Leveraging Gemini CLI and the underlying Gemini LLM to build Model Context Protocol (MCP) AI applications in COBOL deployed to Google Cloud Run.
Why not just use Python?
Python has traditionally been the main coding language for ML and AI tools. One of the strengths of the MCP protocol is that the actual implementation details are independent of the development language. The reality is that not every project is coded in Python- and MCP allows you to use the latest AI approaches with other coding languages.
COBOL? For AI? The 1960’s called and they want their Mainframe back!
The goal of this article is to provide a minimal viable basic working MCP HTTP server in COBOL that can be run locally without any unneeded extra code or extensions. Next- this HTTP server will be staged and deployed to Google Cloud Run.
So how does this Sorcery Work?
It is theoretically possible to implement the stack needed with only valid COBOL language constructs. The key to this solution is that modern compiled COBOL supports importing bindings from C during the compilation process.
A fully functional C MCP library is here:
GitHub - micl2e2/mcpc: Cross-platform C SDK for Model Context Protocol (MCP), in modern🚀 C23.
What Is COBOL?
COBOL (COmmon Business-Oriented Language) is a high-level, English-like programming language developed in the 1950s, designed for business, finance, and administrative systems, especially on mainframe computers, handling large-scale data processing. Despite its age, it remains crucial for critical infrastructure like banking and government systems, powering applications that process massive transaction volumes, due to its reliability and self-documenting syntax.
There are some legacy sites at IBM, but for the modern compiler version- here is the GNU COBOL site:
Installing COBOL
The step by step instructions vary by platform- for a basic Debian system here are the steps:
sudo apt update
sudo apt install build-essential
sudo apt install gcc
sudo apt install gnucobol4
cobc -V
xbill@penguin:~$ cobc -V
cobc (GnuCOBOL) 4.0-early-dev.0
Copyright (C) 2020 Free Software Foundation, Inc.
License GPLv3+: GNU GPL version 3 or later <https://gnu.org/licenses/gpl.html>
This is free software; see the source for copying conditions. There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
Written by Keisuke Nishida, Roger While, Ron Norman, Simon Sobisch, Edward Hart
Built Sep 19 2022 04:28:17
Packaged Jun 06 2020 20:56:36 UTC
C version "12.2.0"
xbill@penguin:~$
What is Firestore?
Google Firestore, also known as Cloud Firestore is a part of the Google Firebase application development platform. It is fundamentally a cloud-hosted NoSQL database for storing and syncing data. Firestore can be directly accessed by mobile and web applications through native SDKs.
Gemini CLI
If not pre-installed you can download the Gemini CLI to interact with the source files and provide real-time assistance:
npm install -g @google/gemini-cli
Testing the Gemini CLI Environment
Once you have all the tools and the correct Node.js version in place- you can test the startup of Gemini CLI. You will need to authenticate with a Key or your Google Account:
gemini
Node Version Management
Gemini CLI needs a consistent, up to date version of Node. The nvm command can be used to get a standard Node environment:
C MCP Documentation
The official MCP C page provides samples and documentation for getting started:
GitHub - micl2e2/mcpc: Cross-platform C SDK for Model Context Protocol (MCP), in modern🚀 C23.
This C library is built and the low level bindings can be used from COBOL code.
Where do I start?
The strategy for starting MCP development with COBOL is a incremental step by step approach.
First, the basic development environment is setup with the required system variables, and a working Gemini CLI configuration.
Then, a minimal Hello World Style COBOL MCP Server is built with HTTP transport using the C library bindings. This server is validated with Gemini CLI in the local environment.
This setup validates the connection from Gemini CLI to the local process via MCP. The MCP client (Gemini CLI) and the MCP server both run in the same local environment.
Next- the MCP server is built and deployed to Google Cloud Run. This remote endpoint is validated with a local copy of Gemini CLI.
Setup the Basic Environment
At this point you should have a working COBOL and C environment and a working Gemini CLI installation. The next step is to clone the GitHub samples repository with support scripts:
cd ~
git clone https://github.com/xbill9/gemini-cli-codeassist
Then run init.sh from the cloned directory.
