Orion

The declarative runtime for AI agents, workflows, microservices, and event processing.

Safe enough to let an AI write your services. Fast enough to run them in production.

Orion is a declarative services runtime. A service is one JSON document holding the logic, the connectors it reaches, and the endpoint it answers on. Post it to a running server and it's live a second later. No rebuild, no restart, no downtime.

Everything around that logic is the runtime's job, and it works the same way for every service you put on it: route and protocol matching, ingress guards, rate limiting, circuit breaking, fault tolerance, connection pooling, zero-downtime hot reload, and end-to-end observability. That's the glue you'd otherwise write again for every microservice, agent backend, stream processor, and data pipeline.

It ships as a single Rust binary on Tokio and Axum, storing your service definitions in an embedded database. There's nothing to containerize and nothing to provision.

Jump to: Quickstart · Why Orion? · What you can build · Is Orion right for you? · Three primitives · The console · What's built in · Connectors · Functions · Performance · Install · Docs

Why Orion?

Open a small internal microservice and count the lines. HTTP server setup, connection pools, a Prometheus exporter, OpenTelemetry wiring, retry loops, a circuit breaker, health checks, a Dockerfile, a deploy manifest. Somewhere in the middle sits the logic you actually cared about, and it's maybe fifty lines long. Orion runs that middle part for you and provides everything around it, the same way, for every service.

• ⚡ No service to build: Idea to live REST or Kafka endpoint in seconds. No Dockerfile, no CI pipeline, no server code.
• 🛡️ Production features included: Rate limiting, circuit breakers, timeouts, caching, and payload validation on every endpoint. You configure them instead of writing them.
• 🤖 Safe for AI-written logic: Models generate JSON reliably. Validation, draft-before-activate, dry-run, percentage rollout, and one-command rollback mean AI output can't quietly break production.
• 🧩 Services that call services: channel_call runs another workflow in-process, so there's no network hop and no serialization cost.
• 📦 One binary, one file: a single Rust binary with an embedded database — with PostgreSQL or MySQL waiting for when you outgrow that.
• 🦀 Measured, not claimed: 5.1K–5.7K workflow requests/sec per instance with single-digit millisecond latency, on the published v1.0.0 benchmark record — run conditions and all.

Your First Service in 2 Minutes

No code. No Dockerfile. No CI pipeline. Just a running service.

1. Start Orion

brew install GoPlasmatic/tap/orion-server   # or: curl installer, cargo install (see Install)
orion-server

2. Deploy your first service (one command)

curl -fsSL https://raw.githubusercontent.com/GoPlasmatic/Orion/main/examples/quickstart.sh | bash

The script talks to the same admin API you'd use in production. It creates a workflow (the logic: flag any order over $10,000 for review) and a channel (the endpoint: POST /orders), activates both, and sends a first test order. Re-running it is safe. Cloned the repo? Run ./examples/quickstart.sh instead.

Create the workflow, with the business logic as JSON (a parse task, then a conditional flag task):

curl -s -X POST http://localhost:8080/api/v1/admin/workflows \
  -H "Content-Type: application/json" \
  -d '{
    "workflow_id": "quickstart-orders",
    "name": "High-Value Order",
    "condition": true,
    "tasks": [
      { "id": "parse", "name": "Parse payload", "function": {
          "name": "parse_json",
          "input": { "source": "payload", "target": "order" }
      }},
      { "id": "flag", "name": "Flag order",
        "condition": { ">": [{ "var": "data.order.total" }, 10000] },
        "function": {
          "name": "map",
          "input": { "mappings": [
            { "path": "data.order.flagged", "logic": true },
            { "path": "data.order.alert", "logic": { "cat": ["High-value order: $", { "var": "data.order.total" }] } }
          ]}
      }}
    ]
  }'

# Activate it (draft → active; the engine hot-reloads)
curl -s -X PATCH http://localhost:8080/api/v1/admin/workflows/quickstart-orders/status \
  -H "Content-Type: application/json" -d '{"status": "active"}'

Create the channel, the endpoint that routes to the workflow, and activate it:

curl -s -X POST http://localhost:8080/api/v1/admin/channels \
  -H "Content-Type: application/json" \
  -d '{ "channel_id": "orders", "name": "orders", "channel_type": "sync",
        "protocol": "rest", "route_pattern": "/orders",
        "methods": ["POST"], "workflow_id": "quickstart-orders" }'

curl -s -X PATCH http://localhost:8080/api/v1/admin/channels/orders/status \
  -H "Content-Type: application/json" -d '{"status": "active"}'

3. Call it. Your service is live

curl -s -X POST http://localhost:8080/api/v1/data/orders \
  -H "Content-Type: application/json" \
  -d '{ "data": { "order_id": "ORD-9182", "total": 25000 } }'

{
  "status": "ok",
  "data": {
    "order": {
      "order_id": "ORD-9182",
      "total": 25000,
      "flagged": true,
      "alert": "High-value order: $25000"
    }
  }
}

That's it. The business logic is a JSON document, deploying it was an API call, and rate limiting, metrics, health checks, and request tracing were already active when it went live. Change the threshold? One API call. No rebuild, no redeploy, no restart.

