
notifyd
io.github.rmzlbPublished by GitHub account rmzlbv0.4.0Updated Nov 2, 2026
Operate a self-hosted notifyd instance: digest, jobs, retries, suppressions, projects, test sends.
Overview
Operate a self-hosted notifyd notification server over MCP: read the digest, inspect and retry jobs, manage projects, suppressions and test sends.
- What you can do
- Read the operator digest with ranked findings, queue, outcomes, latency and deliverability per project
- List and inspect jobs by project, channel, status, recipient or time, with provider, attempts and last error
- Retry or cancel stuck and wrong sends
- List and update projects: sender identity, channels, inbound rate limit and bulk send window
- List, add and release suppressions with all or marketing scope
- Read per-template metrics and send a test message on any channel
- When to use it
- For teams running their own notifyd instance who want an assistant to watch delivery health and act on it: triaging failed jobs, checking bounce rates, adjusting sender settings or proving a channel works. Useful when no admin dashboard exists and the on-call operator is an agent.
- Requirements
- A running notifyd instance reachable over HTTP, with PostgreSQL as its only dependency. The MCP endpoint is the instance's /mcp path and needs an Authorization bearer token: the admin key for full access or a read-only operator key for read tools only. The server itself needs DATABASE_URL, JWT_SECRET and ADMIN_API_KEY, plus at least one provider configuration to actually deliver messages.
Installation
In SourceWeft
- Open notifyd in the dashboard and add it to a workspace.
- Connect Authorization token credentials. They are stored privately and never shown on public pages.
- Enable the server for the chats that should use its tools.
Web executable via Streamable HTTP. Remote servers run from the web runtime once configured in a workspace.
Other MCP clients
Follow the launch instructions in the repository.
README
notifyd
Send email, SMS, WhatsApp, push and in-app notifications from one API call.
One 12 MB Rust binary, PostgreSQL only. Queues, retries, provider failover and quiet hours are handled for you, and an AI agent can run it over MCP.
[License] [Container image] [crates.io] [CI] [Image size] [Memory] [MCP registry] [Agent Skills]
Quick Start • Clients • Examples • Agent operations • API Reference • Architecture • Benchmarks • llms.txt • Contributing
What it is
Your app has to tell people things: a password reset, a shipped parcel, a newsletter, a red badge in the corner. notifyd is the small server that does all of it, so your code makes one call and never has to think about providers, rate limits, retries or time zones again.
- Small and fast. One 12 MB binary, a 14 MB image to pull, 13 MB of RAM idle. It accepts 44 000 notifications per second and drains 3 500 per second on a laptop (method). No Redis, no message broker, no dashboard to host: PostgreSQL is the only dependency.
- Nothing gets lost. A password reset always goes before a campaign. When a provider says "slow down", that channel pauses for exactly the time asked and resumes in priority order; when it fails, a second provider takes over. Retries, idempotency and quiet hours in each recipient's time zone are built in.
- Your providers, your data. Resend, Cloudflare Email, any SMTP, AgentMail, Telnyx, Twilio, smsmode, APNs, Web Push, FCM. Self-hosted, MIT.
- Operated by an API or an AI agent. No admin UI: a digest endpoint says what needs attention and what to do, and the same operations are MCP tools, so the person on call can be an agent.
Three instances run in production today, one per company, operated this way.
[60-second explainer: one send call, priority order under a provider 429, an agent operating the instance over MCP]
60 s explainer, invented data. MP4 · Remotion source
Send in one call
Flat REST, curl works; TypeScript and Python clients
exist. Retries are safe (idempotency_key), scheduling is a field
(scheduled_at), a marketing campaign is POST /v1/batch with thousands of
subscribers per call and it lands in the bulk lane so it never delays a
password reset. Follow any send with GET /v1/jobs/:id. Runnable examples in
every language: examples/.
Run it from a terminal
The same binary is the operator's CLI, against any instance (the server host
has ADMIN_API_KEY in its environment, so it just works there):
Let your agent run it
Most notification tools were designed for a human clicking through a dashboard. notifyd exposes the operator's job as tools, with the detail a human operator would need:
1. One call says what needs attention. GET /v1/admin/digest ranks
findings and tells you the action for each one, in JSON or Markdown:
2. The same operations as MCP tools. POST /mcp is a Streamable HTTP
MCP server (current spec revision, legacy initialize kept). Add it to Claude
Code, Claude Desktop, Cursor or your own agent:
Every tool carries readOnlyHint / destructiveHint annotations and an
outputSchema. A read-only operator key (READONLY_API_KEY) exposes only
the read tools, for an agent that reports but must not act. Every MCP call is
audited.
3. The digest comes to you. DIGEST_NOTIFY=telegram:<chat id> (or a
Slack / Discord webhook) and the findings above land in your chat when
something needs attention; notifyd digest --to telegram:… sends one now.
4. Everything an agent needs to integrate is in the repo. docs/llms.txt
is the whole API in plain text for a context window; three Agent Skills
ship in skills/ (notifyd-operate, notifyd-integrate,
notifyd-deploy):
Published on the official MCP registry as mcp-name: io.github.rmzlb/notifyd
(server.json). Full operator guide: docs/AGENT.md.
