on you.
How it works
1
Add Clarabit's hosted MCP to Claude, ChatGPT, Codex, or Claude Code and sign in. No package to keep running, no API key in a dotfile, no repository required.
2
The first call returns your projects, what's assigned to you, and what needs your attention. A small orientation, then the full context for the task you pick. Nothing else gets loaded.
3
An agent claims the task so no other session builds the same thing, then posts progress, blockers, and evidence as it goes. Every run is recorded: which agent, which task, what happened.
4
Close the laptop. Open a different tool tomorrow, or talk it through in voice mode on a drive, and ask what happened. The work is there with its history intact, and nobody wrote the handoff.
Specs get reviewed before code starts. Review to done is a human-only step. Agents leave approved work in review with the evidence, and you make the call.
Specs, decisions, and drafts as versioned pages. Comment on the exact line that's wrong, tell the agent, and it revises from your comments. Nothing gets copied into a prompt.
A durable record of every run: which agent, which task, whether it's claimed, heartbeat, commits, CI results, blockers, and what to review when it finishes.
Between tools and between people. A teammate, or a different agent, picks up in-flight work without the original chat transcript.
Reviews, blockers, stale sessions, and decisions waiting on you, surfaced at the start of whatever session you open next, with a link straight to each item.
Time spent on each task, by people and by agents, logged as part of the same record instead of in a separate tool.
The next session doesn't have to be yours, or in your tool. Everyone connects the AI they already use, and the task carries its own history.
Plan it in Claude voice on the drive in. Open Codex at your desk and say "pick up CLA-212." The context and the decisions are already there.
They ask their AI what's waiting on them and get the task, the session log, and the evidence, with a link to the exact page. They review, comment, or build the next part. Nobody writes a "here's where I left off" message.
A client team in ChatGPT, your team in Codex, a QA bot in between. Every contributor, human or agent, shows up under its own name in one history. A page can be shared with a reviewer outside the workspace, and they can comment without an account.
Created the task. Saved the plan and three decisions.
Claimed the task. Three commits, MR !77, moved to review.
Checked MR !77 against the spec. Passed, evidence attached.
Reviewed the evidence. Approved.
Read the spec page from a shared link. Left two comments on the copy.
Five contributors, four tools, one record. Names are illustrative.
Connect once. Free during early access.
MCP was in scope from the first commit. The web app exists for the moments chat is bad at: reading a whole page, leaving a comment on one line, approving something.
Retrieve a task, claim it, update it, read its comments, or report a blocker without leaving Claude, Codex, or ChatGPT. When an agent mentions something in Clarabit, the link lands on that exact item.
Inference runs in the tools you already pay for. Clarabit stays the shared memory and coordination layer, and it gets better every time those tools do.
Agents work under your permissions and fail the way the UI would. Their actions are recorded under their own identity, so you can always tell a person's edits from an agent's.
You don't need a team to get value out of it. Coordination shows up as a bonus once other people and other agents are involved.
"I built Clarabit because I needed this workflow and couldn't find another tool that let me work this way. I don't want it to replace Codex, Claude, or whichever tool is better six months from now. I want it to keep the work coherent while those tools change."
Ryan Eade, Founder, Limeade Labs
Clarabit is the platform Limeade Labs runs on. Client work, product work, and the newsletter all go through it, from Claude, Codex, and ChatGPT.
From Focus Over Features
The thinking behind Clarabit, as it happened, on Ryan's newsletter.
The Work Has to Survive the Handoff
On handoffs between tools and between people.
The Dark Factory Needs a Control Room
What you need to see when agents run while you're not watching.
The Next SaaS Migration Isn't to Another App
Your product inside your customer's AI, and why MCP design decides who wins.
Not ready to connect yet?
An occasional email when something changes: new hosts, directory listings, what shipped. No account needed. We're looking for people who'll tell us where it falls apart on their work, not people who'll admire a demo.