create_pinPin it
The agent writes it down while it still has the context: the files and lines, how it came up, the traps, and how risky, valuable and big the job is.
Your coding agents keep finding things that aren't today's job. A flaky test, a retry loop that never gives up, a TODO with teeth. Laterbase gives them one shared place to put it, with enough context that anyone can pick it up next week.
The problem
Halfway through a ticket, your agent spots something off next door, or you decide part of the job can wait. Do it now and your small PR isn't small. Mention it in chat and it scrolls away with the session.
session ended · 3 notes lost
How it works
Laterbase is an MCP server your agents already know how to talk to. They file out-of-scope finds and what you put off, notice when it's been found before, and when you ask for something to do, they start with the best payoff for the least work.
create_pinThe agent writes it down while it still has the context: the files and lines, how it came up, the traps, and how risky, valuable and big the job is.
add_sightingEvery new pin is checked against what's already there. Found it again? It becomes a sighting, not a duplicate. Four sightings is the backlog telling you something.
search_pinsGot a spare afternoon? Ask for something worthwhile. Pins are ranked by priority: payoff over effort, risk, and what each one unblocks. The agent claims it, fixes it and links the PR.
What's in a pin
A pin is the handover note you wish people left. Enough that someone with zero context, or a fresh agent, can read it in thirty seconds and decide.
Duplicate charges under load.
Fewer support tickets, calmer on-call.
An afternoon, tests included.
deliver()send()Stripe returns 409 when an event was already processed. Treat it as success, not a failure, or you'll double-apply refunds on the retry.
Should you pick it up? The verdict weighs risk, payoff, size of job and what it unblocks so you don't have to.
Repo-relative files, line ranges and symbols, pinned at a commit. No vague “somewhere in billing”.
Gotchas the finder learned the hard way, so the fixer doesn't step on the same ones.
What “done” looks like, written down before anyone starts. Agents love a checklist.
The small print, but good
When an agent starts a task it calls pins_near and hears what's already known about that corner of the code.
Everyone's agents pin into the same workspace, so the thing your teammate's agent found yesterday is already there today.
Agents rate risk, payoff and scope as low, medium or high. Leverage is worked out for them, never guessed.
Ask in plain English. Meaning and keywords are both matched, so “slow checkout” finds the N+1 in cart totals.
Pins are grouped by what they affect, like payments or search, not by where the file happens to live this week.
Claude.ai and Claude Code sign in with your account. Other agents take an API key, stored hashed and revocable in a click. Each connection knows which workspace it's in.
The Claude Code plugin
Install the plugin and Claude Code signs in to Laterbase and picks up 10 skills. Ask for a quick win and it finds the pin, fixes it and opens the PR. Wrap up a session and it pins what you left behind.
Plus /laterbase:triage, /laterbase:epic, /laterbase:random and /laterbase:roulette.Every skill, step by step
Connect an agent
Sign in and make a workspace. Many agents sign in when you connect them; the rest take an API key.
Run or paste the snippet for your agent. It's a remote server, so there's nothing to host.
Your agent pins what's out of scope and what you put off, as it goes.
claude plugin marketplace add --scope user \
https://laterbase.dev/claude-code/marketplace.json
claude plugin install --scope user laterbase@laterbaseget_contextsearch_pinspins_nearget_pincreate_pinadd_sightingupdate_pinset_statusupdate_areaTen tools, each described for the agent that reads it. You don't have to learn any of them.
Works with the agents you already use
The dashboard
See what's piling up, which areas are getting worse, and what's worth an afternoon. Claim a pin, hand it to an agent with one copied prompt, close it with the PR. It's optional. Most days your agents do the talking.
Questions
Anything that speaks MCP over HTTP. Claude.ai, Claude Code, Cursor, VS Code, Codex and Copilot CLI sign in with your account; Antigravity, Windsurf and the rest connect with an API key. See every client's setup.
Agents pin only real work: what's out of scope and what you put off. Every new pin is checked for duplicates before it's saved. Repeats become sightings, so noise turns into signal instead of clutter.
If an agent gets overexcited anyway, dismissing a pin takes one click.
It stays where it is. We never clone or scan your repository; we store what your agent writes into a pin: titles, summaries, file paths and any snippets it chooses to include. Pin text goes to OpenAI to make it searchable and to spot duplicates, and is never used to train models. If you install the GitHub app, it shares pull request and commit details so pins can follow the work. The privacy policy has the details.
No, it's the step before one. Most pins aren't worth a ticket yet, and some never will be. When one is, link the issue and carry on. Laterbase is happy being the place things wait.
Yes. A workspace can span as many repos as you like, and you can belong to several workspaces. One connection works in all of them: each call picks its workspace from the repo the agent is in.
Every session leaves something for later. Give it somewhere to go.