The script will attempt to determine your shell environment and set the correct variables:
cd gemini-cli-codeassist
source init.sh
If your session times out or you need to re-authenticate- you can run the set_env.sh script to reset your environment variables:
cd gemini-cli-codeassist
source set_env.sh
Variables like PROJECT_ID need to be setup for use in the various build scripts- so the set_env script can be used to reset the environment if you time-out.
Hello World with HTTP Transport
One of the key features that the standard MCP libraries provide is abstracting various transport methods.
The high level MCP tool implementation is the same no matter what low level transport channel/method that the MCP Client uses to connect to a MCP Server.
The simplest transport that the SDK supports is the stdio (stdio/stdout) transport — which connects a locally running process. Both the MCP client and MCP Server must be running in the same environment.
The HTTP Transport allows the MCP client and server to be in the same environment or distributed over the Internet.
The connection over HTTP will look similar to this:
PROCEDURE DIVISION.
MAIN-LOGIC.
CALL "mcpc_server_new_tcp" RETURNING server-ptr.
IF server-ptr = NULL THEN
CALL "log_info_c" USING BY REFERENCE err-srv-new
STOP RUN RETURNING 1
END-IF.
Package Information
The code depends on several standard libraries for MCP and logging:
IDENTIFICATION DIVISION.
PROGRAM-ID. greet_cb_impl.
DATA DIVISION.
WORKING-STORAGE SECTION.
01 log-msg PIC X(30) VALUE Z"Executed greet tool".
01 err-tool-null PIC X(30) VALUE Z"Error: tool-ptr is NULL".
01 err-msg PIC X(30) VALUE Z"Error retrieving param".
Installing and Running the Code
Run the install make release target on the local system:
xbill@penguin:~/gemini-cli-codeassist/firestore-https-cobol$ make
cc -std=c17 -Wall -Wextra -Imcpc/src -Imcpc -Dis_unix -DMCPC_C23PTCH_KW1 -DMCPC_C23PTCH_CKD1 -DMCPC_C23PTCH_UCHAR1 -DMCPC_C23GIVUP_FIXENUM -O2 -c c_helpers.c
cc -std=c17 -Wall -Wextra -Imcpc/src -Imcpc -Dis_unix -DMCPC_C23PTCH_KW1 -DMCPC_C23PTCH_CKD1 -DMCPC_C23PTCH_UCHAR1 -DMCPC_C23GIVUP_FIXENUM -O2 -c cob_helpers.c
cc -std=c17 -Wall -Wextra -Imcpc/src -Imcpc -Dis_unix -DMCPC_C23PTCH_KW1 -DMCPC_C23PTCH_CKD1 -DMCPC_C23PTCH_UCHAR1 -DMCPC_C23GIVUP_FIXENUM -O2 -c firestore_client.c
make -C mcpc/src
make[1]: Entering directory '/home/xbill/gemini-cli-codeassist/firestore-https-cobol/mcpc/src'
cc -Dis_unix -std=c17 -DMCPC_C23PTCH_KW1 -DMCPC_C23PTCH_CKD1 -DMCPC_C23PTCH_UCHAR1 -DMCPC_C23GIVUP_FIXENUM -Wall -Wextra -Werror -Wno-unused-function -Wno-unused-parameter -Wno-unused-label -Wno-error=unused-variable -Wno-error=unused-but-set-variable -O2 -Os -I.. -fPIC alloc.c log.c errcode.c anydata.c tool.c rsc.c prmpt.c server.c retbuf.c ucbr.c complt.c serlz.c mjson.c -c
ar rcs libmcpc.a alloc.o log.o errcode.o anydata.o tool.o rsc.o prmpt.o server.o retbuf.o ucbr.o complt.o serlz.o mjson.o
cc -s -o libmcpc.so alloc.o log.o errcode.o anydata.o tool.o rsc.o prmpt.o server.o retbuf.o ucbr.o complt.o serlz.o mjson.o -shared ../src/libmcpc.a
