Synaplan

We produce AI freedom.

Open-source, self-hosted AI platform (PHP/Symfony): chat with any model, RAG over your docs, embeddable widget, MCP — see the cost of every answer. Apache-2.0.

Website  ·  Docs  ·  Live instance  ·  iOS  ·  Android  ·  Desktop  ·  Outlook Add-in  ·  Discord

Quickstart

git clone https://github.com/metadist/synaplan.git
cd synaplan
make up

Open http://localhost:5173. Log in as admin@synaplan.com / admin123 and paste one provider key. Free: Groq.

A published image, a Docker GUI, a closed network, or a Linux server: More ways to install.

Related MCP server: laravel-simple-rag

How it compares

People usually arrive comparing Synaplan with a chat window or an app builder. This is what each project is for. It is not a scoreboard.

If you only want a chat window, Open WebUI or LibreChat gets you there faster.

Why Synaplan

• Freedom, not a plan tier. The whole platform — backend, frontend, widgets, plugins — is Apache-2.0 and starts with one command. Self-hosted is the same software as the cloud.

• Hundreds of models, one place. OpenAI, Anthropic, Google Gemini, Groq, Mistral, xAI, HuggingFace, sovereign EU providers, and any local model via Ollama. Swap a provider per task in the UI.

• The planner spends the expensive model only where it matters. A request becomes a small task graph (extract → summarize → generate → reply). Live cards show the steps. Every answer shows what it cost.

• Yours, including offline. On-prem, EU cloud, or fully air-gapped. No training on your data. No forced telemetry.

• Where you already work. Web, iPhone, Android, Desktop, Outlook, a chat widget, WhatsApp, email — plus Microsoft 365, Dropbox, Nextcloud / ownCloud, calendars, Jira, Confluence and OpenCloud.

• Extend it without a fork. Plugins, an OpenAPI REST API, an MCP server and client, and an Anthropic-compatible endpoint for Claude Code and friends. Optional sidecars stay off until you turn them on.

Take the tour

Click any screenshot to see it full size.

Regenerate these assets after a UI change with scripts/build-readme-tour.sh.

More ways to install

The quickstart is a local checkout. The variants below are a published image, a machine with no route to the public internet, and a Linux server.

On this machine

The quickstart is this path. Open http://localhost:5173. A status screen is up within seconds and lists every boot step, then switches to the app when it is ready. First start: 5–15 minutes. Later starts: seconds.

make up answers on :5173 before the rest of the images finish pulling. docker compose up -d starts the same stack; :5173 stays quiet until those pulls finish.

1. Log in as admin@synaplan.com / admin123. The status screen shows this too.

2. Paste one provider key. Free: Groq. The app opens AI provider setup until a key is in place. You do not edit a config file for that.

3. Same notes in the terminal: docker compose logs -f startup-notes.

No cloud key? COMPOSE_PROFILES=local-ai ENABLE_LOCAL_GPT_OSS=true make up pulls a local chat model (gpt-oss:20b, ~14 GB). Chat starts when the download finishes.

Published image, no git checkout

One file is enough. A .env beside it is optional.

mkdir synaplan && cd synaplan
curl -fsSL -o compose.yaml https://raw.githubusercontent.com/metadist/synaplan/main/deploy/quickstart/compose.yaml
docker compose up -d

Open http://127.0.0.1:8000. The first visit opens the setup: create the administrator, paste one provider key (free: Groq), done. The first start pulls the images and takes a few minutes.

Docker GUI (Portainer, Dockge, Synology Container Manager, …): create a stack, paste deploy/quickstart/compose.yaml as its compose file, deploy. Any setting from deploy/quickstart/.env.example goes into the stack's environment variables — nothing is required. When the GUI runs on another machine (a NAS or a server), add SYNAPLAN_HTTP_BIND=0.0.0.0 and SYNAPLAN_URL=http://<that machine's IP>:8000; without them Synaplan only opens on that machine itself. Dockge marks the stack "exited" once the one-time secrets-init step has finished; that is expected, and Synaplan keeps running.

Settings go into .env, then docker compose up -d again (in a GUI: the stack's variables, then redeploy):

• SYNAPLAN_VERSION picks a release tag from the releases page. Empty means the release this file was published with. Never latest.

• SYNAPLAN_HTTP_BIND=0.0.0.0 plus SYNAPLAN_URL=http://<this machine's IP>:8000 opens it to your network. SYNAPLAN_URL must be the address people type, or live chat stays disconnected. Over plain http:// on a network address, browsers block the microphone and copying to the clipboard; chat and files work. For those two, use Closed network HTTPS.