Features
Channels
- Email — Resend, Cloudflare Email Service, AgentMail, any SMTP (
lettre), attachments, per-project sender identity - SMS — Telnyx, Twilio or smsmode, swap with one variable
- RCS — smsmode (
BASIC/TEXTauto-selection, quick replies, click-to-call, rich cards and carousels) - WhatsApp — Telnyx
- Push — APNs natively (
.p8token auth, HTTP/2, badge, sound, thread id, silent pushes, dead tokens dropped), Web Push (VAPID), FCM - In-app inbox — REST + realtime SSE (
EventSource), read / archive / star, unread badge, multi-replica through PostgresNOTIFY
Delivery engine
- Priorities —
critical,normal,bulklanes;/v1/batchand campaign tags land inbulk - Pacing — token bucket per channel (
EMAIL_RATE_PER_SEC); a provider 429 pauses that channel forRetry-Afterwithout consuming an attempt, other channels keep flowing, and when it resumes the claim order putscriticalfirst - Retries — 30 s → 2 m → 10 m → 30 m → 2 h with jitter, 4xx fail fast, rejected batches fall back item by item
- Failover — second email provider with a circuit breaker (
EMAIL_FALLBACK_PROVIDER) - Send windows — quiet hours per project, evaluated in each subscriber's timezone
- Segments —
batchto "plan = pro and country in FR, BE" with a small filter vocabulary compiled to bound SQL; preview the count first - Scheduling, idempotency, stuck-job reaper, batch idempotency
Governance
- Unsubscribe — RFC 8058
List-Unsubscribeone-click on every marketing email, suppression scopesall/marketing - Topics and preferences — a
topicon any send ("tips", "billing"); subscribers opt out per topic and channel, honoured at enqueue; the unsubscribe page offers "this topic only" before "all marketing" - Multi-project — one instance, many projects, isolated by API key, key rotation with a grace period
- PII masking in logs, audit log of every mutation, per-project rate limit
Operations
- Digest, MCP server, Agent Skills,
llms.txt - Metrics —
/v1/metrics,/v1/metrics/prometheus, per-template metrics - Open and click tracking — own pixel and signed redirect links for every email provider, on marketing email by default (transactional links stay untouched),
opened_at/clicked_aton the job, off per project or per request - Webhooks — delivery events to your endpoints
- Workflows — event-triggered multi-step sequences, state in Postgres
- Templates —
{{variable}}substitution, stored per project
Quick Start
Docker (recommended)
The image reads its configuration from environment variables; no config file to mount.
Create a project and get its API key:
Prebuilt image, linux/amd64 and linux/arm64: ghcr.io/rmzlb/notifyd.
Binary, crate, Nix or source
Release binaries: Linux x86_64 and aarch64, macOS Intel and Apple Silicon.
NixOS: services.notifyd.enable = true; with an environmentFile (see
flake.nix).
No provider yet? EMAIL_PROVIDER=log prints emails instead of sending them.
Verify
→ Full setup, TOML alternative, production notes: docs/SETUP.md
API at a glance
Project endpoints take X-Api-Key: sk_<project>_…; operator endpoints take
the admin (or read-only) key, as X-Api-Key or Authorization: Bearer.
Inbox endpoints also accept a subscriber JWT.
→ Every endpoint with examples: docs/API.md, or feed docs/llms.txt to your agent.
vs. the alternatives
Every number in this table was measured by us; the Novu figures come from its own community docker-compose, idle, on the same machine and with the same tool as ours. Method, hardware and bias disclaimer in docs/BENCHMARKS.md.
Clients
Both clients cover the whole API (send, batch, jobs, subscribers, preferences, templates, workflows, suppressions, inbox) and raise a typed error on any non-2xx answer.
TypeScript / JavaScript — npm i github:rmzlb/notifyd until the notifyd-sdk package lands on npm this week; the package name and the import below do not change (Node 18+, browsers, edge runtimes; zero dependencies)
Python — pip install "git+https://github.com/rmzlb/notifyd.git#subdirectory=clients/python" until notifyd-sdk lands on PyPI this week (3.9+, sync and asyncio, httpx only) — clients/python
Any other language: the API is flat JSON over HTTP and docs/llms.txt is the whole contract on one page.
Workflows
Multi-step sequences triggered by an event, state in Postgres, survive restarts. Steps run in order; a condition jumps to a step index.
Step types: send, delay, condition (inbox.is_read or payload.<key>)
and digest (collect events for a while, then send one message). Full
script: examples/welcome-series.sh.
Configuration
Environment variables are the primary interface (that is what the image and
the compose file use); a notifyd.toml is accepted for local development.
Required: DATABASE_URL, JWT_SECRET, ADMIN_API_KEY. Then one provider:
→ Every variable, per provider: docs/CONNECTORS.md and docker-compose.yml. TOML reference: notifyd.toml.example.
Architecture
~14 000 lines of Rust, no unsafe. 12 MB binary (16 MB static musl in the image), 14 MB image to pull and 31 MB on disk, 13 MB RSS
idle, 23 MB while draining 100 000 jobs. → docs/ARCHITECTURE.md
Status
notifyd is a 0.x running in production for three companies. Not done yet: Swift and Kotlin packages; no dashboard, A/B testing or inbound email by design. Order and sizes in docs/ROADMAP.md. Breaking changes are announced in release notes; the queue schema is migrated automatically.
Documentation
Contributing
Issues and pull requests are welcome; good first issue is the place to
start, and big features begin with an issue. Read CONTRIBUTING.md.
License
MIT.
Built with 🦀 in Grenoble, France 🏔️
Source: README.md at commit 3239a54
Tools
0Version history
1- v0.4.0LatestSep 16, 2026
Score
60% of the applicable checks met
How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.
- Reached in the last 30 daysRequired25 pts
Keep the endpoint answering the MCP handshake; SourceWeft probes it every 30 days.
- Auth declaredRequired10 pts
- Tools listed15 pts
Answer tools/list without a credential, or ship a package that starts in a sandbox without one.
- DescriptionRequired15 pts
- Logo10 pts
- Documentation15 pts
- License10 pts