make[1]: Leaving directory '/home/xbill/gemini-cli-codeassist/firestore-https-cobol/mcpc/src'
cobc -x -free -o server-cobol server.cob c_helpers.o cob_helpers.o firestore_client.o mcpc/src/libmcpc.a -lm
To lint the code:
xbill@penguin:~/gemini-cli-codeassist/firestore-https-cobol$ make lint
cc -std=c17 -Wall -Wextra -Imcpc/src -Imcpc -Dis_unix -DMCPC_C23PTCH_KW1 -DMCPC_C23PTCH_CKD1 -DMCPC_C23PTCH_UCHAR1 -DMCPC_C23GIVUP_FIXENUM -Wpedantic -Wshadow -Wpointer-arith -Wcast-qual -Wstrict-prototypes -Wmissing-prototypes -fsyntax-only c_helpers.c
cc -std=c17 -Wall -Wextra -Imcpc/src -Imcpc -Dis_unix -DMCPC_C23PTCH_KW1 -DMCPC_C23PTCH_CKD1 -DMCPC_C23PTCH_UCHAR1 -DMCPC_C23GIVUP_FIXENUM -Wpedantic -Wshadow -Wpointer-arith -Wcast-qual -Wstrict-prototypes -Wmissing-prototypes -fsyntax-only cob_helpers.c
cc -std=c17 -Wall -Wextra -Imcpc/src -Imcpc -Dis_unix -DMCPC_C23PTCH_KW1 -DMCPC_C23PTCH_CKD1 -DMCPC_C23PTCH_UCHAR1 -DMCPC_C23GIVUP_FIXENUM -Wpedantic -Wshadow -Wpointer-arith -Wcast-qual -Wstrict-prototypes -Wmissing-prototypes -fsyntax-only firestore_client.c
xbill@penguin:~/gemini-cli-codeassist/firestore-https-cobol$
To test the code:
xbill@penguin:~/gemini-cli-codeassist/firestore-https-cobol$ make test
python3 test_server_http.py
Starting server on port 8081...
Sending initialize...
Initialize response: {
"jsonrpc": "2.0",
"id": 1,
"result": {
"protocolVersion": "2024-11-05",
"serverInfo": {
"name": "firestore-https-cobol",
"version": "0.1"
},
"capabilities": {
"tools": {
"listChanged": false
}
}
}
}
Gemini CLI settings.json
In this example — the source code uses a compiled binary that can be called directly from Gemini CLI.
The default Gemini CLI settings.json has an entry for the source:
{
"mcpServers": {
"firestore-https-cobol": {
"httpUrl": "http://127.0.0.1:8080"
}
}
}
Start the Local HTTP server
Open a new terminal window, and start the Local copy of the the HTTP MCP server:
xbill@penguin:~/gemini-cli-codeassist/firestore-https-cobol$ ./server-cobol
{"asctime": "2026-01-30T16:06:12Z", "name": "root", "levelname": "INFO", "message": "Resolving project ID..."}
{"asctime": "2026-01-30T16:06:15Z", "name": "root", "levelname": "INFO", "message": "Refreshing access token..."}
{"asctime": "2026-01-30T16:06:16Z", "name": "root", "levelname": "INFO", "message": "Firestore GET request to /inventory"}
{"asctime": "2026-01-30T16:06:16Z", "name": "root", "levelname": "INFO", "message": "doc_to_json called"}
{"asctime": "2026-01-30T16:06:16Z", "name": "root", "levelname": "INFO", "message": "doc_to_json: success"}
Validation with Gemini CLI
Next- leave the local server running, and open a new terminal window. Start Gemini CLI and the MCP connection over local HTTP to the COBOL Code is validated. The full Gemini CLI Session will start:
> /mcp list
Configured MCP servers:
🟢 firestore-https-cobol - Ready (8 tools)
Tools:
- check_db
- get_product_by_id
- get_products
- get_root
- greet
- mcpc-info
- reset
- seed
> with the inventory suggest a snack for a hungry CICS programmer
✦ I will check the inventory database status and then list the available products to find a suitable snack.
