mcpgate

This repository contains the self-hosting distribution for mcpgate: Docker Compose, configuration templates, hooks, and operations docs. Published container images are released via the CI/CD pipeline connected to this repository.

A PM finishes a user interview and asks Claude to consolidate his notes in Notion. (Works the same with ChatGPT, Codex, or any MCP-compatible agent.) After reviewing them, he saves the key takeaways to the insights database and frames an opportunity for the next product meeting. What used to take the rest of the day is done in 15 minutes.

Weeks later, the product team decides to prioritize that opportunity. The PM gives the AI the full context, adds constraints, and starts prototyping. The AI pulls the codebase, scaffolds a working prototype, and the PM iterates on the actual problem — not on tooling. A few hours later, the prototype integrates with the existing app and the design system, because the AI had the context to do it right. Changes are saved to a Git branch automatically.

With all that context loaded, the AI drafts Jira tickets for the refinement. Hooks handle the boring parts — converting Markdown to Jira's ADF format, enforcing required fields, and blocking accidental overwrites. When the team meets, they walk through a working prototype, identify gaps, and make it actionable. Design, development, QA — everyone picks up where the last person left off, with full context.

mcpgate connects your tools to your AI — Notion, Jira, GitLab, Figma, HubSpot, Pipedrive, Windmill, and many more. 40 integrations are built in, and you can add your own through OpenAPI import. Company hooks enforce your policies, while user hooks let individuals fine-tune rules directly from their AI client — hot-reloaded in seconds. mcpgate works as an MCP gateway, but also as a gate: your rules, your data. Eliminate loops between teams, safely manage context across handoffs, and let your team focus on building.

AI transformation is happening. Your tools, your data, and your context need to be connected — mcpgate is how you do it on your terms.

docker compose up -d open http://localhost:8642

That's it. No.envfile needed. The setup wizard walks you through login, branding, team, and connecting services. Secrets are auto-generated on first start.

New here?Clone the repo to get the pre-configureddocker-compose.yml:

git clone https://gitlab.com/mcpgate/mcpgate.git && cd mcpgate

Already have an.env?It still works — environment variables take priority over wizard config.

After setup, connect your AI client from the dashboard:

Name: mcpgate URL: https://your-gateway-url/mcp

claude mcp add mcpgate https://your-gateway-url/mcp -s user -t http

Settings → Apps → Add App → OAuth → enter your MCP URL.

codex mcp add mcpgate --url https://your-gateway-url/mcp gemini mcp add --transport http mcpgate https://your-gateway-url/mcp

CI/CD for what your AI knows about your company. A curated Markdown corpus — which systems and repos exist, who owns what, how things connect — served from the gateway to every connected AI client. The gateway syncs and serves it; AI runs on the consumers, not on the map itself.

- Anchor— point the gateway at a git repo (audited, diffable, revertable) or paste content inline (zero git). A configurable poll keeps it fresh; an optional webhook with a per-tenant secret triggers immediate reindex. A content-push REST endpoint lets other systems POST entries directly.
- map_read— a new MCP tool every connected AI client gets. BM25 search over the page index; fetch a single page by id. A one-line session-hint nudges the model to consult the map forwhere does X live / who owns it / how does it connectquestions instead of guessing.
- Freshness rides every tool response— a tiny version-key compare on each tool call; on change, the tag-scoped delta attaches to the tool's response once per session per change. Covers passthrough tools too (Jira, Slack, Notion, GitLab, …), so a working session notices a mid-session map change on its very next call — no client interrupt, no re-send.
- Gap feed—map_readqueries that come back empty are logged to a bounded, de-duplicated, PII-scrubbed feed with frequency aggregation. The raw material for proposing new pages; nothing is written automatically.
- map_write— AI agents can correct stale pages directly. Every write (inline edit, content-push REST,map_write) passes the same gate: leak-scan + size/format validation + attribution. In repo mode the gateway commits and pushes through its own identity; the served copy reindexes immediately.
- Operator-only by default— themap_read/map_writetools and the freshness beacon are hidden fromexternal/viewerroles. The admin viewer is XSS-hardened; git credential is tokenless at rest; only allowlisted git hosts accepted; symlink escapes refused and logged.

