Every execution tool now accepts an optional server parameter. Pass a server name and
the call routes to that instance without touching the default connection:
{ "tool": "iris_execute", "server": "prod", "code": "Write $ZV" }
iad maintains its own server registry at ~/.config/iris-agentic-dev/servers.json.
Passwords go to the OS keychain — never to the config file — using the same service name
as VS Code Server Manager, so both tools share credentials automatically.
Five new tools manage the pool: iris_servers, iris_add_server, iris_remove_server,
iris_test_server, and iris_import_servers.
Three new tools give Claude a persistent ObjectScript terminal over WebSocket:
iris_ws_open, iris_ws_exec, iris_ws_close. Variables and process state persist
between calls. Before this, every iris_execute call was a fresh context — a sequence
like "set X, do some work, read X back" had to be one call or use a global to carry
state. Now each step can be separate. Requires IRIS 2023.2+ (Atelier V7 API).
Twenty-two new tools across global management, namespace/database admin, observability,
and cross-instance comparison. Most of this section is ported from Pierre Abdelsayed's
Server Manager MCP work — the global confirmation pattern, namespace/database admin,
observability tools, HL7 schema tools, Mermaid diagrams, and resolve_storage all
originate from his design. The data safety gates (PHI policy, system globals blocklist,
environment template) are also his, ported to Rust in an earlier release.
global_preview + global_kill. Preview returns a confirmationiris_namespace_list, iris_namespace_create,iris_database_list, iris_database_stats.journal_search, query_audit_log, stream_inspect, my_access,capability_matrix.hl7_schema_list, hl7_schema_inspect. Return HL7_NOT_AVAILABLEEnsLib.HL7.Schema (requires HealthShare or IRIS for Health).mermaid_class, mermaid_production, resolve_storage.compare_document (unified diff of a single documentcompare_namespace (full namespace diff).All 64 tools now carry MCP ToolAnnotations. 57 are tagged read_only_hint = true
(queries, introspection, list tools). The 7 destructive tools — global_kill,
iris_admin, iris_credential_manage, iris_lookup_manage, iris_namespace_create,
iris_remove_server, skill_forget — are tagged destructive_hint = true. Claude Code
and other MCP clients that read these hints can show a confirmation step before
destructive calls and run read-only tools without prompting.
iris_production, iris_production_item, and all skill tools returned errors or empty
results on connections without IRIS_CONTAINER set: VS Code extension, remote servers,
anything not running in a named local container. These tools have working Atelier REST
paths; the docker-exec fallback was routing around them. All affected tools now use
Atelier REST.
None. All existing tool calls work identically. The server parameter defaults to
omitted, so hot-reload and active-connection behavior are unchanged.
iris_select_container still works.
MCP server and CLI that connects AI coding agents to a live InterSystems IRIS instance.
Compile ObjectScript, run SQL, search namespaces, run unit tests, inspect productions —
from any MCP-compatible agent or from the command line.
Requires IRIS 2023.1 or later. Works with native IRIS on Windows or Linux, and with Docker.
90+ tools exposed over the Model Context Protocol. Connect once; your agent calls tools
by name. Works with Claude Code, GitHub Copilot, OpenCode, or any MCP client.
iris-agentic-dev mcp
To expose only the IRIS tools without the skill/KB/learning-agent surface, pass
--no-skills (or set IRIS_NO_SKILLS=true). Useful when skills are installed
separately and you want to avoid duplicate tool names in the MCP client:
iris-agentic-dev mcp --no-skills
→ Quick start: VS Code + Copilot |
Quick start: Claude Code / OpenCode |
Full connection docs
One tool call, one subprocess, no server. Good for agents that run shell commands,
CI scripts, and skill repos that need IRIS access without a persistent MCP process.
iris-agentic-dev tool iris_query '{"sql":"SELECT TOP 5 ID FROM Sample.Person"}'
iris-agentic-dev tool iris_compile '{"target":"MyApp.Foo.cls"}'
iris-agentic-dev tool iris_execute '{"code":"Write ##class(%SYS.ProcessQuery).GetInfo()"}'
→ Tool reference |
Annotated session: tracing an ADT^A03 through a production
No IRIS connection, no server. Installs ObjectScript instruction files into your agent’s
skill directory. The binary and skills are independent — installing one doesn’t install
the other.
iris-agentic-dev skill install
→ Skill docs |
Full skill reference
brew tap intersystems-community/tap
brew install iris-agentic-dev
brew upgrade iris-agentic-dev updates the binary only — installed skills are
not touched by Homebrew. Run iris-agentic-dev skill install after upgrading to
pick up new skills. Files installed before the managed_by marker was introduced
(or installed by other means) will be skipped as unrecognized; pass --force to
overwrite them and stamp them for automatic updates going forward.
