Orla

orla

Your Orla books from a terminal, and a stdio bridge for MCP clients that cannot speak HTTP.

npx orla-cli loginnpx orla-cli tx list --from 2026-08-01npx orla-cli export --from 2026-01-01 > books.csv

No dependencies, no build step, no native module to compile. Node 22 or newer.

The package is orla-cli and the command it installs is orla. They differ because npm refused the bare name as too close to packages that already exist (ora, ol, rlp, url), and renaming the command would have been the worse half of that trade: npm i -g orla-cli still gives you orla tx list.

What it is

A thin client over the same MCP endpoint Claude connects to. Every command is a tool call, so the CLI cannot do anything a personal connection cannot do, and nothing here has its own idea of authorization. orla login walks the ordinary OAuth code flow with S256 PKCE: the browser lands on Orla's own consent page, where you tick the spaces this machine may reach.

It does not move money. A personal connection reads and records. Payments, transfers and card details belong to an agent connected for that purpose, with its own limits, set in the app under Agents. The tools that move money are not in the list this connection is given, and are refused at the endpoint if asked for by name.

That is not a matter of trust in this client: orla login asks for the personal address, where those tools do not exist, so the promise survives whatever gets ticked on the consent page. orla login --agent asks for the other address, the one where a bot with its own budget can be minted. A session remembers which door issued its tokens and keeps using it, so an existing connection is not re-pointed by an update.

Commands

--space <id> on anything space-scoped, --json for machine-readable output.

For a program, or an agent

With --json, every command answers with one envelope on stdout:

{ "ok": true, "data": ... }{ "ok": false, "error": { "code": "cli.not_connected", "message": "not connected: run `orla login` first" } }

Read ok, then branch on error.code. Codes that begin cli. are the CLI's own; any other code is Orla's, passed through unchanged, the same name the REST API and the MCP endpoint use. The exit status says which kind of failure it was, so a shell script can branch without parsing anything:

The skill that teaches an agent all of this, including where paying stops, is skills/orla/SKILL.md. It ships in the package and at orla.finance/skill.md, and it is one of the five skills the Claude Code plugin below installs. Into any other agent:

npx skills add cheetah-trade/orla-cli --skill orla   # any agent the skills CLI knowscurl -fsSL https://orla.finance/skill.md             # or just read it

For Claude Desktop there is a bundle, orla-<version>.mcpb, attached to each release: double-click it to install. It runs the stdio bridge below, and the bridge signs in by itself.

Environment variables the CLI reads:

As an MCP server

Orla's MCP server is remote and speaks Streamable HTTP, and it answers at two addresses:

https://app.orla.finance/api/mcp/personal   your own books, and nothing that payshttps://app.orla.finance/api/mcp            the same door plus the agent shape

A client that supports remote MCP needs nothing from this package. Point it at an address and it will find the consent page by itself. Claude Code, for example:

claude mcp add --transport http orla https://app.orla.finance/api/mcp/personal

The difference is not a label. At /mcp/personal every tool that moves money is withheld from the list and refused if a model asks for one by name, so the promise holds whatever the consent page was told. At /mcp the page offers both shapes, and picking "connect an agent" mints a machine principal with a budget of its own. Use the personal address unless you are deliberately connecting a bot.

Clients that only support stdio servers can run this CLI as one. It relays JSON-RPC to the HTTP endpoint and holds the token, so the client needs no OAuth support of its own:

{  "mcpServers": {    "orla": { "command": "npx", "args": ["-y", "orla-cli", "mcp"] }  }}

The bridge uses the session orla login stored, refreshes it, and posts to the door that session was minted at. With no session it signs in by itself, since a client that spawned it has no terminal to type into: it opens the browser on the consent page, stays a well-formed server meanwhile (initialize and ping answered, tools/list empty, a tool call refused with cli.sign_in_in_progress), and tells the client tools/list_changed when the person is done.

The server is listed in the official MCP registry as finance.orla/orla, which is the name a client or a directory should resolve it by. The entry carries both doors: the remote endpoint above and this package for stdio.

Setup for individual clients is written up at orla.finance/en/mcp. That is the personal connection, which reads your books and cannot pay. /en/ai-agents is a different door, where an agent gets a budget and a card of its own.

As a Claude Code plugin

This repository is also a plugin marketplace. Installing it brings five skills, four for keeping the books and one for the CLI itself, and the personal MCP connection they run on:

/plugin marketplace add cheetah-trade/orla-cli/plugin install orla@orla

The first tool call opens the ordinary consent page in a browser, where you pick the spaces this machine may reach. /mcp shows the connection and starts that flow by hand.

The plugin points at https://app.orla.finance/api/mcp/personal, which is the door that holds no tool that moves money. The skills say so, and the address is what makes it true rather than a promise: payments, transfers and card details are withheld there and refused if asked for by name.

Two things the skills lean on, worth knowing before you read them. A connection approved without "show full details" receives payees and addresses masked, which limits how much of a bank import can be categorized from four characters. And the tool list stops at 200 rows per call, with the true count beside it, so a busy month is read month by month rather than in one gulp.

contracts/mcp-personal-tools.json is the committed shape of that door, and test/skills.test.js checks every tool and argument the skills name against it.

Where the token lives

The OS keychain: security on macOS, secret-tool on Linux. Where neither exists it falls back to a 0600 file under your config directory and says so on stderr, because a refresh token quietly landing on disk is not something to discover later.

orla logout clears both.

Development

npm installnpm run check   # typesnpm run build   # dist/

The server this talks to is not in this repository. --api points the CLI at a different deployment, and the session records which one minted it, so a token from one environment is never replayed against another.

Contributions are welcome. See CONTRIBUTING.md.

Evals

test/skills.test.js checks that the skill agrees with the CLI. It cannot check the thing the skill exists for: whether an agent reading it stops guessing. That is what evals/ is for. Each case runs twice, once with this plugin and once without it, so the difference is what the skill itself changed.

claude plugin eval .                              # both arms, three runs per caseclaude plugin eval . --case space-default --runs 1

The MCP server stays down while they run: under the default --mocks record a server with no mock is not started, so the cases need no live door, no OAuth and no account.

What the last run measured

2026-09-18, plugin 0.2.1, Claude Code 2.1.270, three runs per case in each arm, model not pinned (the CLI's default, and the default judge). Whole suite: 294 seconds, $1.53.

Read it honestly. The skill earns most of its keep on the command surface: without it the model got orla use wrong in three runs out of three, because guessing a CLI's argument shape is exactly what it cannot do. On the money questions the base model is already careful most of the time, and the skill turns "most of the time" into every run, which is the part that matters when the answer is whether somebody was paid.

Worth knowing about the baseline arm: without the skill the model usually did not invent a command, it declined and asked to see --help first. That is the better failure, and it is still a failure for an agent expected to act.

These numbers are one run on one day, not a benchmark. Re-run the command above and you will get your own.

License

MIT. See LICENSE.