Jaz AI · Agent stack for the ledger

The complete agent surface for Jaz accounting. 358 tools, 7 skills, 13 IFRS recipes, 13 calculators, 12 close playbooks. Works with any agent: Claude, GPT, Gemini, Copilot, Cursor. Token-lean discovery, first-try tool selection, structured errors an agent can recover from.

Also fully compatible with Juan Accounting (same API surface).

Contents

• Install · 30 seconds

• What you get

• Three layers

• Quick start

• Token economics + speed

• For AI agents

• For accountants

• Jaz Kit · run your practice

• Reference

• Troubleshooting

• Privacy & security

• Support

Related MCP server: MCP Multi-Agent Server

Install · 30 seconds

Stdio MCP config (Claude Desktop, Cursor, Windsurf, OpenAI Codex CLI / Agents SDK, any host that runs MCP servers as local processes):

{
  "mcpServers": {
    "jaz": {
      "command": "npx",
      "args": ["-y", "jaz-clio@5.36.12", "mcp"],
      "env": { "JAZ_API_KEY": "jk-your-api-key" }
    }
  }
}

VS Code MCP config (.vscode/mcp.json, workspace-scoped, read by VS Code and the GitHub Copilot Chat extension):

{
  "servers": {
    "jaz": {
      "command": "npx",
      "args": ["-y", "jaz-clio@5.36.12", "mcp"],
      "env": { "JAZ_API_KEY": "jk-your-api-key" }
    }
  }
}

Pin jaz-clio@5.36.12 for stability, or jaz-clio@latest for auto-updates. Multi-org: comma-separated keys, e.g. "JAZ_API_KEY": "jk-aaa,jk-bbb". Personal access tokens (pat-...) also work for multi-org.

Remote connector · no install

Bring Jaz into Claude (claude.ai, Desktop, mobile, Cowork) and ChatGPT with no install and no API key in any config. Just sign in.

1. In Claude: Settings → Connectors → Add custom connector (ChatGPT: Add a connector).

2. Enter the URL https://mcp.jaz.ai/mcp.

3. Sign in with your Jaz account (email one-time code or passkey) and Allow.

It uses OAuth 2.1 + PKCE: the agent receives a scoped, time-limited token tied to your account, never your password. One sign-in reaches every organization you belong to; name the org in your request (e.g. "in Acme Pte Ltd, list unpaid invoices"), and access to each is checked on every call. The same operations as the local server, packaged as namespace tools — one tools/list entry per accounting area, each routing to the operations inside it — with honest read-only / write / destructive hints. Ask "what can you do?" and the connector answers from its own live capability map. Bookkeeping only: it records entries and reads data. It moves no money.

Microsoft 365 Copilot · Copilot Studio

Bring Jaz into Microsoft 365 Copilot through a Copilot Studio agent. Cloud to cloud: no install, no API key in any config.

1. In Copilot Studio, open your agent and go to Tools → Add a tool → New tool → Model Context Protocol.

2. Enter the server name (Jaz), a short description, and the URL https://mcp.jaz.ai/mcp.

3. Authentication: OAuth 2.0 → Dynamic discovery → Create → Next.

4. On the Add tool dialog, select Create a new connection, sign in with your Jaz account, and Allow.

5. Select Add to agent, then publish the agent to Microsoft 365 Copilot or Teams as usual.

Same tool surface and per-call organization checks as the remote connector; the sign-in here uses OAuth 2.0 dynamic client registration. If your tenant restricts custom connectors, a Power Platform admin needs to allow this one.

Prefer a local install on the Microsoft stack? Copilot Studio is cloud-only and cannot run local MCP servers. Use VS Code with GitHub Copilot Chat instead: the VS Code MCP config runs Jaz locally with an API key, and npx jaz-clio init --platform copilot installs the skills to .github/copilot-instructions.md.

