Architecture

A detailed look at how OpenCompany works under the hood. For contributors, system designers, and the curious. You do not need this page to use OpenCompany. For a navigable tour with hyperlinks, see the DeepWiki page.

System Overview

OpenCompany is three loosely-coupled tiers talking over a single persistent WebSocket connection:

+------------------------------------------------------------------------+
|  Frontend tier (client/)                                               |
|    React 19 + TypeScript + Vite + React Flow + Zustand                 |
|    Tailwind v4 + shadcn/ui + Radix + TanStack Query                    |
|    Themed canvas, parameter panel, credentials modal                   |
+------------------------------------------------------------------------+
                             |
                 WebSocket   |   /ws/status  (single long-lived connection)
                             v
+------------------------------------------------------------------------+
|  Backend tier (server/)                                                |
|    FastAPI + WebSocket handlers + dependency injection container       |
|    Plugin-first nodes (BaseNode folders) -> backend NodeSpec is SSOT   |
|    WorkflowService (facade) -> NodeExecutor -> per-node execute()      |
|    ParameterResolver -> template {{node.field}} substitution           |
|    AuthService -> encrypted credentials (Fernet + PBKDF2)              |
+------------------------------------------------------------------------+
                             |
                             v
+------------------------------------------------------------------------+
|  Execution tier                                                        |
|    Temporal workers (distributed)  or  local asyncio decide loop       |
|    Redis (cache, locks, streams)   SQLite (workflow.db, credentials.db)|
|    Node.js server (JS/TS code exec, port 3020)  WhatsApp RPC (9400)    |
|    Android relay WebSocket         Temporal server (ports 7233/8080)   |
+------------------------------------------------------------------------+
                             |
                             v
   External services: OpenAI, Anthropic, Gemini, OpenRouter, xAI, DeepSeek,
   Kimi, Mistral, Groq, Cerebras, plus local Ollama / LM Studio servers,
   Google Workspace, WhatsApp, Telegram, Twitter/X, Brave, Serper,
   Perplexity, Apify, Stripe, residential proxies, webhooks.

100+ Workflow Nodes

AI, agents, social, Android, documents, Google Workspace, code, browser, proxies, utilities

12 LLM Providers

OpenAI, Anthropic, Gemini, OpenRouter, xAI, DeepSeek, Kimi, Mistral, Groq, Cerebras, plus local Ollama and LM Studio

16 Specialized Agents

Android, Coding, Web, Task, Social, Travel, Tool, Productivity, Payments, Consumer, Autonomous, Orchestrator, AI Employee, RLM, Claude Code, Codex

WebSocket-First

A single persistent connection carries all frontend-backend RPC and push updates

Built-in Skills

Dozens of skills across many categories, DB-backed with SKILL.md defaults

Three Execution Modes

Temporal distributed, Redis parallel, sequential fallback

Execution Engine

Execution flow diagram — How a workflow run flows through the execution engine

Conductor’s Decide Pattern

Workflow orchestration is a single function with fork/join parallelism:

_workflow_decide(ctx):
Find ready nodes (all dependencies satisfied)
asyncio.gather() the ready layer -> run in parallel
Checkpoint state to Redis/SQLite
Recurse until every node terminal (completed / skipped / failed)

Each workflow run has its own isolated ExecutionContext. No shared global state between concurrent runs. Decide loops serialize per execution via Redis SETNX distributed locks.

Three Execution Modes

workflow.execute(workflow_id, workflow_data)
                    |
                    v
    TEMPORAL_ENABLED and server reachable?
                    |
         yes -> _execute_temporal() -- per-node activities,
                                        retries, horizontal scaling
                                        (primary production mode)
          no -> Redis available?
                    |
             yes -> _execute_parallel() -- decide loop + Kahn layers
                                            + Prefect-style input hash cache
                                            + DLQ + heartbeat recovery
              no -> _execute_sequential() -- topological walk
                                             (fallback for minimal env)

Layer Computation via Kahn’s Algorithm

Before execution, the DAG is sorted into layers. Layer 0 is the set of nodes with no dependencies; each subsequent layer depends only on earlier layers. Parallel execution runs each layer with asyncio.gather():

Layer 0: [start, cronScheduler]
Layer 1: [httpRequest, whatsappReceive]
Layer 2: [aiAgent]
Layer 3: [whatsappSend, console]

Toolkit sub-nodes (e.g., Android service nodes connected to androidTool) are detected and excluded from execution layers — they run only when the parent toolkit invokes them via tool calling.