╭──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╮
│ ✓ check_db (firestore-https-cobol MCP Server) {} │
│ │
│ Database running: true (Project: comglitn) │
╰──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯
✦ I will retrieve the list of products from the inventory.
✦ For a hungry CICS programmer who needs to keep their energy up while debugging legacy code, I recommend:
Peanut Butter and Jelly Cups
* Price: $9.00
* Quantity Available: 21
* ID: RT307lg10Zi5gOkK5dVj
It provides a good mix of protein and sugar for a quick boost. If you need some caffeine to go with it, there is also Cola ($7.00, ID:
IcQMnaZNHGIfkOXmWlv5) or Coffee Beans ($4.00, ID: EL2qcJF0vp6ziOhijtxd) available.
Deploying to Cloud Run
After the HTTP version of the MCP server has been tested locally — it can be deployed remotely to Google Cloud Run.
First- switch to the directory with the HTTP MCP sample code:
cd ~/gemini-cli-codeassist/firestore-https-cobol
Deploy the project to Google Cloud Run with the pre-built cloudbuild.yaml and Dockerfile:
cd ~/gemini-cli-codeassist/firestore-https-cobol
xbill@penguin:~/gemini-cli-codeassist/firestore-https-cobol$ make deploy
The Cloud Build will start:
xbill@penguin:~/gemini-cli-codeassist/firestore-https-cobol$ make deploy
echo "Submitting build to Google Cloud Build..."
Submitting build to Google Cloud Build...
gcloud builds submit . --config cloudbuild.yaml
Creating temporary archive of 91 file(s) totalling 405.2 KiB before compression.
Some files were not included in the source upload.
Check the gcloud log [/home/xbill/.config/gcloud/logs/2026.01.30/11.08.19.367148.log] to see which files and the contents of the
default gcloudignore file used (see `$ gcloud topic gcloudignore` to learn
more).
Uploading tarball of [.] to [gs://comglitn_cloudbuild/source/1769789299.523841-bbc23e1b9dc943f8bdb8a5246d6e2152.tgz]
Created [https://cloudbuild.googleapis.com/v1/projects/comglitn/locations/global/builds/4031b0ab-70e0-47ba-b869-b8be3021c549].
Logs are available at [https://console.cloud.google.com/cloud-build/builds/4031b0ab-70e0-47ba-b869-b8be3021c549?project=1056842563084].
Waiting for build to complete. Polling interval: 1 second(s).
------------------------------------------------------------- REMOTE BUILD OUTPUT --------------------------------------------------------------
starting build "4031b0ab-70e0-47ba-b869-b8be3021c549"
The cloud build needs to pull in all the COBOL and C libraries in the build environment and generate the entire package from scratch.
When the build is complete- an endpoint will be returned:
Starting Step #1
Step #1: Already have image (with digest): gcr.io/cloud-builders/gcloud
Step #1: Deploying container to Cloud Run service [firestore-https-cobol] in project [comglitn] region [us-central1]
Step #1: Deploying...
Step #1: Setting IAM Policy..............done
Step #1: Creating Revision.................................................................done
Step #1: Routing traffic.....done
Step #1: Done.
Step #1: Service [firestore-https-cobol] revision [firestore-https-cobol-00004-4nb] has been deployed and is serving 100 percent of traffic.
Step #1: Service URL: https://firestore-https-cobol-1056842563084.us-central1.run.app
Finished Step #1 │
The service endpoint in this example is :
https://firestore-https-cobol-1056842563084.us-central1.run.app
The actual endpoint will vary based on your project settings.
Review Service in Cloud Run
Navigate to the Google Cloud console and search for Cloud Run -
and then you can detailed information on the Cloud Run Service:
Cloud Logging
The remote server writes logs to stderr in standard JSON format. These logs are available from the deployed Cloud Run Service:
Validate HTTP connection
Once you have the Endpoint — you can attempt a connection- navigate to in your browser:
https://firestore-https-cobol-1056842563084.us-central1.run.app
You will need to adjust the exact URL to match the URL returned from Cloud Build.