The whole point: shared organisational knowledge that personal AI memory can't solve, in plain text you own, kept fresh because the gateway notices when it gets used and what it misses.

Built-in safeguards that don't need configuration:

- PII Sanitization with Pseudonym Rehydration— sensitive data (emails, names, phone numbers) is replaced with stable pseudonyms before it reaches the LLM, then rehydrated when the agent calls a tool. Mapping stays on-prem, encrypted at rest, and expires after 24h. Preserves write-flows that simple redaction would break.
- Write-Safety Defaults— destructive actions (delete, archive, dashboard PUTs) require explicitconfirmed=trueorforce=true. Response size caps prevent accidental mass operations.
- Stores nothing in transit— mcpgate is a pass-through. Tool actions are auditable in your own tools (Jira, GitLab, Slack) where they happen. The only data we hold is the encrypted pseudonym mapping for PII rehydration, with a 24-hour TTL.
- Highly available— runs as multiple replicas behind your load balancer. Config changes propagate to all replicas in seconds.

A quick read against three named neighbors (figures verified 2026-05-17 via GitHub API):

The ❌ cells above are about what each projectships out of the box, not an architectural ceiling — Obot, Docker MCPG, and ContextForge are all open enough that any of those features can be built on top of them with engineering investment. The trade-off is who does the engineering and who carries the maintenance. Each comparison page on the website walks through that trade-off explicitly.

Detailed honest comparisons live on the website:

Where another project is the better fit for your team, we say so.

Policy and enrichment hooks inconfig/tool_hooks.yaml:

- Policy(validation): destructive action confirmation, API endpoint guards, transition checks
- Enrichment(mutation): Markdown → ADF conversion, text normalization, auto-linking, templates
- Post-processing(observability): response capping, cross-service automation, auth error handling

curl -X POST http://localhost:8642/admin/reload

Branding, access control, and hooks are configurable through the setup wizard or config files. White-label the dashboard with your company name, logo, and colors.

docker compose pull docker compose up -d

For advanced configuration, create a.envfile from the template:

See.env.examplefor all available options including OIDC, service credentials, AI features, and error reporting.

Humanizer PRO is an MCP server that transforms AI-generated text into natural, human-sounding content. It provides 4 tools: - humanize_text: Rewrite AI text to bypass detectors like GPTZero, Turnitin, Originality.ai, Copyleaks, and ZeroGPT. Three modes: Stealth (highest bypass rate), Academic (Turnitin-optimized), SEO (marketing content). - scan_ai_detection: Analyze text for AI patterns. Returns AI probability score, human-likeness percentage, and verdict. - check_word_balance: Check remaining word credits and subscription plan details. - get_subscription_plans: Browse plans - Free (500 words), Starter ($9.99/mo, 30K words), Creator ($14.99/mo, 100K words), Pro Annual ($119.88/yr, 100K words/mo). Authentication: OAuth 2.0. Works with ChatGPT, Claude, Cursor, and all MCP-compatible clients.

Task/project MCP server with OAuth for Claude, Cursor bridge, semantic search, staged write approval

Give Claude, ChatGPT, Claude Code, and other AI assistants secure access to your Wave meeting transcripts, summaries, and semantic search — with one URL and OAuth.

13-tool MCP server for AI agent authorization. Manage agents, grants, tokens, and audit logs from Claude Desktop, Cursor, or Windsurf. Plus @grantex/mcp-auth for adding OAuth

Ask Claude if any AI provider is down — a hosted MCP server with real-time status, incident history, and 30-day uptime for 75+ AI services, no auth.

Security-first WordPress MCP server for Claude, ChatGPT, and Gemini. API key + OAuth 2.1 auth, per-tool capability gating, full activity log. 127 tools covering posts, pages, media, WooCommerce, Elementor, ACF. Free on wp.org.

Ask Claude or ChatGPT about your own credit cards: balances, best card for a purchase, missed rewards. Read only, OAuth 2.1.

AI traffic control plane (chaos governor): Redis prompt replay, compliant web ingest, SSO org ledger, Agent Shell. BYOK OpenAI-compatible ingress. Cursor optional; MCP is a compatibility client.