Prefer to describe the service instead of writing it? Workflow JSON is easy for LLMs to generate. Tell your AI assistant "flag orders over $10,000 for manual review with an alert message" and deploy what it returns. AI Writes Services, Not Code shows the safe path from prompt to production.

What You Can Build

Orion carries the same infrastructure across five kinds of service:

• Microservices: one channel and one workflow make a service, and Orion answers the request in-process — nothing you built sits in the path.
• AI agent tools: an agent calls your channels as tools over HTTP, and through the MCP server in orion-cli an assistant drafts, dry-runs, activates, and rolls back those workflows itself.
• Business rules & decision APIs: pricing tiers, eligibility checks, routing decisions — written as JSONLogic conditions over the request and returned as the response body.
• Kafka event consumers: a topic is the ingress — consume records, transform and enrich them as they arrive, publish results onward, and send poison messages to a dead-letter topic instead of letting one stall the partition.
• Webhook & data ingestion: normalize payloads from Stripe, GitHub or Shopify, then read and write across PostgreSQL, MySQL, SQLite, MongoDB and Elasticsearch through one portable dialect.

See Worked Examples for complete, tested examples, or grab a ready-to-deploy example package from examples/packages/ and run ./examples/deploy.sh <name> against a local instance.

Is Orion Right for You?

Longer, tool-by-tool discussion (Temporal, Kong, Drools, n8n, dataflow-rs): Is Orion Right for You?

Three Primitives

You build services in Orion with three things:

graph LR
    C["Channel<br>(endpoint)"] --> W["Workflow<br>(logic)"]
    W --> Co["Connector<br>(external)"]

    style C fill:#21252b,stroke:#61afef,stroke-width:2px,color:#abb2bf
    style W fill:#21252b,stroke:#61afef,stroke-width:2px,color:#abb2bf
    style Co fill:#21252b,stroke:#61afef,stroke-width:2px,color:#abb2bf

Design-time: define channels, build workflows, configure connectors, test with dry-run, manage versions, all through the admin API.

Runtime: Orion routes traffic to channels, executes workflows, calls connectors, and handles observability automatically.

The Console

Orion is API-first, and everything it does is also point-and-click. Orion UI is the operations console for a running instance. It gives you live dashboards, a system map of every channel → workflow → connector, workflow logic visualization, trace drill-downs, and a data console for firing test requests. Run docker compose up next to the server, or npm run dev.

AI Writes Services, Not Code

When AI generates a microservice, you still need to add health checks, metrics, retries, and error handling. When AI generates an Orion workflow, all of that is already there. The platform guarantees it.

Use the Orion CLI's MCP server to give your AI assistant full Orion context. No manual prompt engineering needed. The MCP server exposes tools covering the full Orion API: workflow syntax, available functions, connector types, and API operations. One config block and you're done (Claude Code .mcp.json, Claude Desktop, or any MCP client):

{
  "mcpServers": {
    "orion": {
      "command": "orion-cli",
      "args": ["mcp", "serve"],
      "env": { "ORION_SERVER_URL": "http://localhost:8080" }
    }
  }
}

No MCP client? Paste the prompt pack into any LLM and it can write and deploy workflows through the plain REST API. It's a self-contained context block with Orion's schemas, conventions, and API calls.

You: "Classify orders into VIP (>=500, 15% discount), Premium (100-500, 5%), and Standard tiers"

AI:  → generates valid workflow JSON
     → creates it via the API
     → tests with dry-run
     → activates when you approve

The safe path from AI output to production, every time:

graph TD
    A["1. Generate<br>(AI Workflow JSON via MCP)"] --> B["2. Validate<br>(Verify JSON Syntax & Schema)"]
    B --> C["3. Create Draft<br>(Saved in DB, Offline)"]
    C --> D["4. Dry-Run Test<br>(Verify with Sample Data)"]
    D --> E["5. Activate<br>(Engine hot-reloads)"]
    E --> F["6. Canary Rollout<br>(10% → 50% → 100% traffic)"]

    style A fill:#21252b,stroke:#5c6370,color:#abb2bf
    style B fill:#21252b,stroke:#5c6370,color:#abb2bf
    style C fill:#21252b,stroke:#5c6370,color:#abb2bf
    style D fill:#21252b,stroke:#5c6370,color:#abb2bf
    style E fill:#2e3f2f,stroke:#98c379,stroke-width:2px,color:#98c379
    style F fill:#293c4e,stroke:#61afef,stroke-width:2px,color:#61afef

Every AI-generated workflow gets version history, draft-before-activate, dry-run testing, rollout control, structured FieldError validation feedback, and audit trails. It's the same governance hand-written workflows get. Roll back to any previous version instantly.

