Agent Center

A self-hostable notification layer for AI agents. One feed for everything your agents did — and everything waiting on you.

Agents open pull requests, file tickets, send messages, run on schedules. What they lack is a decent way to tell you about it: pasting into a chat channel doesn't group, doesn't distinguish "FYI" from "I'm blocked", and buries the one thing that needs you under twelve that don't. This is a small server with one job — agents report to it, you read a clean feed and click through to where the work actually lives.

A notification layer only

It sits beside your agents, not around them. It does not run or orchestrate them, hold conversations, gate anything behind approvals, or replace a channel you already use. Traffic is one-way by design: agents write, you read and click out. If an agent needs an answer, its notification says where it's waiting — you answer it there, in its own channel, not here.

 MCP client ──▶  /mcp  ──────────────┐
                                     ├──▶  agent-center  ──▶  your browser
 cron / CI  ──▶  POST /api/v1/…  ────┘           │
                                        groups repeats, splits
                                        "activity" from "attention"

Four things a chat channel can't do:

• Grouping. Repeats sharing a group_key fold into one entry with a count — "raised 10 times over 19 days" is one row, not ten pings.
• Attention vs activity. "I opened a PR" and "I'm blocked and need you" are different kinds of message, and the feed knows the difference.
• Regression semantics. A grouped item that fires again after you've read it comes back unread. Archived stays archived — a re-fire opens a fresh entry.
• A link out, not a copy. Every card deep-links to the source app. This is the index of the work, not the destination.

Quickstart

pip install agent-center-app       # or pipx install agent-center-app
agent-center serve                 # → http://127.0.0.1:8765 (SQLite in ~/.agent-center/)

The package installs as agent-center-app — PyPI rejects names that differ from an existing project only by punctuation, and agentcenter was taken. The command you run, and everything else, is agent-center.

If 8765 is taken the server says so and refuses — pick another with agent-center serve --port 8766.

Register an agent (in another shell):

agent-center agent add my-agent
# prints a bearer token (shown once) and a ready-to-run curl

Connect a harness — Claude Code, for example:

claude mcp add --transport http agent-center http://127.0.0.1:8765/mcp \
  --scope user --header "Authorization: Bearer <token>"

Or click Connect an agent in the UI for copy-paste steps per client — including a single prompt to hand an agent that configures itself. Send one by hand:

curl -X POST http://127.0.0.1:8765/api/v1/notifications \
  -H "Authorization: Bearer <token>" \
  -H "Content-Type: application/json" \
  -d '{
    "type": "input.needed",
    "category": "attention",
    "priority": "high",
    "group_key": "PROJ-182",
    "title": "Which entity paid the Meridian invoice?",
    "body": "Inc or LTD — needed to issue the correct invoice.",
    "source": {"app": "linear", "link": "https://linear.app/acme/issue/PROJ-182"},
    "actions": [{"label": "Open in Linear", "url": "https://linear.app/acme/issue/PROJ-182"}]
  }'

Open http://127.0.0.1:8765 — it's in the feed. Fire the same group_key again and the card shows ×2 instead of duplicating.

Local mode is the default: loopback bind, SQLite, no password, UI open. The loopback bind is the security boundary; agent tokens are still required for writing, because they are identity (who sent this), not just auth.

To keep it running on a Mac — start at login, restart on crash — docs/launchd.plist is a ready LaunchAgent: fill in the two paths per its comments, copy it to ~/Library/LaunchAgents/, launchctl load it.

Hosting it

Run the server on something always-on and your browser and remote agents reach the same URL. Two rules, enforced rather than suggested:

• ADMIN_PASSWORD is required. On a non-loopback bind (or Vercel) the server refuses to start without it — and a passwordless server answers loopback peers only, whatever it ended up bound to. ALLOW_INSECURE_BIND=1 opts out behind a tailnet or authenticating proxy.
• TLS is the platform's job. Managed platforms terminate it; on a VPS put Caddy in front (see docker-compose.yml). Never expose plain HTTP.