• Provider keys, a bootstrap administrator and mail are listed in the example with one line each.

Data lives in named Docker volumes — the database, uploads, vectors and the generated secrets. Keep them together: a database without its secrets volume does not start. docker compose down keeps everything; docker compose down -v deletes it all.

Change the release: back up the volumes, set SYNAPLAN_VERSION, run docker compose up -d (or redeploy the stack in the GUI). Going back to an older tag after a newer one migrated the database means restoring that backup. The steps are in Update a self-hosted deployment.

Office documents, local AI, file work, closed-network TLS and scripted backups use the full deploy/compose.yaml with deploy/selfhost.env.example instead.

Closed network

Opt-in. The default stays http://127.0.0.1:8000 on this machine.

Turn it on when the network has no route to the public internet and people open Synaplan by the machine's address. Chat needs https://<address>/. The machine creates the certificate. The browser warns once.

Any IPv4 address on that network works, including 10.0.0.0/8, 172.16.0.0/12, 192.168.0.0/16, 100.64.0.0/10 and 169.254.0.0/16, plus a block you assigned and do not announce.

cp deploy/selfhost.env.example deploy/.env
# Set SYNAPLAN_VERSION. Leave SYNAPLAN_HTTP_BIND at 127.0.0.1.
deploy/scripts/local-tls.sh 10.0.0.15
deploy/scripts/prepare.sh
docker compose --env-file deploy/.env -f deploy/compose.yaml pull
deploy/scripts/validate-release.sh
docker compose --env-file deploy/.env -f deploy/compose.yaml up -d

Colleagues open https://10.0.0.15/. Ports 80 and 443 must be free. The app itself stays on 127.0.0.1:8000. A second interface is another argument: deploy/scripts/local-tls.sh 10.0.0.15 192.168.1.20. A public name keeps your own HTTPS proxy and leaves this off.

Details: docs.synaplan.com/local-network · deploy/README.md.

On a server

The commands above are the development stack (source build, Vite, MailHog, phpMyAdmin) or the published image. For a Linux server, the installer writes deploy/.env, pins the latest release, and runs prepare → pull → validate → start → smoke-test. Secrets land in deploy/data/secrets.env.

curl -fsSL https://raw.githubusercontent.com/metadist/synaplan/main/install.sh | \
  bash -s -- --mode server --domain https://ai.example.com

The same steps by hand are in Installation and deploy/README.md. Flags: bash install.sh --help.

One-liner for the development stack, if you would rather not clone by hand:

curl -fsSL https://raw.githubusercontent.com/metadist/synaplan/main/install.sh | bash

One AI, everywhere you work

The same assistant, the same knowledge base, the same model policy — on every channel your team already uses. Connect a system once under Channels; the planner can then read from it and deliver results into it.

Conversation surfaces

Connected systems

Set these up under Manage → Connections (or Manage → Connections → MCP Servers / Manage → Channels → Email). In chat, use the channel word shown as a pill on the Connections page — for example nextcloud, dropbox, outlook.

Details and channel words: docs/CONNECTIONS.md.

The Synaplan ecosystem

Everything below is the same platform, packaged for different homes. Pick what fits — nothing else is required.

Prerequisites

• Docker + Docker Compose v2 (Docker Desktop on macOS/Windows, or Docker Engine + the Compose plugin on Linux)

• Git (the one-line installer also works with curl + tar when git is missing)

• 8 GB RAM minimum (16 GB recommended once you add the local-ai profile)

• ~4 GB free disk for the standard install (includes file work + spoken answers; +~1 GB for local-ai, +~14 GB if you also enable the local chat model)

• A few host ports must be free. The list, and how to change one without editing the Compose file, is in Installation — Port conflicts.

Apple Silicon (M1–M4) Macs — build the backend image, don't pull it. The three-step start above already does this: make up builds the backend and worker locally from a multi-arch base image, so PHP/FrankenPHP runs natively on arm64 with no emulation tax. That is by far the fastest setup, and it is the default — you don't have to do anything special. (The published ghcr.io/metadist/synaplan image is multi-arch too, so pulling it also runs natively.) The first local build takes a few minutes; every later start is a cache hit. Two optional dev tools (phpMyAdmin, MailHog) are still amd64-only upstream images — if you keep them, enable Docker Desktop → Settings → General → "Use Rosetta for x86/amd64 emulation on Apple Silicon" (macOS 13+) so those two emulate quickly.

Install Options

A plain docker compose up -d starts the same stack but :5173 stays silent until every image is pulled — make up exists so the status page answers within seconds.

