aihub-eap

InterSystems AI Hub — EAP Reference (Build 162)

Current build: 2026.2.0AI.162 (community image available; enterprise pending) LLMConfig property is GONE. The pattern is now Parameters + %Init() + ConfigStore. Build 161 has MCP server breaking changes — see section 13. Build 162 adds 13 new %AI classes vs 161 — see section 14.

---

1. Getting the Build

EAP Portal: https://evaluation.intersystems.com/Eval/early-access/AIHub

Current build: 2026.2.0AI.162.0 (as of 2026-04-26)

Available downloads:

iris-community-2026.2.0AI.158.0-docker.tar.gz          # x86_64 Docker (community, no key needed)
iris_arm64-community-2026.2.0AI.158.0-docker.tar.gz    # ARM64 Docker (community)
iris-container-x64.key                                   # License key (removes community limits)
iris-container-arm64.key                                  # ARM64 license key
iris.key                                                  # General license key
langchain_intersystems-0.0.1-py3-none-any.whl           # Python SDK
IRIS_Community-2026.2.0AI.158.0-macx64.tar.gz           # macOS Intel
IRIS_Community-2026.2.0AI.158.0-win_x64.exe             # Windows

Community builds (iris-community-*) run without a license key but have connection/data limits. Use a key to remove limits — request from the EAP portal or from your ISC contact.

---

2. Docker Setup

docker image load -i iris-2026.2.0AI.159-docker.tar.gz

docker run --name iris-ai-hub \
  -p 1972:1972 -p 52773:52773 \
  -d \
  --volume /path/to/key-dir:/external/keys \
  intersystems/iris:2026.2.0AI.159 \
  -k /external/keys/iris-container-x64.key

Fix expired default password

docker exec -it iris-ai-hub iris session iris -U %SYS
# In session:
#  write ##class(Security.Users).UnExpireUserPasswords("*")

iris-devtester integration

from iris_devtester import IRISContainer
container = IRISContainer("intersystems/iris:2026.2.0AI.159") \
    .with_name("iris-ai-hub") \
    .with_bind_ports(1972, 52773) \
    .with_license_key("/path/to/iris-container-x64.key") \
    .start()

---

3. %AI.Agent — Build 159 Patterns

Pattern A: Declarative subclass (recommended)

Class MyApp.AI.MyAgent Extends %AI.Agent
{
    Parameter PROVIDER = "openai";           // provider name
    Parameter MODEL = "gpt-4o";             // model id
    Parameter APIKEY;                        // empty = read from env OPENAI_API_KEY
    Parameter TOOLSETS = "%AI.Tools.SQL";   // comma-separated toolsets

    XData INSTRUCTIONS [ MimeType = text/markdown ]
    {
    You are a helpful IRIS database assistant.
    Use SQL tools to answer questions about the data.
    }

    Method %OnInit() As %Status
    {
        // Optional: additional setup after %Init() completes
        // Set ..Temperature = 0.7
        Return $$$OK
    }
}

Usage:

Set agent = ##class(MyApp.AI.MyAgent).%New()
$$$ThrowOnError(agent.%Init())           // REQUIRED — initializes provider, tools, prompt

Set session = agent.CreateSession()
Set response = agent.Chat(session, "How many patients are in the database?")
Write response.Content

DO NOT call Chat() without %Init() first — tools and provider won't be wired.

Pattern B: Programmatic (for dynamic provider selection)

Set settings = {}
Do settings.%Set("api_key", apiKey)
Set provider = ##class(%AI.Provider).Create("openai", settings)

Set agent = ##class(%AI.Agent).%New(provider)
Set agent.Model = "gpt-4o-mini"
Set agent.SystemPrompt = "You are a helpful assistant."

Set session = agent.CreateSession()
Set response = agent.Chat(session, "Hello!")
Write response.Content

Pattern C: ConfigStore integration (production pattern)

Class MyApp.AI.ProdAgent Extends %AI.Agent
{
    Parameter MODELCONFIGNAME = "opsreview";   // name in ConfigStore

    Method %OnInit() As %Status
    {
        Set sc = $$$OK
        Try {
            If ..Provider = "" && ..#MODELCONFIGNAME '= "" {
                Set sc = ..GetProviderForConfig(..#MODELCONFIGNAME, .provider, .model)
                Quit:$$$ISERR(sc)
                Set ..Provider = provider
                Set ..Model = model
            }
        } Catch ex {
            Set sc = ex.AsStatus()
        }
        Return sc
    }