OpenAI Responses API

The Responses API only accepts HTTP MCP (no stdio). Point it at the hosted Jaz connector:

{
  "type": "mcp",
  "server_label": "jaz",
  "server_url": "https://mcp.jaz.ai/mcp",
  "headers": { "Authorization": "Bearer <your-oauth-token>" },
  "require_approval": "never"
}

The endpoint is OAuth-gated (no static API key). Obtain a token through the remote connector sign-in and pass it as a bearer token. To use a plain API key instead, take the OpenAI Codex CLI / Agents SDK stdio path above.

Just want skills (no MCP, any agent on the Agent Skills standard):

npx jaz-clio init                  # auto-detects your agent
npx jaz-clio init --platform cursor
npx jaz-clio init --no-rules       # skills only, skip the agent-rules file
npx jaz-clio init --check          # report drift between installed agent-rules version and current

Skills install to .agents/skills/ (Agent Skills standard, used by Cursor, Copilot, Codex, Antigravity, Windsurf, Goose, Roo Code, Junie, Amp, and more) or .claude/skills/ (Claude Code).

init also writes a one-page jaz-agent-rules.md block to the path your platform reads on workspace open, so any agent (Claude / GPT / Gemini / Copilot / Cursor) starts every session with the meta-tool discovery flow, the 6 API gotchas, and the recipe-engine carve-outs.

The block is wrapped in version-stamped markers (<!-- BEGIN jaz-agent-rules vX.Y.Z --> / <!-- END jaz-agent-rules -->), so re-running init updates only the Jaz block; your own rules above and below stay untouched. Run clio init --check to report drift between your installed version and the current package version (exit 1 = drift, exit 0 = current).

How many tools is it?

One catalog, three packagings. Every install reaches the same 358 operations — they are presented differently because hosts have different context budgets.

A directory listing that says "358 tools" and a client that shows 3 or 43 are describing the same server. Nothing is missing.

Ask the agent "what can you do?" on any of them — it answers from the connector's own live capability map (describe_capabilities), never from the length of its tool list.

What you get

• 358 tools covering every Jaz endpoint. Each tool description disambiguates against similar tools, lists enum values inline, and notes which operations are safe to retry. The model picks the right tool on the first call.

• 7 skills with the production-grade rules and playbooks any agent needs:

• 3 meta-tools (search_tools, describe_tools, execute_tool) for deferred discovery so the full catalog never has to load into context.

• Help center mirror at help-center-mirror/ synced weekly from Intercom.

• Structured-search DSL for natural-feeling queries (status:unpaid amount:>500 contact:Acme).

Three layers

The stack is one binary plus markdown skills, exposed through three layers that compose. The default agent setup is Skills + MCP. The CLI is the same binary as the MCP server with a different transport flag.

Skills layer on top of either. Most installs (Claude Code plugin, Claude Desktop MCPB, Cursor + MCP, Gemini extension) load Skills + MCP together. The MCP server runs locally (stdio, via the CLI binary) or hosted (the remote connector at mcp.jaz.ai, no install). The same operations either way — packaged differently. See How many tools is it?.

Quick start

Once installed, skills load automatically when an agent works with Jaz. Describe what you need:

Close the books for Acme for January. Bank-recon DBS Current first.
Then file GST for Q1.

Or call the CLI directly:

clio invoices list --json
clio calc loan --principal 100000 --rate 6 --term 60 --json
clio jobs month-end --period 2025-01 --json

Or via MCP from any agent:

search_tools("anomalous bills")
  → download_export at rank 1
execute_tool("download_export", { exportType: "analysis-anomalous-bills" })
  → { fileName, fileUrl }

Token economics + speed

Built so any model sees the right tool fast and calls it once.

For AI agents

• Skills load automatically from .claude/skills/, .agents/skills/, the Claude Code marketplace, the Gemini CLI extension, or the Claude Desktop MCPB.

• Discovery is one-shot. 68 canonical-query lock-in tests guarantee the right tool at rank 1 for the queries that matter.

• Disambiguation is explicit. Tools that look similar carry "USE THIS, not X" preambles. No more guessing.