The workflows, channels, and connectors of one service form a package — Orion's unit of shipping, and what makes one instance a modular monolith: many services side by side, each promoted and rolled back independently. orion-server package is the promotion story: export computes the dependency closure from a source instance into one JSON artifact (git is the registry), lint and plan check it with zero writes, apply stages and activates everything in dependency order with a single engine reload and a version-immutable package receipt, and diff reports drift. The bulk import endpoints (POST /api/v1/admin/{workflows,channels,connectors}/import?dry_run=true, then drop dry_run to commit) remain the low-level primitive when you need to script a single batch. See Packages & Promotion.

See CI/CD with Packages for CI/CD integration and GitHub Actions examples.

Before & After

Before: every piece of business logic is its own service to build, deploy, and operate — a pricing service, a fraud service, a routing service, a notification service, each with its own repo, pipeline, and pager entry.

After: one Orion instance replaces all of them. It routes traffic, runs the workflows, and polices its own ingress — rate limits, validation, deduplication — while each channel and workflow stays independently versioned, testable, and deployable. The modularity of microservices with the operational simplicity of a monolith: change one workflow without touching the others, roll back a single channel without redeploying anything.

The architecture overview draws both topologies side by side.

What's Built In

Every channel gets production-grade features without writing a line of code. Configure per channel or use platform defaults:

A minimal channel needs only a name and a workflow. Everything else has sensible defaults.

Observability deep dive: health endpoints, full Prometheus metrics list, Kubernetes probes, and OpenTelemetry tracing config. See Observability Guide.

Sync and Async

Sync     POST /api/v1/data/{channel}         → immediate response
Async    POST /api/v1/data/{channel}/async   → returns trace_id, poll later

REST     GET /api/v1/data/orders/{id}        → matched by route pattern
Kafka    topic: order.placed                 → consumed automatically

Sync channels respond immediately. Async channels return a trace ID; poll GET /api/v1/admin/traces/{id} for results. Kafka channels consume from topics configured in the DB or config file, no restart needed when you add new ones.

Bridging is a pattern, not a feature. A sync workflow can publish_kafka and return 202. An async channel picks it up from there.

REST channels support parameterized route patterns (/orders/{order_id}) with path, query, and header injection into the workflow context. See Data API.

Service Composition

Most platforms require HTTP calls between services, adding latency, failure modes, and serialization overhead. Orion's channel_call invokes another channel's workflow in-process with zero network round-trip:

graph TD
    Req["POST /orders (Workflow)"] --> P["parse_json<br>(Extract order data)"]
    P --> C1["channel_call<br>(inventory-check)"]
    C1 --> C2["channel_call<br>(customer-lookup)"]
    C2 --> M["map<br>(Compute pricing)"]
    M --> Res["publish_json<br>(Combined response)"]

    classDef task fill:#21252b,stroke:#5c6370,color:#abb2bf;
    classDef inprocess fill:#61afef,stroke:#61afef,color:#1e222b;
    class P,M,Res task;
    class C1,C2 inprocess;

Each composed channel has its own workflow, versioning, and governance, but calls between them are function calls, not network hops. Cycle detection prevents infinite recursion.

Connect to Anything

Connectors are named, reusable connections to external systems. Configure once, reference by name in any workflow. Credentials stay out of your logic:

Every connector gets circuit breaker protection automatically: failures trip the breaker, subsequent calls fast-fail, and the breaker auto-recovers. Database and Elasticsearch connectors also carry per-operation gates (operations: { read, insert, update, delete, upsert, raw_write }). Set "delete": false and no workflow can delete through that connector, no matter what its tasks say. Secrets are stored in the database and masked in API responses, and any string field can use an env://VAR_NAME reference to pull the value from the process environment at startup so production credentials never sit in the saved config. See Connectors Guide for configuration examples and auth options.