curl -fsSL https://github.com/intersystems-community/iris-agentic-dev/releases/latest/download/iris-agentic-dev-macos-arm64 \
-o /usr/local/bin/iris-agentic-dev && chmod +x /usr/local/bin/iris-agentic-dev
xattr -d com.apple.quarantine /usr/local/bin/iris-agentic-dev 2>/dev/null
curl -fsSL https://github.com/intersystems-community/iris-agentic-dev/releases/latest/download/iris-agentic-dev-linux-x86_64 \
-o /usr/local/bin/iris-agentic-dev && chmod +x /usr/local/bin/iris-agentic-dev
Windows: Download iris-agentic-dev-windows-x86_64.exe from the
releases page
and place it on your PATH.
The VS Code extension handles binary discovery and connection config automatically.
Prerequisites: VS Code, GitHub Copilot,
InterSystems ObjectScript extension
iris-agentic-dev (IRIS) now appears in Copilot Chat → Agent mode → tools. It reads
your existing objectscript.conn or intersystems.servers configuration — no additional
setup needed.

To verify the connection, ask Copilot: “Call check_config and show me the result.”

If the
InterSystems Server Manager
extension is installed, iris-agentic-dev reads your server list and retrieves credentials
from the OS keychain automatically — no additional config needed. Set IRIS_SERVER_NAME
if you have multiple servers configured.
Windows users: iris-agentic-dev works with native IRIS on Windows — Docker is not
required. If you hit a 404 on/api/atelier, see
Windows IIS setup below.
After installing the binary, configure your agent:
Claude Code — add to ~/.claude.json:
{
"mcpServers": {
"iris-agentic-dev": {
"command": "iris-agentic-dev",
"args": ["mcp"],
"env": {
"IRIS_HOST": "localhost",
"IRIS_WEB_PORT": "52773",
"IRIS_USERNAME": "_SYSTEM",
"IRIS_PASSWORD": "SYS",
"IRIS_NAMESPACE": "USER"
}
}
}
}
OpenCode — add to ~/.config/opencode/config.json:
{
"mcp": {
"iris-agentic-dev": {
"type": "local",
"command": ["/usr/local/bin/iris-agentic-dev", "mcp"],
"enabled": true,
"environment": {
"IRIS_HOST": "localhost",
"IRIS_WEB_PORT": "52773",
"IRIS_USERNAME": "_SYSTEM",
"IRIS_PASSWORD": "SYS",
"IRIS_NAMESPACE": "USER"
}
}
}
}
Note: OpenCode uses "type": "local" and "environment" (not "type": "stdio" and "env").
WSL2: The Windows OpenCode GUI cannot spawn Linux ELF binaries. Use the Windows .exe
or invoke the Linux binary via wsl.exe:
"command": ["wsl.exe", "-e", "/usr/local/bin/iris-agentic-dev", "mcp"]
Cursor is configured like other stdio MCP clients (~/.cursor/mcp.json), not like the
VS Code Copilot extension auto-registration path. Server Manager keychain and the Marketplace
extension behave differently (or not at all) in many Cursor setups.
See https://github.com/intersystems-community/iris-agentic-dev/blob/master/docs/cursor-quickstart.md for install, mcp.json, fleet
server= routing, and an explicit Cursor vs VS Code + Copilot comparison.
Add a .iris-agentic-dev.toml file to your project root:
host = "localhost"
web_port = 80 # IIS default for IRIS 2024.1+; use 52773 for pre-2024.1
namespace = "USER"
username = "_SYSTEM"
password = "SYS"
| IRIS version | Web server | Default port |
|---|---|---|
| 2024.1+ on Windows | IIS | 80 |
| 2024.1+ on Linux | Apache | 80 |
| Pre-2024.1 (any OS) | PWS | 52773 |
/api web application requiredThis is the most common failure on Windows. IIS needs an explicit /api web application
mapped to the IRIS Web Gateway module. Without it, /api/atelier returns 404 — even when
the Management Portal loads correctly.
To fix:
api, physical path:C:\InterSystems\IRIS\CSP\bin (adjust to your install path)CSPms.dll, no verb restrictionCSP.ini contains an [APP_PATH:/api] sectionSee the iris-windows-iis-setup skill
for full step-by-step instructions with verification commands.
localhost vs 127.0.0.1: On some older Web Gateway builds, using localhost causes
a brief connection error before each request. If you see connection delays, change to
host = "127.0.0.1".
