Okama MCP

okama-mcp

MCP (Model Context Protocol) server that exposes the okama investment portfolio toolkit and okama Planner to AI assistants — Claude Desktop, Claude Code, Cursor, Codex, and any other MCP-compatible client.

With okama-mcp installed, you can ask an AI things like:

*"Backtest a portfolio of 30% gold and 70% real estate over the last 15 years."*

*"Run a Monte Carlo retirement forecast on that portfolio, withdrawing $1,000/month

indexed to inflation, over 25 years."*

*"What's the tangency portfolio of SPY, BND, and GLD with a 3% risk-free rate?"*

…and the AI uses the MCP tools to call okama directly — no Python code needed.

Built on FastMCP. Single codebase, two transports: stdio (for local clients) and streamable-http (for self-hosting). okama-mcp is free and open source. Run it locally or on your own server. The HTTP transport is intended for self-hosting; mcp.okama.io serves the project landing page, not a public MCP endpoint. Client registry tools and Excel exports require a compatible Planner companion and a configured local stdio installation.

Install

Development status (26 October 2026): main prepares MCP 3.1.0 with Planner 0.5.0; neither version has been published yet. Use the source setup below. PyPI still provides MCP 3.0.0 with Planner 0.4.1 and its earlier contracts.

Requires Python ≥ 3.11 (same floor as okama itself); okama ≥ 4.0.0 and okama-planner[reports]>=0.5.0,<0.6.0 are required.

Client data modes (source checkout): local MCP defaults to anonymous reads. client_data_mode_set({"mode":"personal_data"}) enables full client CRUD; {"mode":"anonymous"} disables client writes again. The setting persists per database across connections. uv sync installs a pinned Planner 0.5.0 source revision with the required access API. PyPI still provides Planner 0.4.1, so the prepared release remains blocked until the compatible companion is published; see installation and release gates. Market and finplan_* tools remain available independently. See the 3.0 migration notes for the earlier release behavior.

The easiest way — no clone, no venv — is uv or pipx:

uvx okama-mcp stdio          # run straight from PyPI

or

pipx install okama-mcp

Plain pip works too:

pip install okama-mcp

[!WARNING]

With pip, prefer a dedicated virtual environment: on most modern Linux distros the

system Python is marked externally managed (PEP 668), so pip install outside a venv

fails, and a shared environment risks dependency conflicts. In your MCP client config,

point command at the absolute path of the okama-mcp script inside the venv — GUI

clients don't see your shell PATH. uvx and pipx avoid all of this by isolating

the install automatically.

To work on the code, install from source instead:

git clone https://github.com/mbk-dev/okama-mcp

cd okama-mcp uv python install 3.14 uv sync --python 3.14 --all-groups

Until Planner 0.5.0 is published, tool.uv.sources pins an immutable development companion. Release gates use uv sync --all-groups --no-sources and require published dependencies; no source substitution is used for publication. Remove the development pin after published-pair acceptance passes.

Run

# stdio — for Claude Desktop, Claude Code, Cursor (local IPC)

okama-mcp stdio

streamable HTTP — for self-hosting on your own server

okama-mcp http --host 127.0.0.1 --port 8765

When running from a source checkout, prefix each command with uv run. Each worktree has its own .venv; test and dev groups are installed by default. uv.lock is generated locally and is not committed, so fresh checkouts resolve compatible versions independently. Orca runs the bootstrap in orca.yaml before starting the agent; its source MCP configuration uses uv run --no-sync after setup.

Connect a client

Claude Desktop

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

{

"mcpServers": { "okama": { "command": "uvx", "args": ["okama-mcp", "stdio"] } } }

Restart Claude Desktop; the server appears in the tools menu.

Claude Code

To make the server available in every project (works from any directory):

claude mcp add --scope user okama -- uvx okama-mcp stdio

Developers running from a source checkout can use claude mcp add okama -- uv run okama-mcp stdio from the project root instead.

Or commit a .mcp.json at the project root so the whole team picks it up:

{

"mcpServers": { "okama": { "command": "uvx", "args": ["okama-mcp", "stdio"] } } }

Cursor

Add the server to .cursor/mcp.json in your project (or ~/.cursor/mcp.json to make it global):

{

"mcpServers": { "okama": { "command": "uvx", "args": ["okama-mcp", "stdio"] } } }

Codex (CLI & Desktop)

Add the server with one command:

codex mcp add okama -- uvx okama-mcp stdio

Or declare it in ~/.codex/config.toml (or a project-scoped .codex/config.toml in trusted projects):

[mcp_servers.okama]

command = "uvx" args = ["okama-mcp", "stdio"]

The Codex CLI, desktop app, and IDE extension share this configuration — set it up once and it works in all three.

Self-hosting (streamable HTTP)

Run okama-mcp on your own server and share it across your MCP clients:

okama-mcp http --host 127.0.0.1 --port 8765 --path /mcp

(From source: uv run okama-mcp http ...)

Then point your MCP client at http:// :8765/mcp. For a production setup put nginx + TLS in front; ready-made examples live in deploy/:

• deploy/systemd/okama-mcp.service — systemd unit (hardened, runs as a dedicated user)
• deploy/nginx/self-hosted.conf — nginx vhost: TLS, SSE-friendly proxying of /mcp

The server is open by design — free to run, no registration. If your instance must not be public, restrict access at the nginx level (allow-list, VPN, or HTTP basic auth).

Tool catalog

A multi-stage financial plan (contribute $1,000/month for 20 years into a 70/30 SPY/AGG portfolio, then withdraw $6,000/month indexed to inflation for 25 years), a Monte Carlo retirement forecast (30% gold / 70% real estate, withdrawing $1,000/month indexed to inflation over 25 years) and the efficient frontier of SPY/BND/GLD:

Financial tools are stateless — pass the full portfolio specification with every call. Optional local client tools use stable registry codes and a persistent, database-scoped access mode. Client creation is not idempotent: inspect the registry before retrying an ambiguous creation result. The server caches expensive okama objects (Portfolio, EfficientFrontier) by content hash, so repeated calls on the same spec are fast.

Nested portfolios. Wherever a list of assets is accepted — the assets field of PortfolioSpec/FrontierSpec, or the portfolios argument on the comparison tools — an entry may be a ticker string or a nested portfolio object (the same spec shape). This lets you treat a whole portfolio as a single component: e.g. compare a 60/40 portfolio against gold, or put a sub-portfolio on the efficient frontier.

Search & metadata

Single asset & comparisons

Portfolio backtest

Monte Carlo DCF

| `find_the_largest_withdrawa