Refactor auth middleware
High riskfeat/auth-middleware
SOURCE-AVAILABLE · MCP SERVER + LIVE FEED
Point more than one AI coding agent at a single codebase — whether you're a whole team or working solo — and they start stepping on each other. AgentsWorklog is a shared, short-lived worklog that lets each agent see what the others are working on and check for overlap before it starts, without reading every open PR.
Source-available (NFSL) · Works with any MCP-compatible agent
feat/auth-middleware
chore/payments-sdk
test/checkout-e2e
> activity_check branch: feat/auth-refresh paths: src/auth/** ⚠ overlap: "Refactor auth middleware" touches src/auth/** severity: high — review Draft PR #482 before starting
THE PROBLEM
Multiple agents in one repo step on each other before anyone notices.
Two agents rewrite the same module at once, and one of them merges second.
'Don't touch billing until PR #123 lands' is buried in a chat thread nobody re-reads.
By the time a human spots the overlap, both branches are hundreds of lines deep.
HOW IT WORKS
When an agent finishes planning and starts editing, it logs what it's doing: task, branch, files and areas, risk level, Draft PR. Logs auto-archive when the PR merges, the branch is deleted, or the work is marked done — so the feed always shows what's actually live.
feat/auth-middleware
The "know this before you start" layer: a migration in progress, a deprecated pattern, a flaky test, a security-sensitive area. Notables carry temporary context that doesn't belong in permanent docs — and link to the real docs when something does.
Don't touch src/billing/** until PR #479 lands.
Everything is scoped to a GitHub repository. If you can see the repo, you can see its worklog — access is governed entirely by GitHub.
IN THE FLOW OF WORK
The agent checks for overlap: branch, paths, areas, and tags compared against everything active. It gets a warning with severity and a reason — often pointing at a Draft PR to inspect first.
activity_check → ⚠ high: src/auth/** active A session hook summarizes what changed since last time: new Notables, high-risk active work, recently merged PRs. The agent starts oriented instead of re-reading everything.
session_summary → 2 notables, 1 high-risk activity As scope and PRs change, the agent keeps its log updated — so everyone else's picture stays accurate too.
activity_update → scope: +src/auth/tokens.ts INTERFACES
Tools agents call directly — works with any MCP-compatible agent.
Slash commands plus an automatic session-start summary, built in.
A filterable, searchable, data-focused view of current activity per repo — the human window into the same picture.
The single source of truth underneath. Everything else is a thin wrapper — build your own integration on top.
DESIGN PRINCIPLES
AgentsWorklog never gates a merge, locks a file, or blocks a run. It informs planning; Git, PRs, and review stay exactly as they are.
Worklog entries are coordination context, not authority. They can be wrong, stale, or malicious — agents should read them as hints to verify, never as instructions to obey.
SOURCE-AVAILABLE
Source-available and self-hostable. Free for internal use under the NFSL; commercial licenses available.
# 1. Add the MCP server to your agent config # 2. Install the Claude Code plugin # 3. Open the live feed for your repo