Run iris-agentic-dev init in your project directory — it detects any running IRIS
containers and writes .iris-agentic-dev.toml automatically:
iris-agentic-dev init
Or configure manually:
container = "myapp-iris"
namespace = "MYAPP"
Enterprise IRIS images (intersystems/iris, intersystems/irishealth) ship without a
built-in web server. Run the ISC Web Gateway container alongside IRIS:
services:
iris:
image: containers.intersystems.com/intersystems/iris:2026.1
ports: ["4972:1972"]
webgateway:
image: containers.intersystems.com/intersystems/webgateway:2026.1
ports: ["52773:80"]
entrypoint: ["/bin/sh", "/init.sh"]
volumes: ["./webgateway-init.sh:/init.sh:ro"]
See the iris-vscode-objectscript skill
for a working webgateway-init.sh.
If the
InterSystems Server Manager
extension is installed, iris-agentic-dev reads your server list from VS Code’s
settings.json and resolves credentials from the OS keychain — no .iris-agentic-dev.toml
needed.
Single server configured: auto-connects, no extra setup.
Multiple servers configured: set IRIS_SERVER_NAME to the map key from
intersystems.servers:
export IRIS_SERVER_NAME=dev-local
Use check_config to see which servers were detected and whether credentials resolved.
Add [policy.<server-name>] blocks to .iris-agentic-dev.toml to restrict which tool
categories are permitted on a given server:
[policy.prod]
allow = ["query", "search", "docs"]
Blocked calls return error_code: "POLICY_GATE" with the list of allowed categories.
Available categories: compile, execute, query, search, docs, source_control,
debug, admin, skill, kb.
iris-agentic-dev resolves the IRIS connection in this order — first match wins:
--host, --web-port, --scheme).iris-agentic-dev.toml in the workspace rootIRIS_HOST, etc.)settings.json (objectscript.conn / intersystems.servers)intersystems.servers + OS keychain credential)| Variable | Default | Description |
|---|---|---|
IRIS_HOST |
localhost |
IRIS web gateway hostname |
IRIS_WEB_PORT |
52773 |
Web gateway port |
IRIS_SCHEME |
http |
http or https |
IRIS_WEB_PREFIX |
(empty) | URL path prefix for non-root gateway installs |
IRIS_USERNAME |
_SYSTEM |
IRIS username |
IRIS_PASSWORD |
SYS |
IRIS password |
IRIS_SERVICE_USERNAME |
(empty) | Least-privilege account for execute/query/write tools |
IRIS_SERVICE_PASSWORD |
(empty) | Password for IRIS_SERVICE_USERNAME |
IRIS_NAMESPACE |
USER |
Default namespace |
IRIS_CONTAINER |
(empty) | Docker container name (required for ✦ tools) |
IRIS_SERVER_NAME |
(empty) | Server Manager server name (multiple servers) |
OBJECTSCRIPT_WORKSPACE |
$PWD |
Workspace root for .iris-agentic-dev.toml lookup |
IRIS_ENABLED_TOOLS |
(empty) | Comma-separated allowlist — expose only these tools |
iris_execute, iris_execute_method, iris_query (mode="write"), and iris_global
(set/kill) can run arbitrary ObjectScript/SQL. Under a %All account these can edit
class and routine code — even by indirection — bypassing the SCM lock.
Set IRIS_SERVICE_USERNAME / IRIS_SERVICE_PASSWORD to a least-privilege IRIS account
(no %Development resource, code database mounted read-only). Those four tools then
authenticate as that account, so code edits fail with <PROTECT> at the IRIS privilege
layer. Code-writing tools (iris_doc put, iris_source_control, iris_compile) keep
using the primary IRIS_USERNAME so audit stays attributed to the real user.
Skills are concise instruction files that teach your AI assistant ObjectScript-specific
patterns and common mistakes. They work with or without the MCP server and require no IRIS
connection.
Skills and the MCP server are independent — installing the binary installs no skills.
Tested with Claude Sonnet 4.6 on the ObjectScript repair suite (22 tasks):
| Benchmark suite | Baseline | With top skill | Lift |
|---|---|---|---|
| ObjectScript repair (22 tasks) | 73% | 100% | +27% |
The top skill is objectscript-review — a 205-word checklist that catches the 10 most
common ObjectScript mistakes before the AI writes any code.
Read the +27% as a rough signal — one run, one model, 22 tasks that may be in training
data. BENCHMARKING.md covers the caveats and lets you run it
yourself.
