OCI MCP Server

Created and maintained by Sarthak Pansare · open-source under Apache-2.0

Drive Oracle Cloud Infrastructure from any MCP-aware LLM client.

A Model Context Protocol server that exposes the OCI control plane — compute, networking, storage, databases, Kubernetes (OKE), load balancers, KMS vaults, functions, monitoring, and cross-service search — to any LLM tool that speaks MCP: Claude Code, Claude Desktop, Cursor, Codex CLI, Continue, VS Code, Zed, Antigravity, and others.

Token-efficient by design: one-line tool descriptions, compact summaries, pagination defaults, null-stripped responses.

What you can do

> Deploy a public web server called demo using my SSH key.
> Show me every resource in compartment XYZ grouped by type.
> Create a flexible Load Balancer fronting subnets A and B.
> List my OKE clusters and their kubernetes versions.
> Summarize CpuUtilization on instance ocid1.instance... for the last hour.
> Tear down the VCN we just created and the instance attached to it.

Tool surface (50+ tools across 12 modules)

Install

1. Prerequisites

• Python 3.10+

• uv (recommended) or pip

• An OCI account (cloud.oracle.com — Always-Free tier works)

2. Configure OCI credentials

In the OCI Console: avatar → My profile → API keys → Add API key → Generate API key pair.

Save the files:

~/.oci/oci_api_key.pem      # private key
~/.oci/config               # config (paste the snippet from the console)

Example ~/.oci/config:

[DEFAULT]
user=ocid1.user.oc1..xxxx
fingerprint=aa:bb:cc:...
tenancy=ocid1.tenancy.oc1..xxxx
region=us-ashburn-1
key_file=~/.oci/oci_api_key.pem

3. Clone + install

git clone https://github.com/sarthak-pansare/oci-mcp-server
cd oci-mcp-server
uv sync

Connect to your LLM client

The server is a standard MCP server that speaks stdio by default. Most clients only need a small JSON snippet.

Claude Code (CLI)

claude mcp add oci -s user -- uv --directory /absolute/path/to/oci-mcp-server run oci-mcp-server

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "oci": {
      "command": "uv",
      "args": ["--directory", "/absolute/path/to/oci-mcp-server", "run", "oci-mcp-server"]
    }
  }
}

Cursor

Settings → MCP → + Add new MCP server:

{
  "mcpServers": {
    "oci": {
      "command": "uv",
      "args": ["--directory", "/absolute/path/to/oci-mcp-server", "run", "oci-mcp-server"]
    }
  }
}

Codex CLI

Edit ~/.codex/config.toml:

[mcp_servers.oci]
command = "uv"
args = ["--directory", "/absolute/path/to/oci-mcp-server", "run", "oci-mcp-server"]

Continue (VS Code / JetBrains)

In your .continue/config.yaml:

mcpServers:
  - name: oci
    command: uv
    args: ["--directory", "/absolute/path/to/oci-mcp-server", "run", "oci-mcp-server"]

VS Code (native MCP)

In .vscode/mcp.json:

{
  "servers": {
    "oci": {
      "type": "stdio",
      "command": "uv",
      "args": ["--directory", "/absolute/path/to/oci-mcp-server", "run", "oci-mcp-server"]
    }
  }
}

Zed

In ~/.config/zed/settings.json:

{
  "context_servers": {
    "oci": {
      "command": {
        "path": "uv",
        "args": ["--directory", "/absolute/path/to/oci-mcp-server", "run", "oci-mcp-server"]
      }
    }
  }
}

Antigravity / generic MCP clients

Any client that supports stdio MCP can launch:

command: uv
args:    ["--directory", "/abs/path/to/oci-mcp-server", "run", "oci-mcp-server"]

Remote mode (HTTP/SSE)

For shared deployments or web-based clients:

uv run oci-mcp-server --transport sse --host 0.0.0.0 --port 3000
# or
uv run oci-mcp-server --transport streamable-http --host 0.0.0.0 --port 3000

Put it behind an authenticating reverse proxy. See SECURITY.md.

Docker

docker build -t oci-mcp-server .
docker run --rm -p 3000:3000 \
  -v ~/.oci:/root/.oci:ro \
  -e OCI_MCP_TRANSPORT=sse \
  oci-mcp-server

Configuration

Reduce token surface further

If you only need a subset of OCI services, register fewer tools — every tool's description is sent to the LLM on every turn:

uv run oci-mcp-server --enable iam,compute,networking,composite

Token-efficiency strategy

Designed to minimize the cost of using OCI tools with an LLM:

1. One-line docstrings. Every tool's description fits on a single line. Across all 50+ tools the schema fits in roughly 2 000 tokens.

2. Compact summaries by default. list_* calls return {id, name, state, ...} — not the full SDK object. Use verbose=True or get_* for full detail.

3. Pagination + limits. Every list call defaults to limit=50 (max 200) and accepts a page token.

4. Null-stripping. None / empty fields are removed from every response.

5. Module gating. --enable lets users register only the services they need.

Smoke test

After installing and restarting your client:

1. Run mcp__oci__whoami         → confirms credentials work
2. Run mcp__oci__list_regions   → confirms network path
3. Run mcp__oci__find_resources query: "query all resources"

For a full end-to-end deployment:

Deploy a web server called demo with this SSH key: ssh-ed25519 AAAAC3...

Roadmap

• v0.2: Logging, DNS, Events, Notifications, Streaming, Vault crypto, full Tagging UX

• v0.3: Instance-principal / resource-principal auth, work-request polling helpers

• v0.4: Terraform-style diffs, cost estimation tool, multi-region fan-out

Contributing

See CONTRIBUTING.md. PRs welcome.

Author

Sarthak Pansare — @sarthak-pansare

If this project saves you time, a ⭐ on the GitHub repo is appreciated.

License

Apache-2.0 — see LICENSE. Copyright © 2026 Sarthak Pansare.

Acknowledgements

• Anthropic Model Context Protocol

• Oracle Cloud Infrastructure Python SDK