• Errors are structured. Server validation failures return field-level details so the agent can self-correct.

• Multi-org is native. Comma-separated keys (jk-aaa,jk-bbb) or PATs unlock cross-org tools (list_organizations, per-call org_id).

• CONTEXT.md captures runtime rules-of-engagement (bootstrap with clio context --json, search before create, mutate as draft first, never echo API keys).

For accountants

Run period work conversationally. Describe it to any agent:

Close March for Acme. Bank-recon DBS Current first. File GST for Q1.

13 calculators (clio calc loan / depreciation / lease / ecl / fx-reval / provision / fixed-deposit / asset-disposal / prepaid-expense / deferred-revenue / accrued-expense / leave-accrual / dividend), 12 job blueprints (clio jobs month-end / quarter-end / year-end / bank-recon / document-collection / gst-vat / payment-run / credit-control / supplier-recon / audit-prep / fa-review / statutory-filing), all with --json for structured blueprint output.

In Claude Code, ten commands walk the jobs that have real steps to follow:

Every one of them also works by just asking — the commands are a shortcut, not a requirement.

Jaz Kit · run your practice

A close is not one conversation. Month-end runs eighteen steps over one to three days, and an accountant serving eight clients runs it eight times with eight different sets of bank accounts, materiality thresholds, and recurring accruals. Jaz Kit gives each organization a folder that remembers all of it.

/jk-setup                  set up the kit, connect an organization
/jk-open acme             load its context, verify the connection
/jk-close 2026-06         run the close — resumable across sessions
/jk-review                approve the drafts waiting on you
/jk-status                every organization, what's due, what's pending
/jk-exit                  journal the session, sweep scratch

/jaz-* runs a workflow. /jk-* runs your practice. Also /jk-keys, /jk-policy, /jk-teach, /jk-save, /jk-help.

~/Documents/Jaz Kit/
  orgs/<slug>/
    ORG.md                  bank accounts, materiality, recurring accruals, FY end
    policies/ rules/        how this organization works
    closes/2026-06/
      CLOSE.md              exactly where the close stopped
      workpapers/           the permanent audit file

Work is created as drafts and every record carries a link into Jaz, so you review in the UI and finalize when you're ready — the agent never posts live behind you. Interrupted closes resume where they stopped, and reconciliation steps verify against the ledger before retrying, so a crash never doubles a journal.

Each company's key lives in its own folder's .env. A jk- key is scoped to one company, so the folder you open decides which books you touch — no profiles, no labels, nothing to switch. Keys are gitignored, so a kit shared through a private git repo carries the context and policies but never the keys; each machine pastes its own. (On a default Mac ~/Documents syncs to iCloud, so keys sync too — bounded and revocable; set JAZ_KIT_HOME elsewhere to keep them off the cloud.)

One organization per session, enforced. Every call names its organization explicitly rather than relying on whichever one happens to be active. If something in your shell would silently override that choice — an exported JAZ_API_KEY, or several comma-separated keys — the command stops instead of posting to the wrong company's books.

Multi-organization work needs the CLI (npm i -g jaz-clio); a single organization works through MCP tools alone. Windows is supported by design but not yet verified — tell us if you hit something.

Start with /jk-setup, or just say "set up Jaz Kit for my company" — the skill triggers the same flows in Codex CLI, Cursor, and Copilot, which have no slash commands.

Slash commands share one global namespace across installed plugins, so /jk-* and /jaz-* could collide with another plugin using the same names. If that happens, ask for the flow in words instead ("open acme", "close the books for June") — it triggers on intent, not on the command name.