Install skills:
iris-agentic-dev skill install # full pack, Claude Code + OpenCode
iris-agentic-dev skill install objectscript-review # selective
iris-agentic-dev skill install --agent copilot # repo-scoped Copilot instructions
iris-agentic-dev skill list # check install status
VS Code Copilot: The extension installs the binary only. Run
iris-agentic-dev skill install --agent copilot from a git repo root to install skills
into .github/instructions/.
| Skill | What it does | Benchmark |
|---|---|---|
objectscript-review |
Hard-gate checklist: 10 most common AI mistakes in ObjectScript | 100% repair |
objectscript-guardrails |
All-in-one hard gate, works without MCP | 86% repair |
objectscript-sql-patterns |
IRIS SQL quirks: reserved words, SQLCODE, table naming, NULL handling | 100% SQL |
objectscript-unit-test |
Generates %UnitTest scaffolding from live class introspection |
86% repair |
objectscript-list-patterns |
%List, $LISTBUILD, $LISTNEXT, $LISTTOSTRING patterns |
91% repair |
objectscript-navigation |
Codebase discovery using MCP introspection tools | 82% repair |
objectscript-tdd |
Compile-test-fix loop for iterative development | |
objectscript-debugging |
Maps .INT offsets to .CLS source lines, reads error logs |
|
objectscript-repair |
Coordinated fixes across multiple dependent classes | |
iris-docs |
Fetches live IRIS class reference before implementing any API — eliminates hallucinated methods | |
iris-vector-ai |
IRIS vector search syntax (HNSW, VECTOR_COSINE, TO_VECTOR) |
domain |
iris-connectivity |
IRIS connection APIs from Python, Java, JDBC, ODBC | domain |
ensemble-production |
Interoperability production lifecycle, logs, queues | domain |
iris-devtester |
IRISContainer factory methods and test fixture patterns |
domain |
Note: some skills hurt if loaded globally.
objectscript-loop-patternsmeasured
−19% lift when loaded for all tasks. Domain skills should only load when working in
those areas. See BENCHMARKING.md.
See skills/ for the full list and how to contribute a skill.
Most tools work over the Atelier REST API and connect to any IRIS instance — no Docker
required unless noted. Tools marked ✦ require IRIS_CONTAINER. Tools marked 🔒 are
write-gated (suppressed on Live instances unless IRIS_ALLOW_PROD=1).
Tools that can reach patient data or IRIS internals are blocked before they run unless
you opt in. See Data safety gates for what is
blocked and how to permit it.
| Tool | What it does |
|---|---|
iris_compile |
Compile a class, routine, or wildcard. Returns errors with line numbers. |
iris_doc |
Read, write, delete, or check any IRIS document. |
iris_execute |
Run ObjectScript, return output. |
iris_execute_method |
Invoke a ClassMethod directly by class+method+args, no boilerplate. |
iris_query |
Execute SQL, return rows as JSON. mode=explain|count|write for plans, counts, and gated DML. |
iris_test |
Run %UnitTest tests, return structured pass/fail results. |
iris_global |
Read, write, kill, or list IRIS global nodes. Patient-data and system globals are gated. |
iris_coverage |
Measure ObjectScript line coverage via %Monitor.System.LineByLine. |
iris_source_control ✦ |
Check lock status, checkout, execute SCM actions. |
| Tool | What it does |
|---|---|
iris_symbols |
Search classes and methods via %Dictionary. |
iris_symbols_local |
Search .cls/.mac/.inc files on disk by glob — no IRIS connection required. |
docs_introspect |
Deep class inspection: methods, properties, XData, superclasses. |
iris_search |
Full-text search across the namespace. Supports regex and category filters. |
iris_info |
Namespace discovery: documents, jobs, CSP apps, metadata. |
iris_macro |
Macro inspection: list, signature, definition, expand. |
iris_table_info |
Inspect a SQL table: class-projected vs. DDL, backing storage globals. |
resolve_dynamic_dispatch |
Resolve $classmethod/##class({var}) polymorphic dispatch to candidate classes. |
extract_message_map_routing |
Extract a compiled Ensemble MessageMap routing table. |
find_subclass_implementations |
Find all concrete subclass implementations of a method. |
| Tool | What it does |
|---|---|
iris_debug |
Map INT offsets to source lines, fetch error logs, capture error state. |
iris_get_log |
Retrieve a full result by log_id when a tool returns truncated: true. |
check_config |
Show active connection state — host, container, config file, write tool status. |
| Tool | What it does |
|---|---|
iris_generate |
Build a context-rich prompt for generating ObjectScript. No API key required. |
iris_generate_class |
Generate and compile a class from a description (requires LLM API key). |
iris_generate_test |
Generate %UnitTest scaffolding for an existing class. |
