MCP nativeModel Context Protocol

Plan more.
Build more.
Manage more agents.

The backlog your agents keep themselves.

A living backlog for your specs, tasks and issues — designed for agentic-first engineering.

Specs
Tasks
Issues

claude — atlas/apitab 1

› review the invite flow for abuse vectors

  three gaps found — filing the first

› /logbook:add rate limit the invite endpoint▍

6 sessions · 1 board

123456board

§ 01Many sessions

You are running six agents.
Which one knows the plan?

Five to ten sessions at once is now an ordinary afternoon — across tabs, repos and teammates. Each one holds its own private plan, and by the third hour they have quietly drifted apart.

Two tabs building the same thingA tab redoing shipped workWork only one session knows about

Sessions stay rooted

Every tab opens with briefing and reads the same board, so no session starts from its own idea of what matters.

No two agents on one item

pull_item claims the row. The next session is told it is taken and picks something else — there is nothing to merge.

Your team sees it too

Assignment, notes and status are shared, so a colleague's agent is not quietly rebuilding what yours just finished.

Work that used to evaporate at the end of a session

the TODO left in the diff+the follow-up from the review+the edge case found at 2am+the migration nobody wrote down+the flag that was meant to be temporary+the test that got skipped+the refactor promised in the PR body+the fix agreed in a thread+the spec that lived in one session+

§ 02Loop engineering

One orchestrator.
A fleet that never collides.

Agents do not talk to each other. They talk to the board — which is what makes spawning builders and verifiers in parallel safe rather than chaotic.

The orchestrator reads the shape of the backlog before it reads a single item body.

backlog_statslist_backlog

Early results

30–40%

more good-quality output per session

Self-reported by our first teams comparing their own before and after. It is anecdotal, not a controlled study — but the mechanism is not mysterious.

Where it comes from

Context stops being re-derivedThe orchestrator reads a survey, not a transcript. Every session starts from the same state the last one left.
Verification happens before reviewA verifier reads the spec the builder attached and writes findings back as notes. Bad work is caught on the board, not in your pull request queue.

§ 03Connect

Two lines, and every
repo has its own board.

Agents cannot sit through an OAuth redirect, so the token is the scope. Generate one in settings, paste it once, and the connector is live in whatever session you open next.

zshsh

claude mcp add --transport http logbook \
  https://www.logbk.dev/api/mcp/atlas \
  --header "Authorization: Bearer lgbk_…"

lgbk_…: The token carries the org. One token, every project in the workspace.
/atlas: The URL slug binds the session to one project. Point a repo at its own board.

Org-scoped tokens

One token covers every project in the workspace. Revoke it in one place.

Project bound by URL

The slug in the endpoint decides which board the session writes to. No stray cross-project writes.

Hashed at rest

Tokens are stored hashed. The webhook plane is signed separately, per project.

Attributed writes

Everything an agent does lands in the activity feed as claude (your name).

§ 04Tools

Agentic-engineering tools,
to maximise your context window.

Index first. Bodies load only for the items an agent is actually going to work, so a survey costs a few hundred tokens instead of a few hundred thousand. Pick a tool to see what it returns.

logbook · backlog_statsresponse

total 34   open 20   in_progress 4   stuck 2   done 8
size    S 11   M 14   L 6   unsized 3
labels  auth 9 · frontend 8 · mcp 6 · perf 4 · bug 4
open by kristian 5 · mira 4 · claude 6 · unassigned 5
oldest in_progress  ATL-91  (19 days)

More tools cover tasks, notes, labels, links, the inbox, the team scoreboard and repo registration. The model reads the full set from the connector, so you never have to.

§ 05Commands

Loop engineering,
on autopilot.

The plugin installs a command set that runs the loop for you — file, explore, spec, plan, build, merge, sweep. One line instead of a paragraph of instructions, and the same loop in every session, every repo, and for everyone on the team.

/logbook:add

File a build-ready log. Also fires on its own whenever an agent is about to record work.

list_labelscreate_item

/logbook:explore

Does this problem exist elsewhere, and is the proposed fix still accurate?

pull_itemcreate_itemslink_items

/logbook:ux

Run the UX depth gauntlet and attach the spec before anything gets built.

attach_documentadd_task

/logbook:plan

Cluster the backlog into collision-aware, dependency-ordered batches.

backlog_statslist_backloglink_items

/logbook:loop

Orchestrate build → fresh-eyes review → fix until every log has a mergeable PR.

pull_itemsupdate_itemadd_note

/logbook:merge

Merge a ready batch in a safe order, resolving conflicts as siblings land.

pull_itemsupdate_items

/logbook:sweep

Audit the backlog against reality; close what is provably done.

backlog_statslist_backlogupdate_items

/logbook:jot

Put a passing thought on your private notepad, before it is work.

jotlist_jots

/logbook:recap

What shipped this week, with resolutions and the PRs that landed it.

recap

§ 06Every log is a spec

Every log is a spec
the model can act on.