Result Caching, Recovery, and DLQ

Prefect-style caching: every node result is stored in Redis/SQLite keyed by hash_inputs(inputs). Re-running an identical node returns the cached result with status TaskStatus.CACHED.
Heartbeat recovery: RecoverySweeper scans executions:active every 60s; nodes with stale heartbeats (> 5 min) are marked stuck and recovered on next startup.
Dead Letter Queue: failed nodes (after retry exhaustion) are quarantined with full input snapshot. Inspect, replay, or purge via the get_dlq_* / replay_dlq_entry / purge_dlq WebSocket handlers.

Edge Conditions

Edges carry optional conditions for runtime branching with 20+ operators (eq, neq, gt, lt, contains, exists, matches, in, …). Unmatched branches are marked TaskStatus.SKIPPED.

Event-Driven Deployment

Deployments are event-driven: each trigger event spawns an independent concurrent execution run. There is no iteration loop.

deploy_workflow(workflow_id)
        |
        v
  Set up triggers, return
        |
        +-> cronScheduler fires   -> ExecutionRun 1  (isolated context)
        +-> cronScheduler fires   -> ExecutionRun 2  (isolated context)
        +-> whatsappReceive fires -> ExecutionRun 3  (isolated context)
        +-> webhookTrigger fires  -> ExecutionRun 4  (isolated context)
        +-> telegramReceive fires -> ExecutionRun 5  (isolated context)
        +-> taskTrigger fires     -> ExecutionRun 6  (from delegated agent)

Multiple runs execute simultaneously with no interference. The firing trigger is marked complete before downstream execution starts; every other trigger node in the same run is auto-marked _pre_executed with {not_triggered: True} so they never block as event waiters.

Push vs Polling Triggers

Push triggers (fire on an inbound event):
  whatsappReceive, webhookTrigger, chatTrigger, taskTrigger,
  telegramReceive, emailReceive

Polling triggers (poll an API on an interval):
  googleGmailReceive  (Gmail push requires paid Google Cloud setup)
  twitterReceive      (X API has no webhook on the free tier)

Scheduler:
  cronScheduler       (cron-expression schedule)

Under the default Temporal execution path, each deployed trigger runs as a listener workflow that spawns an isolated child run per event. On the legacy local path, push triggers register a waiter with a filter closure and suspend until an external service dispatches a matching event (in-memory asyncio.Future, or Redis Streams for multi-worker deployments).

AI Agent System

AI agent routing diagram — Native SDK chat path vs LangChain agent path

12 LLM Providers

Each provider below has its own chat-model node and appears in every agent’s provider dropdown. (xAI is the exception: it is reachable on the native chat path through the shared OpenAI-compatible client, but it is not a standalone chat-model node.)

Provider	Native SDK Path	LangChain Path	Notes
OpenAI	Yes	Yes	GPT-5.5/5.4/5.2, GPT-4.1 family, o3/o4-mini reasoning, GPT-4o/4o-mini
Anthropic	Yes	Yes	Claude Fable 5, Opus 4.8/4.7, Sonnet 4.6, Haiku 4.5 (extended thinking via `budget_tokens`)
Gemini	Yes	Yes (fallback)	Gemini 3.5-flash, 3.1-pro/flash-lite, 3-flash, 2.5-pro/flash/flash-lite; native path bypasses a LangChain Windows hang
OpenRouter	Yes	Yes	200+ models through one API
xAI	Yes (shared OpenAIProvider)	No (native chat only)	Grok models; no standalone chat-model node
DeepSeek	Yes (shared OpenAIProvider)	Yes	deepseek-v4-flash, deepseek-v4-pro (chat/reasoner legacy)
Kimi	Yes (shared OpenAIProvider)	Yes	Moonshot kimi-k2.6, kimi-k2.5, kimi-k2.7-code
Mistral	Yes (shared OpenAIProvider)	Yes	Large / Medium / Small / Codestral (latest)
Groq	No (LangChain only)	Yes	Llama 3.3-70b, Llama 3.1-8b, GPT-OSS-120b/20b, Qwen3-32b
Cerebras	No (LangChain only)	Yes	GPT-OSS-120b, Qwen-3-235b, GLM-4.7
Ollama	Yes (shared OpenAIProvider)	Yes	Local server; whatever models the user has pulled
LM Studio	Yes (shared OpenAIProvider)	Yes	Local server; whatever models the user has loaded