Built-in Task Functions

All functions are built into every binary. The dataflow-rs runtime contributes the data-shaping core (parse_json, parse_xml, filter, map, validation, publish_json, publish_xml, log); Orion adds the connector-backed handlers (http_call, data_query, data_write, db_read, db_write, cache_read, cache_write, mongo_read, publish_kafka) and the in-process channel_call. data_query/data_write speak the portable data dialect — write the query once and switch between SQL, MongoDB, and Elasticsearch by switching connectors, with db_read/db_write as the raw-SQL escape hatch. See the Function Reference for every function's exact input schema, or browse them at runtime via GET /api/v1/admin/functions.

When Things Go Wrong

Production services fail, and Orion handles the standard failure modes without you writing retry loops or fallback logic: a downed external API trips its circuit breaker, slow workflows time out with a 504, traffic spikes hit the rate limiter (429) and backpressure (503), failed async tasks land in a dead-letter queue with automatic retry, and duplicate requests are caught by idempotency keys. Each behaviour is configurable per channel or connector — the Resilience Guide covers every failure mode and its knobs.

Debugging is built in. Every request gets a x-request-id propagated through the entire pipeline, structured JSON logs show what each task received and produced, and OpenTelemetry traces http_call/channel_call chains end to end. Inspect circuit breakers, DLQ traces, and debug endpoints via the API Reference.

Deploy Anywhere

flowchart LR
    subgraph Standalone ["Standalone"]
        direction TB
        S["./orion-server"]
    end

    subgraph Container ["Docker / Kubernetes"]
        direction TB
        D["docker run ghcr.io/goplasmatic/orion:latest"]
    end

    subgraph Cluster ["HA Cluster (cluster mode)"]
        direction TB
        LB["Load balancer"] --> N["Orion × N replicas"]
        N --> B["PostgreSQL/MySQL + Redis (shared)"]
    end

    style Standalone fill:#21252b,stroke:#5c6370,color:#abb2bf
    style Container fill:#21252b,stroke:#5c6370,color:#abb2bf
    style Cluster fill:#21252b,stroke:#5c6370,color:#abb2bf
    style S fill:#61afef,stroke:#61afef,color:#1e222b
    style D fill:#61afef,stroke:#61afef,color:#1e222b
    style LB fill:#4b5263,stroke:#5c6370,color:#abb2bf
    style N fill:#61afef,stroke:#61afef,color:#1e222b
    style B fill:#4b5263,stroke:#5c6370,color:#abb2bf

Single binary. SQLite by default, no database to provision, no runtime dependencies. Need more scale? Swap to PostgreSQL or MySQL by changing the storage.url. No rebuild needed.

Need more than one node? Turn on cluster mode. N identical replicas behind a load balancer share one PostgreSQL/MySQL and one Redis and behave as a single logical system: a config change made through any node reaches every node in about two seconds, idempotency keys and rate limits hold across the whole fleet (one execution per key, limits counted fleet-wide, cache hits shared), and rolling deploys are zero-downtime — replicas drain gracefully while the rest keep serving. It ships two packaged ways:

# Kubernetes — published to GHCR on every release
helm install orion oci://ghcr.io/goplasmatic/charts/orion

# Anywhere else — nginx + 2 replicas + Postgres + Redis, migrations included
docker compose -f docker-compose.ha.yml up

Same channel definitions work in any topology: one instance, an HA cluster — or dedicated capacity by splitting channels across instance pools with include/exclude filters. The definition doesn't change; only the deployment config does.

Performance

5.1K–5.7K workflow requests/sec on a single instance, and a 58K req/s health-check baseline, as measured on v1.0.0 (Apple M2 Pro Mac Mini, release build, 30s per scenario, 50 concurrent connections — full raw record with run conditions in tests/benchmark/results/v1.0.0/):

Tail latency improved over the v0.2.0 record (simple-workflow P99 12.6 ms vs 16.7 ms) while straight-line throughput reads lower; the record's SUMMARY.md carries the honest comparison, including what changed on the hot path in 1.0 (always-on per-task Prometheus timing) and the capture conditions. Zero errors across every scenario, including 56 engine hot-reloads under sustained load. Run ./crates/orion-server/tests/benchmark/bench.sh to reproduce the single-instance scenarios, and ./crates/orion-server/tests/benchmark/bench.sh cluster to drive the HA compose stack through its load balancer (not part of this record — the capture host had no Docker).

Pre-compiled JSONLogic, zero-downtime hot-reload, lock-free reads, SQLite WAL mode, async-first on Tokio.

