MCP Server (ordo mcp)

ordo mcp runs Ordo as a Model Context Protocol server over stdio, so a coding agent (Claude Code, Cursor, Windsurf, …) gets Ordo's tools natively — it can read, write, validate, test, and ship decision rules for you without leaving your editor.

Register

# Claude Code
claude mcp add ordo -- ordo mcp

For other clients, add a stdio MCP server whose command is ordo mcp, run from inside a decision project (the folder created by ordo init).

Tools

The server exposes nine tools. Read/edit/check tools operate on the local project files and the embedded engine — offline and instant; only publish reaches the platform.

Tool	What it does
list_files	List the project's files
read_file	Read a file
grep	Search files for a substring
write_file	Create/overwrite a file
delete_file	Delete a ruleset/tests/contracts file
validate	Compile a ruleset, structured errors
run_tests	Run a ruleset's test cases
trace	Execute an input and return the step-by-step path
publish	Deploy a ruleset to an environment

Safety

The server is local-first and git-backed, so file edits are reversible and allowed by default. High-risk actions are gated by flags:

ordo mcp --allow-publish     # permit the publish tool
ordo mcp --allow-delete      # permit deleting ruleset files

Without --allow-publish, the publish tool returns a blocked result rather than deploying — so an agent can propose a release but a human stays in control.

Typical flow

1. You tell your agent: "add a rule: approve if amount ≤ 10000, else reject."
2. The agent uses list_files / read_file to understand the project, then write_file to add rulesets/loan-approval.json.
3. It calls validate and run_tests, fixing anything that fails.
4. It calls trace to confirm a sample input takes the expected path.
5. With --allow-publish, it can publish — otherwise it hands off to you.

Because validate/test/trace are offline and sub-second, the agent's edit → check loop is tight, and its results match what the platform would produce (concepts are materialized identically).