Dual-Path Architecture

execute_chat(parameters)                        execute_agent(parameters)
  direct chat completions                         tool-calling agent loop
        |                                               |
        v                                               v
is_native_provider(provider)?              create_model(provider, ...)
        |                                               |
   yes -+-- create_provider() -> LLMResponse            v
        |                                       LangChain ChatOpenAI /
    no -+-- create_model() -> chat_model                ChatAnthropic /
            .invoke() -> LLMResponse                    ChatGoogleGenerativeAI
                                                        |
                                                        v
                                                   bind_tools(...)
                                                        |
                                                        v
                                                 _run_agent_loop
                                                 (LLM step <-> tool exec loop)

The native path returns a normalized LLMResponse dataclass across all providers. Agent tool-calling still uses the LangChain chat-model wrappers (bind_tools) driven by the plain-async _run_agent_loop; the native SDK path powers direct chat completions and model discovery. Groq and Cerebras remain LangChain-only for chat as well.

16 Specialized Agents

All specialized agents share the same handle architecture (input-main, input-memory, input-skill, input-tools, input-task) and dispatch through the same BaseNode.execute() plugin path. They mostly differ in icon, title, theme color, and default prompt.

android_agent      coding_agent        web_agent         task_agent
social_agent       travel_agent        tool_agent        productivity_agent
payments_agent     consumer_agent      autonomous_agent  orchestrator_agent
ai_employee        rlm_agent           claude_code_agent codex_agent

Routing:

Most agents run the shared plain-async agent loop (_run_agent_loop) for tool calling and delegation.
rlm_agent routes to RLMService (REPL-based recursive LM with llm_query / rlm_query / FINAL).
claude_code_agent and codex_agent run through the multi-provider CLI agent framework (Claude Code / Codex CLIs) instead of the standard LLM loop.

Agent Teams Topology

orchestrator_agent and ai_employee have an extra input-teammates handle. Connected agents become delegate_to_<agent_type> tools automatically:

                   +-------------------------+
                   |     AI Employee         |
                   |   (orchestrator_agent)  |
                   +-----+------+------+-----+
                         |      |      |         input-teammates
           +-------------+      |      +------------+
           |                    |                   |
    +------v------+     +-------v-----+     +-------v-----+
    | coding_agent|     |  web_agent  |     | task_agent  |
    +-------------+     +-------------+     +-------------+
       delegate_         delegate_             delegate_
       to_coding_        to_web_               to_task_
       agent tool        agent tool            agent tool

The team lead’s LLM decides when to delegate based on task context. Under the default Temporal execution path, delegation spawns the teammate as a child workflow whose progress mirrors onto the parent’s canvas badge and whose result flows back into the parent’s loop. On the legacy local path, delegation is fire-and-forget: the child spawns as asyncio.create_task(), the parent continues, and the child broadcasts its own status independently. Either way, child completion can be consumed by taskTrigger nodes elsewhere in the workflow.

Agent Loop Flow

             +---------+
  START ---> |  agent  |
             | (LLM)   |
             +----+----+
                  |
          any tool_calls?
              /        \
       tools /          \ end
            v            v
       +---------+      END
       |  tools  |
       | (exec)  |
       +----+----+
            |
            +---------> loop back to agent

The loop routes purely on the model’s tool_calls. A recursion limit guards against infinite tool loops, and a token-based compaction threshold summarizes the transcript before it overflows the model’s context. Tools are built as Pydantic-schema StructuredTool instances with the node’s parameter schema.