Install

# Docker (quickest way to try)
docker run -p 8080:8080 ghcr.io/goplasmatic/orion:latest

# Docker Compose (with persistent storage)
docker compose up  # uses docker-compose.yml from this repo

# Homebrew (macOS Apple silicon, Linux; Intel Macs build from source via `cargo install` below)
brew install GoPlasmatic/tap/orion-server

# macOS (Apple silicon) / Linux (shell installer)
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/GoPlasmatic/Orion/releases/latest/download/orion-server-installer.sh | sh

# Windows (PowerShell)
powershell -ExecutionPolicy ByPass -c "irm https://github.com/GoPlasmatic/Orion/releases/latest/download/orion-server-installer.ps1 | iex"

# From crates.io
cargo install orion-server

# From source — this repo is a two-binary workspace, so the package must be named
cargo install --git https://github.com/GoPlasmatic/Orion.git --locked orion-server

# ...or install the server and the CLI in one go
cargo install --git https://github.com/GoPlasmatic/Orion.git --locked orion-server orion-cli

Verify with orion-server --version. Swagger UI available at http://localhost:8080/docs. See Configuration for deployment options.

The server binary also ships diagnostic and promotion subcommands you can run without booting the HTTP listener:

orion-server validate-config --format summary       # Parse + validate the config, one-screen summary, secrets redacted
orion-server lint workflow.json                     # Strict-validate a workflow JSON file
orion-server dry-run -w wf.json -i input.json       # Execute a workflow offline (--stubs answers connector calls)
orion-server test examples/workflow-tests           # Run offline *.case.json workflow regression tests
orion-server preflight                              # Scan stored channels/workflows before upgrading
orion-server package apply -s <url> -f pkg.json     # Promote a package (export/lint/plan/apply/diff)

The full list — migrate, test-connectivity, dump-openapi, every flag — is in the CLI Commands reference. ${VAR} / ${VAR:-default} placeholders inside config.toml are substituted from the environment when any of these subcommands load the config, so the same file works across dev, staging, and prod without templating.

CLI Tool

Manage workflows, channels, and connectors without writing curl commands:

# Install — versioned in lockstep with the server and shipped in the same release
brew install GoPlasmatic/tap/orion-cli                # Homebrew
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/GoPlasmatic/Orion/releases/latest/download/orion-cli-installer.sh | sh
cargo install --git https://github.com/GoPlasmatic/Orion --locked orion-cli  # From source

# Deploy a workflow from a JSON file
orion-cli workflows create -f order-processing.json
orion-cli --yes workflows activate high-value-order
orion-cli channels create -f orders-channel.json
orion-cli --yes channels activate orders

See the CLI reference for the full command list — the CLI is developed in this repo at crates/orion-cli.

Documentation

The full book lives at docs.goplasmatic.io — getting-started tutorials, architecture, per-feature guides (observability, resilience, scalability, security, availability, maintainability, deployability), the Data API, and the portable data dialect. The five a newcomer reaches for first:

Built With

Axum (HTTP), Tokio (async runtime), SQLx (database), sea-query (portable SQL builder), SQLite/PostgreSQL/MySQL (storage), datalogic-rs (JSONLogic), dataflow-rs (workflow orchestration).

Ecosystem & Roadmap

Orion ships with two companion projects:

• Orion UI: the admin dashboard. Manage workflows, channels, and connectors, visualize workflow pipelines, inspect audit trails, and monitor engine health from the browser.
• Orion CLI: the command-line interface and MCP server, developed in this repo. Manage everything from your terminal or AI assistant.

Under consideration: workflow marketplace (community templates), cron-based scheduling, WASM task functions, and language SDKs. Have an idea or want to push one of these forward? Open an issue or start a discussion.

Who's Using Orion?

Using Orion in a project, a company, or a side quest? Add yourself to ADOPTERS.md with a one-line PR, or share what you built in Discussions. Real-world usage reports directly shape the roadmap.

Contributing

Contributions welcome! Whether it's a bug fix, new connector, documentation improvement, or feature request, we'd love to hear from you. CONTRIBUTING.md has everything: dev setup, how to run the container-gated tests, the PR checklist, and commit conventions. The project follows the Contributor Covenant; notable changes are tracked per package in the server CHANGELOG and the CLI CHANGELOG.

• Report bugs: Open an issue
• Ask questions: GitHub Discussions
• Report security issues privately: see SECURITY.md
• Submit code: Fork, branch, PR — see CONTRIBUTING.md for the full gate
• Docs recordings: the README GIFs and mdBook asciinema casts are generated from real sessions. See docs/recordings/ to regenerate them.

Support the Project

If Orion looks useful, a ⭐ on this repo is the easiest way to help other developers find it. Beyond that: share what you build in Discussions, add yourself to ADOPTERS.md, or send this to a colleague who's tired of building a new service for every bit of business logic.

License

Apache-2.0. See LICENSE for details.