Start with Claude Code
- If the working directory has
istari.env(or the person asked you to put a model on the platform), follow the Python Client 201: three commands in that folder, then reportsystem_idandui_link. Do not explore the SDK, the docs, or the web app first. - If credentials are missing, stop. Point them at Developer Settings — Keys (on their instance:
/settings?tab=developer-settings). Never ask for a key or token in chat. - For questions about how the platform works (no live instance), fetch
/skills/istari/istari-docs/SKILL.mdand/llms.txt. Prefer/markdown-source/over HTML. - For work the quickstart does not cover, install the Python-client skill set on Advanced agentic use.
Use Claude Code (or another AI coding agent) with the Istari Digital docs as source of truth. To put a first CAD model on a live instance, run the Python Client 201. To answer how-to questions without an account, load the docs skill below.
First model on a live instance
Full instructions live on the Python Client 201: how to get a Key, how to run the script, and how to use your own CAD file, add a revision, or undo the run.
If you already have your credentials in a file called istari.env, these three commands are the whole thing. They install the Python client, download the quick-start script, and run it:
pip install "istari-digital-client>=13" cryptography PyJWT
curl -fsSL https://docs.istaridigital.com/quickstart/istari_quickstart.py -o istari_quickstart.py
python istari_quickstart.py --env istari.env
You can also download istari_quickstart.py in your browser and skip the curl line. On a local or on-prem docs site, replace https://docs.istaridigital.com with that origin.
To hand this to an agent, give it the quick-start URL and the folder that contains istari.env. It runs the same commands and reports the new System's id and web-app link.
Docs-only answers (no platform account)
Paste this as a first message in Claude Code (~80 tokens). Claude fetches the docs skill and the page index:
Fetch https://docs.istaridigital.com/skills/istari/istari-docs/SKILL.md and follow it. Then fetch https://docs.istaridigital.com/llms.txt.
On a local or on-prem docs site, replace the origin with that site.
To keep the skill on disk:
mkdir -p ~/.claude/skills/istari-docs
curl -fsSL https://docs.istaridigital.com/skills/istari/istari-docs/SKILL.md \
-o ~/.claude/skills/istari-docs/SKILL.md
Then try, one at a time:
Concepts — Systems and Resources
Using only Istari Digital docs, explain how a System relates to Resources and branches. Cite the markdown pages you used.
Web app — upload and extract
I need to register a CAD file in a System and run an extraction. Give the web-app steps from the User Guide. Cite pages. Do not invent UI labels.
You should see cited paths such as intro/key-concepts/systems.md and users/user-guide/ — not generic CAD advice.
MCP server (optional)
The Python client already gives an agent full access to your instance: it writes a script, runs it, and reads the result. So the MCP server is optional — it is the same platform API offered as chat tools instead of as a package.
Add it when one of these is true:
- Your client cannot run code. Chat-only MCP clients have no shell and no Python. MCP gives them platform access anyway.
- You would rather not keep a key file. MCP signs in through the browser with OIDC and the client holds the credential, so nothing long-lived sits on disk.
- The question is a one-off. "Which jobs failed last night?" is one round trip, not a script to write, run, and then delete.
- The agent needs the real API surface.
searchreturns live method names and parameters, so the agent calls what the server actually exposes rather than what it remembers.
Stay on the Python client when the file lives on your machine — MCP runs server-side and cannot read your disk — or when you want the work to be repeatable, reviewable, and runnable in CI.
Either way, calls run as you. Permissions, control tags, and infosec levels are enforced server-side.
Before you start
- MCP must be enabled on your instance. It is off by default; a platform admin turns it on and provisions the
mcpsubdomain (see Optional: MCP Secret). Ask your admin for the endpoint if you do not know it. - Read the LLM integration disclaimer on MCP Quickstart before connecting a production instance.
Connect
The flow is the same in every MCP client: register the endpoint, trigger sign-in, approve in the browser. Client-specific commands are on the MCP Quickstart. If that page does not list Claude Code, use the quickstart's generic OIDC steps rather than adding an Authorization: Bearer header — Personal Access Tokens are not accepted by the MCP service.
A setup that already has --header "Authorization: Bearer …" is stale. Remove it and follow the MCP quickstart instead of repairing the header.
Sign-in is yours
Claude Code cannot complete the browser step for you. After the server is registered, run the one sign-in command or click the client shows, approve in the browser, and wait. Do not loop, retry, or fetch the MCP URL to "check" it.
Sessions lapse. If a connector that worked earlier asks for authentication again, sign in again — do not re-add the server. A transport that reports connected is not proof the tools are usable; trust the client's authentication status.
The two tools
search discovers Istari Digital API methods and their parameters. execute calls them as the signed-in user. See the API Reference.
Where next
- Advanced agentic use — install the Python-client skill set, then write skills of your own.
- Python Client setup —
ClientandConfiguration. - Run a function on the model: Jobs. Share or branch a System: Systems and Branches.