开源MCP工具

System Requirements

Synkora runs ~20 Docker containers. The stack includes Elasticsearch (2 GB hard cap), Redis (2 GB configured), a sentence-transformers ML service, Playwright-based scraper, ClickHouse, and multiple Celery workers. Size accordingly.

Required Software

The ./install.sh script checks all of the above automatically and will offer to install Docker if it's missing.

Prerequisites

• Docker and Docker Compose v2
• Node.js 20+ and pnpm (for local frontend development)
• Python 3.11+ and uv (for local backend development)

Install dependencies (using uv)

uv sync

Install dependencies

pnpm install

Adding Dependencies

Backend:

cd api
uv add package-name              # Production dependency
uv add --dev package-name        # Development dependency

Frontend:

cd web
pnpm add package-name            # Production dependency
pnpm add -D package-name         # Development dependency

What You Can Build

• Customer support agents - 24/7 query handling from your knowledge base, with HITL escalation to humans
• Engineering assistants - Code review, bug triage, and documentation via GitHub, GitLab, and Sentry
• Marketing automation - Content drafting, campaign analysis, and social scheduling
• Data agents - Natural language queries over your databases, automated reports, anomaly detection
• Internal copilots - HR, finance, and ops agents connected to your existing tools via 50+ integrations
• Personal AI assistants - General-purpose assistants with voice, browser control, and messaging integrations

One-Line Install (Recommended)

Run the interactive installer — it handles everything: prerequisite checks, .env generation, database migrations, seeding, and starting all services.

curl -fsSL https://raw.githubusercontent.com/getsynkora/synkora-ai/main/get.sh | bash

Installs into ~/synkora-ai by default. To choose a different directory:

curl -fsSL https://raw.githubusercontent.com/getsynkora/synkora-ai/main/get.sh | \
  SYNKORA_INSTALL_DIR=~/my-synkora bash

For CI/CD and server deployments (no prompts):

curl -fsSL https://raw.githubusercontent.com/getsynkora/synkora-ai/main/get.sh | \
  SYNKORA_ADMIN_EMAIL=admin@example.com \
  SYNKORA_ADMIN_PASSWORD=securepass123 \
  SYNKORA_LLM_PROVIDER=openai \
  SYNKORA_LLM_API_KEY=sk-... \
  bash -s -- --non-interactive

The installer will: 1. Check system resources (RAM, CPU, disk) and warn if below minimums 2. Verify Docker, openssl, Node.js/pnpm are present (and offer to install Docker if missing) 3. Detect existing installations and offer Upgrade / Reset / Quit 4. Collect admin account details, LLM provider key, and optional Slack bot tokens 5. Generate all .env files with secure random secrets 6. Pull images, start services, run migrations, seed plans/roles/template agents 7. Print a summary with all URLs and management commands

Manual Setup (Docker Compose)

If you prefer to set up manually:

```bash

Install act (macOS)

brew install act

Deployment

Docker Compose

```bash

Build all images

docker-compose build

Install

helm install synkora ./helm/synkora

Uninstall

helm uninstall synkora ```

Development Setup

See the Quick Start section for development environment setup instructions.

Quick Start

1. Copy and edit environment files

cp api/.env.example api/.env

Edit api/.env with your configuration

Set up environment

cp .env.example .env

Edit .env with your configuration

Set up environment

cp .env.example .env.local

Edit .env.local if needed

Configuration

Environment Variables

#### Backend (api/.env) ```bash

LLM Providers (optional)

OPENAI_API_KEY=sk-... ANTHROPIC_API_KEY=sk-ant-...

Langfuse (optional)

LANGFUSE_PUBLIC_KEY=pk-... LANGFUSE_SECRET_KEY=sk-... LANGFUSE_HOST=http://localhost:3001

#### Frontend (`web/.env.local`)

Run API tests locally

act workflow_dispatch -W .github/workflows/api-tests.yml -j test --matrix python-version:3.11 -P ubuntu-latest=catthehacker/ubuntu:act-latest --container-architecture linux/amd64 ```

CI/CD Pipeline

Our CI/CD pipeline runs on every push and pull request:

Comparison vs Other Platforms

Research methodology: Each platform was investigated against its official docs, GitHub issues/PRs, and changelogs as of April 2026. Every cell is sourced — see footnotes. Legend: ✅ Native built-in feature · ⚡ Partial (manual setup or limited scope) · ❌ Not supported

Footnotes

<a name="fn1"></a>¹ Dify prompt caching — Anthropic only, via official plugin v0.3.10 (Apr 2026). Six opt-in per-call parameters (prompt_caching_system_message, prompt_caching_tool_definitions, etc.) in the LLM node UI. OpenAI prefix caching is transparent at provider level (Dify does nothing special). Gemini explicitly declined as "not planned" (issue #2121). Known bug: Haiku 4.5 cache shows 0% (issue #1946).