    ClassMethod GetProviderForConfig(
        configName As %String,
        Output provider As %AI.Provider,
        Output model As %String) As %Status
    {
        Set sc = $$$OK
        Try {
            Set sc = ##class(%ConfigStore.Configuration).GetDetails(
                "AI.LLM."_configName, .details, 0, 1)
            Quit:$$$ISERR(sc)
            Set provider = ##class(%AI.Provider).Create(details."model_provider", details)
            Set model = details."model"
        } Catch ex {
            Set sc = ex.AsStatus()
        }
        Quit sc
    }
}

ConfigStore entry for "opsreview":

Set config = {
    "model_provider": "anthropic",
    "model": "claude-sonnet-4-5@20250929",
    "api_key": "secret://MySecrets.anthropic#apikey"
}
Do ##class(%ConfigStore.Configuration).Create("AI","LLM","","opsreview", config)

---

4. @{} Variable Substitution

Used inside PROVIDERCONFIG parameters and ToolSet XML — NOT in ObjectScript code directly.

Syntax	Source	Example
@{env.VAR}	OS environment variable	@{env.OPENAI_API_KEY}
@{config.Key}	^%AI.Config global	@{config.VertexSAPath}
@{wallet.Col.Key}	IRIS Secure Wallet	@{wallet.AISecrets.anthropic}

// In a declarative agent Parameter:
Parameter PROVIDERCONFIG = "{
    ""project_id"": ""my-gcp-project"",
    ""region"": ""us-east5"",
    ""service_account_path"": ""@{env.VERTEX_SA_PATH}""
}";

// In ToolSet XData (MCP remote server with token from wallet):
<MCP Name="MyServer">
    <Remote URL="https://mcp.example.com/mcp"
            AuthType="bearer"
            Token="@{wallet.MCPSecrets.token}"/>
</MCP>

@{config.AI.LLM.opsreview.APIKey} is NOT a valid pattern. Use GetDetails() + %AI.Provider.Create() for ConfigStore-backed API keys in code.

---

5. Session Management

// Create session (inherits agent's provider, model, prompt, tools)
Set session = agent.CreateSession()

// With config overrides
Set cfg = {"max_iterations": 10, "temperature": 0.7, "max_tokens": 1000}
Set session = agent.CreateSession(cfg)

// Blocking chat
Set response = agent.Chat(session, "Your question here")
Write response.Content

// Streaming
Set renderer = ##class(%AI.System.StreamRenderer).%New()
Set response = agent.StreamChat(session, "Your question", renderer, "OnChunk")
Do renderer.Flush()

// Multi-modal
Set content = [
    {"type": "text", "text": "What is in this image?"},
    {"type": "image_url", "image_url": {"url": "https://example.com/img.jpg"}}
]
Set response = agent.ChatWithContent(session, content)

// Session stats
Set stats = session.GetStats()
Write stats."total_interactions", " turns, ", stats."total_tool_calls", " tool calls"

// Context inspection
Set messages = session.GetContext()
Set iter = messages.%GetIterator()
While iter.%GetNext(.i, .msg) { Write msg.role, ": ", $EXTRACT(msg.content,1,80), ! }

// Session control
Do session.Reset()           // full reset
Do session.ResetContext()    // messages only
Do session.ResetStats()      // stats only

---

6. %AI.Provider — Supported Providers

// OpenAI
Set provider = ##class(%AI.Provider).Create("openai",
    {"api_key": key, "organization": orgId})

// Anthropic
Set provider = ##class(%AI.Provider).Create("anthropic", {"api_key": key})

// AWS Bedrock — bearer token (ISC SSO)
Set provider = ##class(%AI.Provider).Create("bedrock",
    {"region": "us-east-1", "bearer_token": token})
// MUST use cross-region inference IDs: "us.anthropic.claude-sonnet-4-6"
// NOT raw model IDs: "anthropic.claude-sonnet-4-6"

// AWS Bedrock — SigV4 (AWS keys)
Set provider = ##class(%AI.Provider).Create("bedrock",
    {"region": "us-east-1"})   // reads AWS_ACCESS_KEY_ID etc from env

// Google Vertex AI
Set provider = ##class(%AI.Provider).Create("vertex",
    {"project_id": pid, "region": "us-east5",
     "service_account_path": saPath})

// Google Gemini
Set provider = ##class(%AI.Provider).Create("gemini", {"api_key": key})

// xAI Grok
Set provider = ##class(%AI.Provider).Create("grok", {"api_key": key})

// NVIDIA NIM
Set provider = ##class(%AI.Provider).Create("nim", {"base_url": url})

