workbench MCP server
Commit Nov 14, 2026Aggregates multiple SaaS integrations into a single MCP endpoint using meta-tools to maintain a consistent tool list for AI agents.
Also known as amarthaid/workbench
- Stars
- 7
- Tools
- —
- Downloads, 30 days
- —
- Latest version
- —
- Last commit
- Nov 14, 2026
- License
- MIT
What is workbench MCP?
workbench aggregates multiple SaaS integrations into a single MCP endpoint using meta-tools to maintain a consistent tool list for AI agents.
Summary generated from the project's README
About workbench MCP server
Workbench acts as a self-hosted aggregator that consolidates multiple SaaS integrations into a single MCP endpoint. It is designed for developers who want to provide AI agents with access to various services without overwhelming them with massive tool lists. Instead of connecting to sixteen different servers, an agent connects to one endpoint and interacts with 194 tools through nine specific meta-tools. This approach keeps the tool list size consistent regardless of how many integrations are active.
Users set up the server locally using Node.js and manage connections through a React portal. The system handles per-user OAuth connections and encrypts all tokens at rest. It supports various authentication modes including oauth2, apikey, cookie, and none. While agents use the JSON-RPC MCP interface, other scripts can bypass MCP to run tools via plain JSON through REST endpoints, which avoids the 60,000-character result limit.
What you can do with it
- 1Simplify agent tool managementProvide an AI agent with access to many services through a single connection point.
- 2Maintain stable agent contextsEnsure the number of tools presented to an AI remains constant even as more integrations are added.
- 3Securely manage third-party credentialsStore and use encrypted OAuth tokens for different service providers.
Written from the project's README. Every use case is checked against a quote from it.
Automated checks
Ran Nov 15, 2026Full report →Not verified yet. A server needs a score of 80, no serious finding, and enough data for us to check. How checks work · What to fix for more points
- Source repository linked and reachableamarthaid/workbench is public and reachablegithub.com/amarthaid/workbench ↗Pass
- Package names this repository as its sourceNot checked
- Build provenance (Sigstore attestation)Not checked
- Open source license (OSI)MITspdx.org/licenses/MIT.html ↗Pass
- Commit or release in the last yearLast commit or release 2026-11-08github.com/amarthaid/workbench ↗Pass
- OpenSSF ScorecardNot checked
- Package name not close to a popular packageNot checked
- No critical advisory in the package or its dependenciesNot checked
- No hidden instructions to the model0 tool descriptions and the README read; no override phrases, credential paths or hidden charactersPass
- Destructive tools are labelledNot checked
- Remote endpoint answers the MCP handshakeNot checked
README
View on GitHubSelf-hosted MCP tool aggregator. One endpoint, per-user OAuth, 194 tools across 16 integrations — behind 9 meta-tools.
Website · Documentation · Quickstart · Integrations · Build a plugin · Deploy
Give an agent one MCP endpoint instead of sixteen. workbench holds a separate OAuth connection per user per provider, encrypts every token at rest, and exposes every integration through a fixed set of nine meta-tools — so the agent's tool list stays the same size whether one integration is connected or all of them.
flowchart LR
Agent["Agent"] -->|JSON-RPC| MCP["POST /mcp"]
Script["Script / cron"] -->|plain JSON| REST["POST /rest/:integration"]
MCP --> Meta["9 meta-tools"]
Meta --> Reg["Plugin registry<br/>16 integrations · 194 tools"]
REST --> Reg
Reg -->|credential injected| APIs["Jira · GitHub · Slack · Google · …"]
Portal["Portal"] --> Store[("Encrypted tokens")]
Meta -.-> Store
Anything that is not an agent can skip MCP entirely:
POST /rest/:integration runs the
same tools over plain JSON, on the same credentials, with no JSON-RPC framing and no
60,000-character result cap.
Quickstart
git clone https://github.com/barockok/workbench.git && cd workbench
npm install
cp .env.example .env
# Both are required — the server refuses to boot without them.
echo "ENCRYPTION_KEY=$(openssl rand -hex 32)" >> .env
echo "SESSION_SECRET=$(openssl rand -base64 32)" >> .env
npm run dev
Then open the portal, connect an integration, mint an API key, and point your MCP
client at http://localhost:3000/mcp. The full walkthrough is in the
Quickstart.
Documentation
The docs are the product surface — start there, not here.
| Section | What's in it |
|---|---|
| Get started | What it is, how it works, connecting an agent |
| Guides | Discovering and executing tools, OAuth, browser sessions, raw API calls, troubleshooting |
| Integrations | Every provider: exact scopes, setup steps, full tool list |
| Build plugins | Manifest reference, plugin context API, the four auth modes |
| Deploy | Docker, PostgreSQL, portal SSO, security, observability, releases |
| Reference | All 9 meta-tools, every HTTP route, every environment variable, the tool catalog |
| Field notes | Production failures, root causes, and what changed |
What's in the box
| Integrations | 16 on disk, plus 2 internal (browser, jots) |
| Tools | 194 plugin tools, reached through 9 meta-tools |
| Auth modes | oauth2, apikey, cookie, none |
| Agent auth | Workbench API key or OAuth 2.1 (dynamic registration + PKCE) |
| Portal SSO | Google, Keycloak, or both |
| Database | SQLite or PostgreSQL, with a migration path between them |
| Stack | TypeScript, Fastify, MCP TypeScript SDK, React portal |
Development
npm run dev # start dev servers
npm run test # run tests
npm run build # build all packages
npm run lint # lint all packages
The docs site is generated from Markdown by a script with no framework:
node docs/site/build.mjs # _content/*.md + nav.json → static HTML in docs/site/_site/
Commit the Markdown only. docs/site/_site/ is build output, is not tracked,
and is regenerated by the Docs workflow on every push to main. CI runs the same
build on every pull request, and a broken internal link fails it.
See Contributing for the branch, commit, and release conventions, and SECURITY.md for reporting a vulnerability.
License
workbench MCP questions
How does workbench handle multiple integrations for an agent?
It exposes every integration through a fixed set of nine meta-tools, ensuring the agent's tool list remains the same size regardless of how many integrations are connected.
What authentication methods are supported for plugins?
The platform supports four authentication modes: oauth2, apikey, cookie, and none.
What databases can I use with workbench?
You can use either SQLite or PostgreSQL, and there is a migration path available between them.