<a name="fn2"></a>² Flowise prompt caching — No cache_control implementation. Open feature request #4634 with no maintainer commitment. GitHub code search returns 0 matches for cache_control in the repo.

<a name="fn3"></a>³ LangFlow prompt caching — No native support. A LangFlow blog post titled "Prompt Caching in LLMs" is educational only — it describes no LangFlow feature. LangChain's AnthropicPromptCachingMiddleware exists but is not exposed as a built-in LangFlow node.

<a name="fn4"></a>⁴ n8n prompt caching — Anthropic cache_control (4-breakpoint strategy) was merged for the internal AI Workflow Builder (PR #20484, Oct 2025). The user-facing Anthropic Chat Model node PR (#22318) was still open as of Apr 2026.

<a name="fn5"></a>⁵ Dify response cache — No generic Redis-backed LLM output cache. The nearest equivalent is Annotation Reply: a manually curated Q&A store with vector-similarity matching that short-circuits LLM calls. A workflow-level caching feature request (#23598) was closed "not planned" Sep 2025.

<a name="fn6"></a>⁶ Flowise response cache — Documented, shipped feature. LangChain cache layer with four backends: InMemory, Redis, Upstash Redis, Momento. Connected to LLM nodes as an optional "Cache" input. (docs)

<a name="fn7"></a>⁷ LangFlow response cache — No LLM output deduplication cache exists. The in-memory cache is for internal flow-graph state (component outputs within a run), not LLM response deduplication. Redis is documented as "experimental" for flow caching only.

<a name="fn8"></a>⁸ n8n response cache — No first-class cache node. Achievable via Redis Vector Store node + manual workflow logic (community template), but requires authoring, not a one-click setting.

<a name="fn9"></a>⁹ Dify / Flowise / n8n model routing — All three provide routing primitives (If/Else nodes, Condition nodes, Switch nodes) that can be manually wired to route to different models. None has a built-in automatic complexity classifier that routes to a cheaper model. n8n community has a workflow template using OpenRouter for this.

<a name="fn10"></a>¹⁰ LangFlow model routing — Shipped a real LLM Router / LLM Selector component in v1.7 (PR #5475, Jan 2025) that uses OpenRouter's model-spec API and a judge LLM to pick among attached models by quality/speed/cost/balanced. Requires OpenRouter; incurs a live LLM judgment call per routing decision.

<a name="fn11"></a>¹¹ Dify batch API — Explicitly declined as "not planned" (issue #13126, closed Mar 18, 2025).

<a name="fn12"></a>¹² Flowise batch API — No evidence of Anthropic Message Batches or OpenAI Batch API integration. The batchSize field in the OpenAI node is LangChain-internal concurrency (embedding chunking), not the provider batch endpoint. Async response confirmed "not currently supported" (discussion #1212).

<a name="fn13"></a>¹³ LangFlow batch API — No support. LangFlow's /v1/flows/batch/ endpoint is for creating/deleting multiple flow definitions, not for submitting LLM calls to provider batch APIs.

<a name="fn14"></a>¹⁴ n8n batch API — Anthropic batch API is accessible via an HTTP Request node using a community workflow template. No dedicated node. OpenAI Batch API requires the same workaround.

<a name="fn15"></a>¹⁵ Dify token & cost tracking — Per-node: execution_metadata.total_tokens, total_price (USD), currency in the node_finished SSE event (issue #8873). Visible in Run History / Debug view. Cache-aware: cache_creation_input_tokens (1.25×) and cache_read_input_tokens (0.1×) tracked in Anthropic plugin. Caveat: shows $0 for OpenAI-compatible third-party models without PLUGIN_BASED_TOKEN_COUNTING_ENABLED=true; fixed in v1.9.2+.

<a name="fn16"></a>¹⁶ Flowise token & cost tracking — Token data is generated internally but only surfaced via third-party observability integrations (Langfuse, LangSmith, LunaryAI, LangWatch, Arize, Phoenix, Opik). No native per-call USD cost display in Flowise UI. Non-OpenAI providers (Gemini, DeepSeek) have incomplete cost data in traces (Langfuse issue #8293).

<a name="fn17"></a>¹⁷ LangFlow token & cost tracking — Token counts added to the inspection UI in v1.8 (blog) and available in the OpenAI Responses API path. USD cost requires Langfuse / LangSmith / LangWatch integration; not calculated natively. Earlier feature request (#3261) was closed "not planned" Nov 2024.

<a name="fn18"></a>¹⁸ n8n token & cost tracking — Token data is in the raw model output for providers that return it (OpenAI, Anthropic, Google) but not displayed natively in the n8n execution UI. Per-call tracking and a cost dashboard require a custom subworkflow (community template).

---