You will get an error- this connection is expecting a message in the MCP format:
{"jsonrpc":"2.0","error":{"code":-32700,"message":""}}
Gemini CLI settings.json.cloudrun
Replace the default Gemini CLI configuration file — settings.json with a pre-configured sample- settings.json.cloudrun to use the Cloud Run version of the connection:
{
"mcpServers": {
"mcp-cloudrun-cobol": {
"httpUrl": "https://firestore-https-cobol-$PROJECT_NUMBER.us-central1.run.app"
}
}
}
Copy the Cloud Run version of the Gemini CLI configuration file:
xbill@penguin:~/gemini-cli-codeassist/firestore-https-cobol$ cd .gemini
cp settings.json.cloudrun settings.json
xbill@penguin:~/gemini-cli-codeassist/firestore-https-cobol/.gemini$
Validation with Gemini CLI
The final connection test uses Gemini CLI as a MCP client with the deployed Cloud Run Service in C providing the MCP server. Startup Gemini CLI with the updated settings :
> /mcp list
Configured MCP servers:
🟢 mcp-cloudrun-cobol - Ready (8 tools)
Tools:
- check_db
- get_product_by_id
- get_products
- get_root
- greet
- mcpc-info
- reset
- seed
> suggest a 1960's power lunch based on inventory for hardcore IBM 360 coders
✦ For the hardcore IBM 360 coder—the kind who hand-patches binaries and dreams in EBCDIC—lunch isn't just a meal; it's a refueling operation for
the next 12-hour coding marathon.
Here is a 1960s Power Lunch configuration optimized for maximum throughput and caffeine density:
The "Supervisor Call" (SVC) Special
* The "Batch Job" Entrée: Beef
* Item: Beef (ID: L2LORsCFh00SYWEcKBjA)
* Cost: $7.00
* Why: A raw slab of protein. No time for buns or garnish. Eat it with a fork in one hand while toggling front panel switches with the
other.
* The "Core Dump" Side: Cheddar Cheese
* Item: Cheddar Cheese (ID: SsnJvn9QPfn45MbTCsd2)
* Cost: $1.00
* Why: Dense, yellow blocks of energy. Reminiscent of the punch cards you're trying to debug.
* The "Infinite Loop" Beverage: Cola
* Item: Cola (ID: IcQMnaZNHGIfkOXmWlv5)
* Cost: $7.00
* Why: The closest 1960s substitute for a Red Bull. The sugar and caffeine are essential for maintaining the "High" logical state.
* The "Machine Cycle" Booster: Coffee Beans
* Item: Coffee Beans (ID: EL2qcJF0vp6ziOhijtxd)
* Cost: $4.00
* Why: Don't wait for the brew. Chew them raw for immediate instruction execution.
* The "Success Code" Dessert: Mint Chocolate Cookies
* Item: Mint Chocolate Cookies (ID: HWtLxQjjImVyfQGYJxwj)
* Cost: $3.00
* Why: A refreshing finish to clear the palate of cigarette smoke (metaphorically speaking) and frustration.
---
Total Check: $22.00
Efficiency Rating: 100%. Return to console immediately.
Final Code Review
Gemini CLI was used to validate the project:
> what does this project do?
✦ This project is an experimental Model Context Protocol (MCP) server implemented in COBOL (GnuCOBOL). It is named firestore-https-cobol and
serves as a proof-of-concept for building modern MCP-compliant tools using legacy languages.
Here are the key details:
* Core Function: It exposes a single MCP tool called greet, which takes a name as a parameter and returns a greeting (e.g., "Hello, World!").
* Technology Stack:
* Language: COBOL (main logic) and C (helper bindings).
* Library: It uses mcpc (a C-based MCP implementation) to handle the protocol details.