---

7. ConfigStore API

// Store
Set config = {"model_provider":"anthropic","model":"claude-sonnet-4-5@20250929",
              "api_key":"secret://Secrets.ant#apikey"}
Do ##class(%ConfigStore.Configuration).Create("AI","LLM","","myconfig", config)

// Retrieve (raw)
Set cfg = ##class(%ConfigStore.Configuration).Get("AI","LLM","","myconfig")

// Retrieve with secret resolution (resolveSecrets=1)
Set sc = ##class(%ConfigStore.Configuration).GetDetails(
    "AI.LLM.myconfig", .details, 0, 1)
// details."model_provider", details."model", details."api_key" all resolved

// Delete
Do ##class(%ConfigStore.Configuration).Delete("AI.LLM.myconfig")

---

8. Wallet (Secrets)

// Create security resource first (%SYS namespace)
Set $NAMESPACE = "%SYS"
Do ##class(Security.Resources).Create("AI.Secrets")

// Create collection
Do ##class(%Wallet.Collection).Create("AISecrets",
    {"UseResource": "AI.Secrets", "EditResource": "AI.Secrets"})

// Store secret
Do ##class(%Wallet.KeyValue).Create("AISecrets.anthropic",
    {"Usage": "CUSTOM", "Secret": {"apikey": "sk-ant-...actual-key..."}})

// Reference in ConfigStore: "api_key": "secret://AISecrets.anthropic#apikey"

---

9. Python (langchain-intersystems)

python -m venv .venv && source .venv/bin/activate
pip install ./langchain_intersystems-0.0.1-py3-none-any.whl
pip install mcp langchain-openai langchain-ollama

from langchain_intersystems.chat_models import init_chat_model
from langchain_intersystems import init_mcp_client

llm = init_chat_model("AI.LLM.myconfig")    # uses ConfigStore entry
mcp = init_mcp_client("AI.MCP.my-server")

Common issues:

• Version mismatch: check pip show langchain-intersystems matches the whl filename
• Platform errors: add --force-reinstall if wheel metadata conflicts

---

10. Breaking Changes: Build 141 → 159

Feature	Build 141	Build 159
LLMConfig property	Existed	REMOVED
Agent instantiation	##class(%AI.Agent).%New(provider) only	Also: declarative Parameters subclass
Config Store	Partial WIP	Full %ConfigStore.Configuration support
%Init() requirement	Optional	REQUIRED before first Chat()
Streaming	Not available	StreamChat() + %AI.System.StreamRenderer
Session reset	Not available	Reset(), ResetContext(), ResetStats()
Checkpoints	Not available	AddCheckpoint(), RewindTo()
Session fork	Not available	Fork(), ForkAndSummarize()
Nested agents	Not available	CreateSubAgent(), DelegateTask tool
Skills	Not available	%AI.Agent.Skill with XData
RAG	Not available	%AI.KnowledgeBase, EnableSmartDiscovery()
Prompt caching	Not available	"cache": {"enabled": 1} in session config
@{wallet.*} substitution	Not available	@{wallet.Collection.Key}

DEAD CODE from build 141 (do not use):

// ❌ LLMConfig is gone
Set agent.LLMConfig = "AI.LLM.openai"

// ❌ Must call %Init() first
Set agent = ##class(MyAgent).%New()
Set response = agent.Chat(session, "hi")   // will fail — no provider wired

---

11. Quick Smoke Test

// Verify %AI.* classes are available
Write ##class(%AI.Provider).%IsA("%RegisteredObject"), !   // → 1

// Minimal working agent (env var for key)
Set env("OPENAI_API_KEY") = "sk-..."      // or set in OS env before starting IRIS
Set provider = ##class(%AI.Provider).Create("openai", {"api_key": $SYSTEM.Util.GetEnviron("OPENAI_API_KEY")})
Set agent = ##class(%AI.Agent).%New(provider)
Set agent.Model = "gpt-4o-mini"
Set session = agent.CreateSession()
Set r = agent.Chat(session, "Say hi")
Write r.Content

---

13. Build 161 MCP Server Breaking Changes (2026-04-26)

Proven working in ~/mindwalk-test2/ on EC2 with 18/18 tools discovered and %Invoke executing correctly.

%AI.MCP.Service.Utils is GONE in build 161

LoadToolSetsToManager was available in builds 147–159. It does not exist in 161.

// ❌ Build 161 — class does not exist:
do ##class(%AI.MCP.Service.Utils).LoadToolSetsToManager("My.ToolSet","/mcp/myapp")
// → <CLASS DOES NOT EXIST> *%AI.MCP.Service.Utils

This is fine — iris-mcp-server in build 161 discovers tools directly from the compiled class via wgproto without needing the cache pre-populated. Do not call it, do not add workarounds.