Reference

• CONTEXT.md · runtime rules-of-engagement for agents using the stack

• CHANGELOG.md · release notes

• Skills source · all 7 skills (jaz-api / jaz-cli / jaz-conversion / jaz-jobs / jaz-recipes / jaz-pseudo-sql / jaz-kit)

• OpenAPI spec · full HTTP surface, synced weekly

• README-cli.md · npm-package README, full CLI command catalog

• help.jaz.ai · Jaz product help center

• CLI surface: 66 command groups across the clio binary

jaz-api

jaz-recipes (13 IFRS recipes + 13 calculators)

jaz-jobs (12 close playbooks + Singapore Form C-S)

jaz-conversion

Troubleshooting

command not found: clio

Node.js is not installed, or npm's global bin directory is not on PATH.

node --version    # Need v18+. If missing, install LTS from https://nodejs.org
npm install -g jaz-clio
npm config get prefix    # e.g. /usr/local or ~/.nvm/versions/node/v22.x.x
export PATH="$(npm config get prefix)/bin:$PATH"   # Add to ~/.bashrc or ~/.zshrc

Auth error / 401 Unauthorized

Missing or invalid API key. Keys expire if regenerated in the Jaz app.

clio auth whoami
clio auth add <your-api-key>       # Get from Settings > API in Jaz
clio auth list                     # Confirm the right org is active

If you use env vars, set JAZ_API_KEY in the current shell or your MCP config's env block.

MCP not connecting

The config path is wrong, the command is wrong, or the server crashes on startup.

npx jaz-clio mcp                   # Smoke-test the server (Ctrl+C to stop)
claude mcp list                    # Confirm Claude Code sees "jaz"
claude mcp add jaz -- npx jaz-clio mcp

For Cursor / VS Code / Windsurf, validate the JSON and pin the API key:

{
  "command": "npx",
  "args": ["-y", "jaz-clio@5.36.12", "mcp"],
  "env": { "JAZ_API_KEY": "jk-your-api-key" }
}

Pin JAZ_API_KEY in MCP config rather than relying on the active CLI profile. MCP servers cache credentials at startup so clio auth switch won't take effect until restart. For multi-org, use comma-separated keys.

Skills not loading

Files are in the wrong path, or your agent doesn't auto-discover from there.

ls .claude/skills/                 # Claude Code
ls .agents/skills/                 # Agent Skills standard (Cursor, Copilot, etc.)
npx jaz-clio init                  # Re-install with auto-detection

Each skill folder must contain SKILL.md plus its reference files. Manual copies often miss subdirectories.

EACCES permission denied on npm install

npm's global directory needs elevated permissions on macOS/Linux without nvm.

# Option A: nvm (recommended)
curl -o- https://raw.githubusercontent.com/nvm-sh/nvm/v0.40.3/install.sh | bash
nvm install --lts && npm install -g jaz-clio

# Option B: fix npm permissions
mkdir -p ~/.npm-global && npm config set prefix ~/.npm-global
export PATH="$HOME/.npm-global/bin:$PATH"   # Add to ~/.bashrc or ~/.zshrc
npm install -g jaz-clio

Stale data after org switch

You ran clio auth switch but the MCP server still uses the previous org (it cached auth at startup).

# In Claude Code: remove and re-add
claude mcp remove jaz && claude mcp add jaz -- npx jaz-clio mcp

# In Cursor / VS Code: restart the editor or reload MCP servers

Or use multi-org mode and skip restarts: comma-separated keys (jk-aaa,jk-bbb) and switch orgs in conversation.

Privacy & security

Runs entirely on your machine. API calls go directly from your machine to the Jaz API over HTTPS. No telemetry. The API key lives locally in ~/.config/jaz-clio/credentials.json.

The hosted remote connector (mcp.jaz.ai) is different by design: you connect over OAuth 2.1 + PKCE (no API key stored anywhere, never your password), requests run server-side over HTTPS, and the connector reaches the organizations your Jaz account belongs to; access is checked on every call.

Full policy: jaz.ai/legal. Vulnerability disclosure: SECURITY.md.

Support

• Help center: help.jaz.ai

• Issues: github.com/teamtinvio/jaz-ai/issues

• Email: build.with@jaz.ai

License

MIT · Copyright (c) 2026 Jaz · Clio is a registered trademark.