Memory, Skills, Tokens, and Cost

Markdown-Based Memory

The simpleMemory node stores conversation history as editable Markdown in the parameter panel. The AI Agent handler reads, parses to LangChain messages, executes, appends the new exchange, trims to a window, and archives removed messages to an InMemoryVectorStore (optional) using HuggingFace BAAI/bge-small-en-v1.5 embeddings.

AI Agent reads memoryContent (Markdown)
        |
        v
_parse_memory_markdown() -> LangChain Messages
        |
        v
(optional) vector store similarity_search(prompt, k=retrievalCount)
        |
        v
Execute LLM with full history + retrieved context
        |
        v
Append human + ai messages to Markdown
        |
        v
_trim_markdown_window(windowSize) -> (kept, removed)
        |
        v
If longTermEnabled: store.add_texts(removed)
        |
        v
Save updated Markdown back to node parameters

Skill System

Dozens of built-in skills organized into folders under server/skills/. Each top-level folder appears as an option in the Master Skill node’s folder dropdown:

assistant/            android_agent/        autonomous/
coding_agent/         productivity_agent/   rlm_agent/
social_agent/         task_agent/           travel_agent/
web_agent/            terminal/             payments_agent/

Each skill is a folder with a SKILL.md file containing YAML frontmatter (name, description, allowed-tools, metadata) and Markdown instructions. First load seeds the database; after that the database is source of truth so users can edit skill instructions in the UI. “Reset to Default” reloads the original SKILL.md. The masterSkill node aggregates multiple skills with enable/disable toggles. A split-panel editor shows the skill list on the left and the selected skill’s Markdown on the right. When connected to an agent, the backend expands the skillsConfig parameter into individual skill entries injected into the agent’s system message.

Token Tracking and Compaction

Every AI execution stores a TokenUsageMetric row with input/output/cache/reasoning token counts and calculated costs (USD) based on server/config/pricing.json. Cumulative state per session lives in SessionTokenState. Compaction threshold priority (highest to lowest):

Per-session custom_threshold (user-set)
Per-user UserSettings.compaction_ratio x the model’s context window
Environment COMPACTION_RATIO (default 0.8) x the model’s context window (e.g., ~838K for Claude Sonnet 4.6 with 1M context)
JSON fallback llm_defaults.json ratio x context window

When the cumulative session tokens cross the threshold, CompactionService.compact_context() generates a 5-section summary (Task Overview, Current State, Important Discoveries, Next Steps, Context to Preserve) following the Claude Code pattern and replaces the memory content. Anthropic and OpenAI also have native compaction APIs (context_management edits, compact_threshold) that are configured transparently.

Communication Layer

A single persistent WebSocket at /ws/status handles most frontend-backend communication. Handlers live in the MESSAGE_HANDLERS map in server/routers/websocket.py plus plugin-registered handlers that self-wire through services.ws_handler_registry (each self-contained plugin folder can own its own WebSocket commands). Together they cover: node parameters, tool schemas, node execution, triggers, dead letter queue, deployment, AI operations, API keys, OAuth flows (Claude, Twitter, Google), Android, WhatsApp, Telegram, workflow storage, chat messages, console logs, skills, memory, user settings, pricing, agent teams, and model registry.

Request/Response Pattern

Frontend                                Backend
   |                                       |
   |-- {type: "execute_node", data: ...}-->|
   |                                       |-- dispatch to handle_execute_node()
   |                                       |-- run node handler
   |                                       |
   |<-- {type: "node_status", ... }---------|  (broadcast)
   |<-- {type: "node_output", ... }---------|  (broadcast)
   |<-- {type: "execute_node_response", ...}|  (direct reply, request_id matched)

Broadcast Message Types

node_status         workflow_status      api_key_status
node_output         android_status       token_usage_update
variable_update     compaction_starting  node_parameters_updated
variables_update    compaction_completed

Auto-Reconnect and Keepalive

Frontend sends {"type": "ping"} every 30 seconds. Reconnection uses PartySocket with jittered exponential backoff and message replay, so a dropped socket recovers automatically and queued requests replay on the new connection. Connection is gated on isAuthenticated so logged-out users never connect.