Compile order is critical in build 161

%AI.ToolSet.Specification.Compiler validates all referenced classes at compile time. If ToolSet compiles before its <Include Class="..."> targets are loaded, it discovers 0 tools.

// ❌ WRONG — compiles ToolSet too early (flag "ck" = load + compile):
do $system.OBJ.LoadDir("/src/Mindwalk","ck",,1)

// ✅ CORRECT — three-step pattern:
do $system.OBJ.LoadDir("/src/Mindwalk","k",,0)          // 1. load all, no compile
do $system.OBJ.CompilePackage("Mindwalk","ck")           // 2. compile everything
do $system.OBJ.Compile("Mindwalk.ToolSet","ck")          // 3. recompile ToolSet last

The batch compile (CompilePackage) may still emit <CLASS DOES NOT EXIST> *Mindwalk.GraphToolsPy during the ToolSet sub-compile — this is benign as long as the final explicit Compile("Mindwalk.ToolSet","ck") succeeds. Check that the last compile outputs Compilation finished successfully.

iris session IRIS -U USER "..." hangs in build 161

Inline single-command iris session calls hang indefinitely in build 161.

# ❌ Hangs — do not use for readiness checks:
iris session IRIS -U USER "write 1,! halt" >/dev/null 2>&1 && break

# ✅ TCP probe — works reliably:
bash -c 'cat < /dev/null > /dev/tcp/localhost/1972' 2>/dev/null && break

# ✅ Script file via irissession — works for class loading:
cat > /tmp/init.script << 'EOF'
zn "MINDWALK"
do $system.OBJ.LoadDir("/src/Mindwalk","k",,0)
halt
EOF
/usr/irissys/bin/irissession IRIS < /tmp/init.script 2>&1

iris-mcp-server binary is architecture-specific

The binary at services/iris-mcp-sidecar/iris-mcp-server in some repos is ARM64. It will silently fail on x86_64 EC2. Always use the binary inside the IRIS image:

# ✅ Sidecar pattern — gets the right binary for the right arch:
FROM ${IRIS_IMAGE}    # same image as the IRIS container
ENTRYPOINT ["/usr/irissys/bin/iris-mcp-server"]
CMD ["--config", "/etc/iris-mcp/config.toml", "run"]

AutheEnabled=96 — must patch if iris.script ran at image build time

If your iris.script baked AutheEnabled=64 into the image, patch it at container start:

zn "%SYS"
set appProps("AutheEnabled") = 96
do ##class(Security.Applications).Modify("/mcp/myapp", .appProps)

The iris-mcp-server GET /mcp/myapp/v1/services request arrives unauthenticated. AutheEnabled=64 (Password only) returns HTTP 500. 96 (Password + Unauthenticated) allows it through.

---

12. Verified-on-Build-159 Findings (2026-04-25)

@{} substitution — wallet prefix NOT registered on 159

Only env and config prefixes work. wallet causes an error at runtime.

// ✅ Works on 159:
Parameter APIKEY = "@{env.OPENAI_API_KEY}";

// ❌ Fails on 159 — wallet not registered:
Parameter APIKEY = "@{wallet.AISecrets.openai}";

Use env vars or store keys directly in PROVIDERCONFIG referencing @{env.*}.

%AI.LLM.Response — use .Content, never .%Get("content")

response.Content is a typed %String property. response.%Get() throws <METHOD DOES NOT EXIST>.

// ✅ Correct:
Write response.Content

// ❌ Throws <METHOD DOES NOT EXIST>:
Write response.%Get("content")

Available properties on the response object:

• .Content — %String — the assistant's text reply
• .ToolCalls — %DynamicArray — tool call requests from the model
• .Usage — %DynamicObject — token usage (prompt_tokens, completion_tokens)

LLMConfig property is gone on 159 (confirmed)

The LLMConfig property documented for build 141 does NOT exist on 159. Using it causes <PROPERTY DOES NOT EXIST>. Use declarative Parameters or programmatic %AI.Provider.Create().

irishealth vs iris image — use irishealth for MCP over HTTP

Image	WebServer	CSPServer binary	Use for
irishealth-community-*	1 (enabled)	Yes	MCP over HTTP, REST, web apps
iris-community-*	0 (disabled)	No	ObjectScript/CLI only, no HTTP tools

# ✅ For MCP over HTTP:
docker image load -i irishealth-community-2026.2.0AI.158.0-docker.tar.gz
docker run --name iris-ai-hub -p 1972:1972 -p 52773:52773 -d \
  irishealth/iris-community:2026.2.0AI.158.0

# ⚠️ Plain iris image has no web gateway — iris_execute HTTP path won't work

Empty Parameter APIKEY auto-reads from OS environment

When APIKEY parameter is empty (the default), %Init() automatically resolves @{env.OPENAI_API_KEY} (or the provider-appropriate env var).