* Communication: It runs an HTTP server (default port 8080) to accept JSON-RPC 2.0 requests.
* Architecture: The COBOL code (server.cob) calls C functions (c_helpers.c, cob_helpers.c) to bridge with the mcpc library.
* Purpose: To demonstrate how to integrate COBOL applications into modern AI-driven ecosystems via the Model Context Protocol.
and code review the project:
✦ Code Review Report: firestore-https-cobol
The codebase investigation has been completed. Here is a summary of the findings, focusing on architecture, implementation correctness, and
safety.
1. Architecture & Interoperability
The project employs a robust multi-layered Foreign Function Interface (FFI) architecture to bridge the gap between 1960s-era COBOL and modern
C libraries.
* Bridge Mechanism: The system correctly uses a bidirectional FFI strategy:
* COBOL -> C: server.cob makes direct CALLs to mcpc library functions and custom C helpers. Argument passing handles string termination
and memory by pre-filling COBOL buffers with nulls (LOW-VALUES).
* C -> COBOL: cob_helpers.c implements the necessary wrapper functions (e.g., greet_cb_wrapper) that act as callbacks for the mcpc
library. Crucially, these wrappers correctly format arguments into a void *args[] array and use cob_call to re-enter the GnuCOBOL
runtime, ensuring context is preserved.
* Memory Strategy: The COBOL code uses fixed-size buffers (typically 4096 bytes) for string manipulation. This is standard for the language
but requires strict boundary checks (see "Safety" below).
2. MCP Implementation
The server correctly implements the Model Context Protocol lifecycle:
* Initialization: It follows the standard mcpc_server_new_tcp -> mcpc_server_start flow.
* Tooling: Tools like greet are correctly defined with properties and callbacks.
* Buffer Management: The GEMINI.md note regarding ensure_buf_suffi appears to be handled internally by higher-level helper functions like
mcpc_toolcall_result_add_text_printf8 (or the local helper_add_text_result), meaning the COBOL application code doesn't need to manually
resize buffers, reducing complexity and error risk.
3. Code Quality & Safety
* Security (Critical): The Firestore client (firestore_client.c) uses curl via shell commands.
* Finding: The function impl_get_product_by_id includes explicit validation to prevent shell injection attacks on the id parameter. This
is a critical security feature that must be preserved.
* Buffer Safety: COBOL calls to retrieve strings (e.g., mcpc_tool_get_tpropval_u8str) correctly pass a buffer capacity argument (BY VALUE
BUFFER-CAP), preventing overflows into the fixed-size COBOL variables.
* 64-bit Compatibility: The mjson library issue (segmentation faults with size_t) is a known constraint. Developers modifying C helpers must
continue to cast size_t lengths to int when using mjson format specifiers.
4. Testing
* The project uses test_server_http.py for integration testing, treating the COBOL binary as a black box and verifying behavior via standard
JSON-RPC 2.0 HTTP requests. This is an appropriate strategy for this type of system.
Recommendations
1. Maintain Validation: Ensure the shell injection checks in firestore_client.c are never weakened or removed.
2. Explicit Casting: Continue strictly enforcing (int) casts for all size_t values passed to mjson functions to maintain 64-bit stability.
3. Buffer Monitoring: While current buffer sizes (4KB) are likely sufficient for simple greetings, monitor this if tool outputs grow (e.g.,
large JSON responses from Firestore), as COBOL's static allocation is inflexible.
Summary
MCP development with COBOL using Gemini CLI was validated with an incremental step by step approach.
A minimal streaming HTTP transport MCP Server was built from source code and validated with Gemini CLI running as a MCP client in the same local environment.
Then — the MCP server was wrapped in a container and submitted to Google Cloud Build for deployment to Google Cloud Run. The remote MCP server was validated with a standard browser, and Gemini CLI.
Finally- remote MCP operations were performed from the local Gemini CLI installation to the COBOL MCP server hosted in Google Cloud Run.
This approach can be extended to more complex deployments and Cloud based options.
Top comments (0)