No AI weights are downloaded by default, so the first boot is dominated by the Docker images and npm ci. COMPOSE_PROFILES=local-ai is the same switch a self-hosted install uses in deploy/.env, and it pulls the local embedding model (bge-m3, ~1 GB) in the background for RAG and semantic search; progress is shown in the app.

Prefer the shell to the UI for provider keys? Keys in backend/.env still work — the backend reads that file when the container starts and imports the key into the encrypted store on first use. Write the key before starting, or restart the containers afterwards:

echo "GROQ_API_KEY=your_key" >> backend/.env
make up
# already running? pick up the new key with:
# docker compose restart backend worker

Access

Open http://localhost:5173. The API is at http://localhost:8000, the API docs at http://localhost:8000/api/doc. phpMyAdmin, MailHog, Qdrant, Tika, Ollama and TTS: Installation — Development service URLs.

Default Login Credentials:

Features

• AI Chat — Ollama, OpenAI, Anthropic, Gemini, Groq, Mistral, xAI, TrustedTokens (DE), A2Agent (CN), HuggingFace (provider list)

• Self-aware assistant — Ask "What can you do here?" or type /help; the AI assistant answers from this installation's live capabilities, not a generic brochure

• Multi-Task DAG Routing — An AI planner decomposes complex requests into a directed task graph (extract → summarize → generate → reply), routes each step to the model that fits it, and streams live task cards while the steps execute — cheaper models for simple steps means fewer wasted tokens

• RAG Search — Semantic document search with MariaDB VECTOR or Qdrant

• Chat Widget — Embed on any website (widget guide)

• Mobile Apps — Chat, documents and voice on iPhone and Android, pointed at web.synaplan.com or at your own server (App Store · Google Play)

• Synaplan Desktop — Pair a computer and run skills on it. Open Channels → Synaplan Desktop. No installer yet (synaplan-desktop)

• AI assistants — Saved recipes (instructions, knowledge folders, tools, triggers) that you publish in versions (assistants)

• Tools & approvals — One tool registry; write-class actions pause under Approvals (tools)

• Users & groups — Share folders, chats, assistants and tasks; Operate → Users (people)

• File work — Short Python or Node runs on copies of files you picked, in an isolated sidecar (guide)

• Live Support — Realtime WebSocket layer (Centrifugo + Redis): human takeover of widget chats, typing indicators, operator notifications (realtime guide)

• WhatsApp — Meta Business API integration

• Email — AI-powered email responses, plus live mailbox search (IMAP and Microsoft 365)

• Connections — Microsoft 365, Dropbox, Nextcloud / ownCloud / WebDAV, CalDAV — read mail, file results, write calendar events (connections guide)

• Saved Tasks — Pin a multi-step plan and run it on demand or on a schedule, with a Steps editor and webhook trigger (Manage → Automations → Saved Tasks)

• Watched pages — Save a URL, compare it on a schedule, and mail the diff when it changes

• Audio — Whisper transcription (input) + optional synaplan-tts (output; five baked voices, UI language selects the voice)

• Documents — PDF, Word, Excel, images with OCR; the assistant can also build and revise Office files; optional Collabora CODE sidecar for thumbnails, PDF export, preview and combine (office documents)

• AI Memories — User profiling with Qdrant vector search

• Feedback System — Feedback capture and analysis powered by Qdrant

• Plugins — Non-invasive plugin system (plugin guide)

• MCP Server (early access) — Connect AI clients (Claude, Cursor, …) over the Model Context Protocol; your RAG and memories become tools at POST /mcp (MCP guide)

• MCP Client (early access) — Connect your MCP servers (Jira, Confluence, CRM, wiki, n8n, …) under Manage → Connections → MCP Servers. The planner pulls live data via mcp_fetch and, when you enable allow write actions on that server, can create tickets or pages via mcp_action — destructive tools stay refused. SSRF-guarded, per-topic opt-in. Seeded BCONFIG flags (MCP.CLIENT_ENABLED, MULTITASK.MCP_FETCH_ENABLED, MULTITASK.MCP_ACTION_ENABLED) turn this on; an explicit 0 row is the operator kill switch. See docs/MULTITASK_DATA_NODES.md

• Claude Code & Anthropic-compatible API — Point Claude Code or any Anthropic-protocol client at your instance (POST /v1/messages); configure under Manage → Developer & devices → Coding clients (guide)

AI Providers & Models

Synaplan is provider-neutral: connect the providers you want in Operate → AI infrastructure → Providers & keys (keys are validated live and stored encrypted in the database, active without a restart), or set the env variables below in backend/.env — those are read at container start and imported into the encrypted store on first use. Each user picks a different model per task (chat, vision, image, video, audio, embeddings) — nothing is hardcoded.