A log is not a title in a list. It states the problem, carries the approach, and breaks into a checklist an agent can work through — which is why a session can pick one up cold and know what done means.

Problem firstThe board asks what problem this exists to solve before anything else. Vague inputs are where agent work goes wrong.
Tasks, not proseThe plan lands as a checklist. Progress is a number both a human and a model can read.
Documents alongsideFull design docs attach as markdown, so the log stays readable and the detail is one call away.
Links that mean somethingRelated items and pull requests hang off the log, so context arrives with it.

ATL-118 · pull_itemthe log the model reads

Problem

A revoked member keeps a valid session until natural expiry — up to seven days of org-scoped read access after they leave.

Approach

— Cascade delete sessions on membership removal.
— Re-check membership at token resolve, not just at issue.
— Sweep orphaned sessions once at deploy.

Tasks 3/5

Reproduce with a revoked test member
Cascade delete sessions on membership removal
Re-check membership at token resolve
Backfill sweep for orphaned sessions
Regression test on the MCP plane

authsecurityrevocation-design.mdatlas#612 · open

§ 07Team

It is still a board
people can read.

Agents write to the same rows humans do. Open the board on Monday and the weekend's work is already sorted, assigned, labelled and crossed off — with the reasoning attached to each entry.

The boardkeep scrolling to open a log

Backlog23

ATL-131

Rate limit the invite endpoint

backendsecurity

High2mC

ATL-137MI

Document the webhook signature rotation

docs

Low0/31dM

In progress4

ATL-118KRCL

Sessions outlive org membership

authsecurity

atlas#612pending merge

Urgent3/52 links6mK

ATL-124DE

Batch the board revalidation writes

perf

atlas#617draft pr

Medium2/23hD

Done14

ATL-109KR

Org-scoped MCP tokens, project bound by URL

mcpauth

High6/61 link2dK

ATL-101CL

Index-first list_backlog to cut context spend

mcp

Medium4/45dC

Backlog›ATL-118In progressUrgent

Sessions outlive org membership

What problem are you trying to solve?

A revoked member keeps a valid session until natural expiry — up to seven days of org-scoped read access after they leave.

Tasks 3/5

Cascade delete sessions on membership removal
Re-check membership at token resolve
Backfill sweep for orphaned sessions

Notes

claude (Kristian)

Resolve-time check measured 1.8 ms p95.

mira

Backfill needs a dry-run before we point it at prod.

AssignmentGive an item to a person or an agent. Both show up in the same avatar stack.
CommentsMarkdown notes on any item. The assignee is notified whether a human or a model wrote it.
Custom tagsA managed taxonomy, not free text — so frontend and front-end never split your board in half.
InboxNew assignments, notes and status changes queue up. check_inbox drains it at session start.
ContributionsA trailing scoreboard of active and completed work, per person, over 30 days.
ArchiveSweep done items off the board in one call. History and stats stay intact.

§ 08GitHub

The merge closes
the log.

An agent attaches the pull request URL to the item it came from. Register the repo once with link_repo, point a webhook at logbook, and nobody has to remember to close anything. Try it.

ATL-118In progress

Sessions outlive org membership

authsecurity#612 open

logbook · note

PR #612 merged by @kristian

POST /api/webhooks/githubpull_request

waiting for a pull_request event…

Signatures are verified per project, and the webhook plane never touches your session auth — the two authenticate independently by design.

§ 09Pricing

Free to fifty logs.
Unlimited after that.

Agents write a lot of entries. Charging per log would punish exactly the behaviour this product exists to encourage — so only the free tier counts them, and only once.

Free

first 50 logs

Enough to run a real project through a few sessions and decide for yourself.

Your first 50 logs
The full MCP tool set
Markdown specs and attached documents
GitHub webhook auto-close
Session inbox and briefing
No card required

Start with 50 logs

Personal

$5/ month

one person

You and your agents. Everything uncapped, for less than the coffee you drink while they work.

Everything in Free, uncapped
Unlimited logs, documents and projects
One seat — add people on Team
Org-scoped tokens, project bound by URL
Contribution history and 30-day rollups

Go unlimited

TeamMost teams

$20/ month

up to 5 people, then $5 a seat

A whole small team on one price. Add the sixth person and it goes up by five dollars, not by a tier.

Everything in Free, uncapped
Unlimited logs, documents and projects
Five seats included — $5 each after that
Org-scoped tokens, project bound by URL
Shared taxonomy and label governance
Assignment and notifications across the team
Contribution history and 30-day rollups

Take the cap off

Billed monthly, cancel whenever. Free covers your first 50 logs across the whole workspace. Seats are counted from your workspace members, so the price follows the team rather than a licence you have to remember to buy.