| Tool | What it does |
|---|---|
iris_production ✦ |
Start, stop, update, check, or recover a production. |
iris_interop_query ✦ |
Query production logs, queue depths, or message archive. |
iris_production_item 🔒 |
Enable, disable, or get/set settings on a production config item. |
iris_production_diff |
Diff the running production config against the last source-controlled version. |
iris_message_body |
Read a message body by ID. Blocked unless the connection permits patient data. |
iris_business_rule_info |
List or inspect Ensemble business rules. |
iris_credential_list |
List Ensemble credentials (IDs/usernames only — passwords never returned). |
iris_credential_manage 🔒 |
Create, update, or delete an Ensemble credential. |
iris_lookup_manage |
Read, write, delete, or list Ensemble lookup table entries (write actions gated). |
iris_lookup_transfer |
Export or import an Ensemble lookup table as XML (import gated). |
| Tool | What it does |
|---|---|
iris_admin |
List namespaces, databases, users, roles, web apps; create/delete users; real-time observability. |
journal_search |
Search the IRIS journal for SetKill records by global pattern and/or time range. |
iris_namespace_list |
List all namespaces on the connected instance. |
iris_namespace_create 🔒 |
Create a new namespace. |
iris_database_list |
List databases and their paths. |
iris_database_stats |
Disk usage and block-level stats for a database. |
query_audit_log |
Query the IRIS SQL audit log for recent activity. |
my_access |
Show the current user’s roles and resource permissions. |
capability_matrix |
Show which tools are enabled/disabled and why (gates, policy, container availability). |
iris_containers ✦ |
List, select, or start IRIS Docker containers. |
| Tool | What it does |
|---|---|
agent_history |
Recent tool-call history for the current session. |
agent_stats |
Learning agent status: skill count, pattern count, KB size. |
telemetry_query |
Query the durable telemetry record beyond the in-memory session. |
skill |
Manage the learning agent skill registry: list, describe, search, forget, or propose. |
skill_community |
Browse or install community skills published to subscribed GitHub repos. |
kb |
Index markdown/text into the IRIS knowledge base, or recall content by keyword. |
→ Full tool reference with error codes
| Symptom | Likely cause | Fix |
|---|---|---|
404 on /api/atelier (Windows) |
IIS missing /api web application |
See Windows IIS setup |
check_config works but compile/search fail |
Atelier Recurse=0 |
Management Portal → Security → Web Apps → /api/atelier → enable Recurse |
| All tools fail, namespace listing works | API version mismatch | Verify IRIS supports Atelier v8 |
| 403 on write operations | Insufficient permissions | Use a user with %DB_USER or %All role |
| Connection delays on Windows | localhost DNS issue |
Use host = "127.0.0.1" in .iris-agentic-dev.toml |
SERVER_MANAGER_CREDENTIAL_ERROR |
Credential not in OS keychain | VS Code → Server Manager → right-click server → Reconnect |
SERVER_MANAGER_AMBIGUOUS |
Multiple SM servers, no server name | Set IRIS_SERVER_NAME=<server-key> |
For verbose HTTP logging:
iris-agentic-dev mcp --verbose 2>debug.log
iris-agentic-dev mcp # Start the MCP server
iris-agentic-dev tool # Call a tool directly (no server)
iris-agentic-dev compile MyApp.Foo.cls # Compile from the terminal
iris-agentic-dev skill install [names] # Install skills
iris-agentic-dev skill list # Check skill install status
iris-agentic-dev init # Generate .iris-agentic-dev.toml
iris-agentic-dev benchmark --skill # Run the skill benchmark harness
iris-agentic-dev --version # Print version
| Guide | Contents |
|---|---|
| https://github.com/intersystems-community/iris-agentic-dev/blob/master/docs/connecting.md | Native IRIS, Docker, Server Manager, policy gates, env vars, discovery order |
| https://github.com/intersystems-community/iris-agentic-dev/blob/master/docs/tools.md | Full tool catalog with descriptions and error codes |
| https://github.com/intersystems-community/iris-agentic-dev/blob/master/docs/skills.md | Skill inventory, benchmark results, CLI install reference |
| https://github.com/intersystems-community/iris-agentic-dev/blob/master/docs/troubleshooting.md | Symptom table, CLI commands, verbose logging |
| https://github.com/intersystems-community/iris-agentic-dev/blob/master/docs/ecosystem-integration.md | Patterns for downstream projects and skill repos |
Issues and pull requests are welcome. File bugs at the
Issues tab.
To contribute a skill: write a SKILL.md, run the benchmark, submit a PR with your
results. See BENCHMARKING.md.
Questions: thomas.dyar@intersystems.com