Class MyAgent Extends %AI.Agent {
    Parameter PROVIDER = "openai";
    Parameter APIKEY;   // empty = reads OPENAI_API_KEY from OS env automatically
}

// Set in shell before starting IRIS:
// export OPENAI_API_KEY=sk-...

Set agent = ##class(MyAgent).%New()
$$$ThrowOnError(agent.%Init())   // picks up OPENAI_API_KEY from env

This is the recommended pattern for local dev. For production, use ConfigStore + Wallet.

---

14. Build 162 New Classes (community 2026-04-26; enterprise pending)

Build 162 community has 51 %AI classes vs 38 in enterprise 161. The 13 additions:

RAG stack (new in 162)

%AI.RAG.Embedding
%AI.RAG.Embedding.FastEmbed
%AI.RAG.Embedding.OpenAI
%AI.RAG.KnowledgeBase
%AI.RAG.VectorStore.IRIS

MCP client in ToolSet XData (consume external MCP servers, new in 162)

%AI.ToolSet.Specification.MCP
%AI.ToolSet.Specification.MCP.Remote
%AI.ToolSet.Specification.MCP.Stdio
%AI.ToolSet.Specification.Utils

Built-in tool providers (new in 162)

%AI.Tools.FileSystem
%AI.Tools.ShellTools
%AI.Tools.SQL

Agent composition (new in 162)

%AI.Agent.Skill
%AI.Agent.SubAgent
%AI.Tool
%AI.Tool.Resolver
%AI.Tool.Schema

ConfigStore/Wallet API (new in 162, NOT in enterprise 161)

%AI.Utils.ConfigStore
%AI.Utils.SettingStore
%AI.Utils.WalletStore

Important: The aihub-eap skill documents %AI.Utils.ConfigStore patterns — these only work on community 162+ or when enterprise catches up. On enterprise 161, calling any %AI.Utils.* method throws <CLASS DOES NOT EXIST>.

MCP client in ToolSet (162 only) — consume external MCP servers

Build 162 adds <MCP> elements to the ToolSet XData, letting IRIS act as an MCP client and expose remote MCP server tools as local tools:

<ToolSet Name="MyToolSet">
    <Include Class="MyApp.LocalTools"/>
    <MCP>
        <Remote Name="external" URL="https://example.com/mcp"/>
        <Stdio Name="local-tool" Command="/usr/bin/my-mcp-tool"/>
    </MCP>
</ToolSet>

docker pull for 162 community

The tags list API has a pagination bug — 162.0 doesn't appear but pulls fine directly:

docker pull docker.iscinternal.com/docker-intersystems/intersystems/irishealth-community:2026.2.0AI.162.0

---

13. Linux Docker Volume Permissions (from READY 2026 hackathon — Anthony Master)

Symptom: IRIS container exits immediately on Linux with:

terminate called after throwing an instance of 'std::runtime_error'
what(): Unable to find/open file iris-main.log in current directory /home/irisowner/dev

Root cause: ALL IRIS container editions (community, enterprise, irishealth, ai_hub) run as UID 51773 (irisowner). When you bind-mount a host directory owned by UID 1000 (typical Linux user), the container can read the volume but cannot write to it — and IRIS needs to write iris-main.log at startup.

Not affected: macOS (VirtioFS translates permissions transparently).

Fix options

Option 1 — POSIX ACLs (recommended, minimal footprint)

setfacl -R -m u:51773:rwX <repo-dir>
setfacl -R -d -m u:51773:rwX <repo-dir>

The -d flag makes new files/dirs inherit the rule automatically. Verify with: getfacl <repo-dir>

Option 2 — tmpfs (no persistence)

# docker-compose.yml
volumes:
  - type: tmpfs
    target: /home/irisowner/dev

Option 3 — chown on host (broad)

sudo chown -R 51773:51773 <repo-dir>

Option 4 — Docker named volume (avoid bind-mount entirely)

volumes:
  iris-data:
services:
  iris:
    volumes:
      - iris-data:/home/irisowner/dev

Team note: If you re-clone or a new team member sets up the repo on Linux, they must re-run the setfacl commands. Add this to your project README or Makefile setup target.