Transparent pricing. Every model carries its provider's own rate (USD per 1M tokens in/out, or per image / second / character for media) — no proprietary credit unit in between. The selector shows a Free / Low / Mid / High cost badge next to each model and on every answer, GET /api/v1/config/models returns priceIn / priceOut, and the Statistics page logs the real cost of each call. On the hosted instance at web.synaplan.com that same catalog is what your plan meters against; self-hosted with Ollama, the per-token cost is simply zero. Details: Model pricing & cost transparency.

Model catalog changes (new models, retired generations, price updates) ship as seeders, so an existing install is repointed to a supported successor instead of silently keeping a dead model — a retirement records the successor and every install is switched off on the next deploy. See docs/PRICING_MAINTENANCE.md.

Lean by design: core vs optional building blocks

make up starts a complete platform, but the core is deliberately small: the app, its database and Redis. Everything else is a building block that adds one capability and costs RAM. Switch a block on when you need it and off when you don't — Synaplan keeps running either way and simply hides the matching feature. The boot status screen at http://localhost:5173 lists the live on/off state of every block, and Operate → System status (/admin/features) does the same after login.

Keep a default-on block off across restarts. docker compose stop is undone by the next up -d. To make a block opt-in permanently, give it a profile in a docker-compose.override.yml (not tracked by git) — plain up -d then skips it, --profile optional brings it back:

services:
  qdrant:
    profiles: [optional]

Production follows the same rule set: deploy/compose.yaml always starts web/worker/scheduler plus db/redis/centrifugo/tika/qdrant and spoken answers, with office, local-ai and compute as profiles (COMPOSE_PROFILES=office,local-ai,compute in deploy/.env). Kubernetes installs wire the same services via synaplan-charts. The other dev-only containers (phpmyadmin, mailhog, frontend-widgets, startup-notes) never ship to production.

Realtime & Background Processing

Both compose files also start four internal services (no host ports, no setup needed):

In a multi-node cluster all nodes share one Redis, so WebSocket events published on one node reach browsers connected to any other. Details: docs/REALTIME.md.

Text-to-Speech

Voice output starts with make up — synaplan-tts, image ghcr.io/metadist/synaplan-tts. The image already contains five Piper voices (English, German, Spanish, French, Turkish). Stop the tts container to hide the speaker control.

# Already running after `make up`. Standalone on another machine:
docker run -d --name synaplan-tts -p 127.0.0.1:10200:10200 ghcr.io/metadist/synaplan-tts:latest

The backend looks at SYNAPLAN_TTS_URL (compose default http://tts:10200).

The UI language selects the voice. Chat sends the active frontend locale (en / de / es / fr / tr); if the backend detects a different reply language, that wins. Piper then maps the short code to the matching baked voice (German UI → kerstin, Spanish → davefx, …). There is no separate voice picker. Add more Piper models by dropping .onnx + .onnx.json into the extra-voices volume — see synaplan-tts README and docs.synaplan.com/tts.

Office documents

Thumbnails, PDF export, inline preview and “Combine as PDF” need an optional Collabora CODE sidecar. It is off by default, so make up does not pull it. Chat and DOCX / XLSX / PPTX generation work without it. Setup, including why OFFICE_CONVERT_URL does not belong in backend/.env: Installation — Optional office conversion.

File work

The assistant can run a short Python or Node program on copies of files you already picked. make up starts it. The dev token, and how to turn the sidecar off, are in Installation — File work.

Common Commands

# Startup progress ("please wait..." notes + READY message)
docker compose logs -f startup-notes

# Logs
docker compose logs -f backend

# Restart
docker compose restart backend

# Reset database
docker compose down -v && make up

# Run tests
make test

# Code quality
make lint

Documentation

User-facing & API docs live at docs.synaplan.com. Source: metadist/synaplan-docs.

In-repo guides (for developers working on this codebase):

Related Repositories

Project Structure

synaplan/
├── backend/        # Symfony PHP API
├── frontend/       # Vue.js SPA
├── docs/           # Developer documentation
├── deploy/         # Production self-host compose + lifecycle scripts
├── sidecars/       # Optional sidecars (secure compute)
├── _docker/        # Docker configs
└── plugins/        # Plugin system

Star history

Community & Support

• Discord — chat with the team and community; the fastest place for self-hosting and configuration questions

• GitHub Issues — bugs and feature requests

• Security policy — how to report a vulnerability privately, and what happens next

• www.synaplan.com — product, hosting and enterprise contact

Contributing

See CONTRIBUTING.md. AGENTS.md is the guide for coding agents working in this repository.

License

Apache-2.0