Cache, Persistence, and Security

Cache Fallback Hierarchy

OpenCompany follows the n8n cache pattern with automatic environment-based fallback:

Production:  Redis  ->  SQLite (cache_entries)  ->  in-memory dict
Local dev:          SQLite (cache_entries)  ->  in-memory dict
                    (Redis disabled via REDIS_ENABLED=false)

CacheService in server/core/cache.py checks each backend in order. TTL expiration is supported across all three. A background cleanup_expired_cache() task removes expired SQLite rows.

Encrypted Credentials

API keys and OAuth tokens live in a separate SQLite database (credentials.db) isolated from the main workflow database (workflow.db). Encryption uses Fernet (AES-128-CBC + HMAC-SHA256) with keys derived from a server-scoped config key via PBKDF2HMAC (600K iterations, OWASP 2024 recommendation).

API_KEY_ENCRYPTION_KEY (.env)  +  salt (from credentials.db, 256 bits)
                    |
                    v
           PBKDF2HMAC-SHA256
           (600,000 iterations)
                    |
                    v
        urlsafe_b64encode -> Fernet key
                    |
                    v
           held in memory only
                    |
                    v
      encrypt() / decrypt() inside EncryptionService

Two separate credential systems inside credentials.db:

API key system (EncryptedAPIKey table): provider keys the user enters manually
OAuth token system (EncryptedOAuthToken table): tokens from OAuth flows (Google, Twitter, Claude.ai)

All routers access credentials through AuthService only. Direct database access is forbidden. Decrypted values are cached in AuthService memory dicts and never written to disk or Redis. For cloud deployments, CREDENTIAL_BACKEND can be switched to keyring (OS-native) or aws (Secrets Manager) via the CredentialBackend abstraction.

Authentication

JWT in HttpOnly cookies following the n8n pattern. Two modes:

AUTH_MODE=single - first user becomes owner, registration disabled after
AUTH_MODE=multi - open registration for cloud deployments
VITE_AUTH_ENABLED=false - bypass login entirely for local development

WebSocket connections check the JWT cookie before accepting. Auth context has exponential-backoff retry (5 attempts) to survive race conditions where the frontend starts before the backend is ready.

Node Catalog

Browse all workflow nodes by category

AI Models

12 LLM providers with native SDK and LangChain paths

AI Agents

16 specialized agents with teams and delegation

GitHub

Source code and in-repo docs-internal/ deep dives

​Architecture

​System Overview

100+ Workflow Nodes

12 LLM Providers

16 Specialized Agents

WebSocket-First

Built-in Skills

Three Execution Modes

​Execution Engine

​Conductor’s Decide Pattern

​Three Execution Modes

​Layer Computation via Kahn’s Algorithm

​Result Caching, Recovery, and DLQ

​Edge Conditions

​Event-Driven Deployment

​Push vs Polling Triggers

​AI Agent System

​12 LLM Providers

​Dual-Path Architecture

​16 Specialized Agents

​Agent Teams Topology

​Agent Loop Flow

​Memory, Skills, Tokens, and Cost

​Markdown-Based Memory

​Skill System

​Token Tracking and Compaction

​Communication Layer

​Request/Response Pattern

​Broadcast Message Types

​Auto-Reconnect and Keepalive

​Cache, Persistence, and Security

​Cache Fallback Hierarchy

​Encrypted Credentials

​Authentication

​Related

Node Catalog

AI Models

AI Agents

GitHub

Architecture

System Overview

Execution Engine

Conductor’s Decide Pattern

Three Execution Modes

Layer Computation via Kahn’s Algorithm

Result Caching, Recovery, and DLQ

Edge Conditions

Event-Driven Deployment

Push vs Polling Triggers

AI Agent System

12 LLM Providers

Dual-Path Architecture

16 Specialized Agents

Agent Teams Topology

Agent Loop Flow

Memory, Skills, Tokens, and Cost

Markdown-Based Memory

Skill System

Token Tracking and Compaction

Communication Layer

Request/Response Pattern

Broadcast Message Types

Auto-Reconnect and Keepalive

Cache, Persistence, and Security

Cache Fallback Hierarchy

Encrypted Credentials

Authentication

Related