Browsers sign in with the password. Browserless readers — a retention cron, the menu bar app — use viewer tokens (agent-center viewer add cron, prints once). Viewer tokens read and triage; agent tokens write; neither can manage agents — that takes the password session, so a stolen device token can't mint itself a write credential.

Render — one click: render.yaml provisions the service, a disk, and a generated ADMIN_PASSWORD (read it from Dashboard → Environment). Free-tier variant in the file's comments.

Docker: docker compose up -d — reads ADMIN_PASSWORD from .env (see .env.example), stores SQLite in a volume. One process, one worker: scale the box.

Vercel + Neon: api/index.py, vercel.json, and requirements.txt ship ready. Set DATABASE_URL (Neon's pooled string, rewritten postgresql+psycopg://…), ADMIN_PASSWORD, and CRON_SECRET (a viewer token) so the daily cron can call /api/v1/prune.

Connecting agents

The HTTP API is the only thing that writes; every other channel is a client of it. Remote MCP is mounted in the same process at POST /mcp — streamable HTTP, stateless, authenticated by the same per-agent bearer token — so connecting any harness is one URL and one header. Four tools:

Category is carried by the tool name rather than an enum field: choosing between two named tools is a decision a model actually makes; filling in a category field is one it defaults through. The full field rules live in one markdown file served at GET /api/v1/guide.md and as the MCP initialize instructions, so no harness keeps a copy that can go stale.

Why MCP alone isn't enough — the skill

An mcp_servers entry means an agent can notify, not that it will. Harnesses decide what to do by scanning their skill index before they act, and a server that exists only in a config file is invisible to that pass. So the install flow ends with the agent fetching GET /api/v1/skill.md and writing it into its own skill directory: MCP delivers the capability, the skill delivers the policy.

Its triggers are deliberately post-conditions on the agent's own work — a notification skill that fires only when you already thought to ask is the failure it exists to prevent:

triggers:
  - you created, shipped, merged, filed, or sent something the user would want to know about
  - you stopped because you need a decision, an answer, or a credential from the user
  - a scheduled, cron, or background run finished, succeeded, or failed

(It also tells agents not to satisfy "notify me" with osascript or notify-send — a desktop toast that nothing collects.)

The notification model

Retries: send an Idempotency-Key header and a re-delivery of the same fire counts once.

Configuration

API

Interactive docs at /docs. Errors are RFC 7807 application/problem+json; the feed is cursor-paginated. GET /api/v1/notifications filters on category, type, agent, source_app, priority, tag, unread, archived, and q; /notifications/facets returns the available values with counts — it's what the filter menus are built from. POST /api/v1/agents/{slug}/token rotates a credential (only hashes are stored, so tokens can never be re-shown).

Optional: the macOS menu bar app

clients/macos/ is a native menu bar client (pictured above) — one more reader of the same read API, for the feed and its banners without keeping a browser tab open. It authenticates with a viewer token, polls conditionally (an idle minute costs a handful of 304s), and reads, archives, and snoozes in place. macOS 14+, builds with the Command Line Tools alone:

cd clients/macos && make install

See clients/macos/README.md.

Status

Early, and deliberately narrow: one-way visibility is the whole scope, now and later. On the roadmap: SSE live updates (the UI polls every 30s today; /facets answers 304 when nothing changed), snooze/mute rules, per-harness lifecycle hooks so a run cannot finish silently, and Web Push for urgent. Not on the roadmap: replies, approvals, or anything that would make this a place work happens rather than a place work is reported.

Development

See CONTRIBUTING.md for scope, checks, and PR guidance.

python3 -m venv .venv && .venv/bin/pip install -e ".[dev]"
.venv/bin/pytest
.venv/bin/agent-center serve            # → http://127.0.0.1:8765

The UI is a React app in frontend/ (Vite, TypeScript, Tailwind, shadcn/ui, TanStack Query), built into agent_center/static/ so the wheel ships it and pip install users never touch npm:

cd frontend
pnpm install
pnpm dev        # dev server on :5173, proxies /api to :8765
pnpm build      # emits into agent_center/static/

MIT licensed.