AI clients (MCP)

Connect edisyl to your AI client and ask a Knowledge Pack directly from chat or the tools you already use. Ask a pack a question, Teach it a fact, and Task its agents with longer work.

In addition to the CLI, edisyl also lets you reach your packs through MCP (Model Context Protocol). Everyone in your organization can ask questions and bring a pack’s knowledge into their existing workflow, in the AI tools they already use. Every client needs one URL:

https://api.<your-domain>/mcp

Use your edisyl install’s API host in place of api.<your-domain>. Your edisyl admin has it.

Add it, sign in with your edisyl account in the browser window that opens, and you’re connected. No API keys.

Set up your client

Claude

claude.ai and the Claude apps. Team and Enterprise owners can add it once for everyone. Set up Claude →

ChatGPT

Turn on Developer mode, add the URL as an app. Set up ChatGPT →

Microsoft 365 Copilot

A Copilot Studio agent, or an org-wide connector from the admin center. Set up Copilot →

Gemini Enterprise

Added once by a team admin under Connected apps. Set up Gemini Enterprise →

Working with a pack from chat

What to say, what comes back, and what a refusal means. Read the guide →

Using a coding agent?

Claude Code, Cursor, Codex, Antigravity, and other coding tools should drive the edisyl CLI instead of MCP. They get every verb plus source setup and steering, and edisyl onboard teaches the agent the whole flow.

What you can do once connected

Your client’s assistant gets these tools and calls them as you talk:

Say	Tool	Result
“Which packs can I use?”	list_packs	The packs in your org you can read.
“How many active wallets last week?”	ask	One cited answer over the pack’s knowledge and data.
“Show me the source for that.”	read_document	The document the answer cited.
“Why did it answer that way?”	explain_answer	The plan, sources, and definitions behind an answer.
“Remember that active means a swap in 30 days.”	teach	That fact, added to the pack for good.
“Research the new grant program and add it.”	task	The pack’s agents on longer work.
“Undo that teach.”	forget	The document and everything learned from it, removed.
“Check on that task.”	action_get, action_list, action_events, action_cancel	Status and progress of any ask, teach, or task, and a way to stop one.
“Ask this every Monday.”	schedule_*	Standing work the pack runs on its own.
“What does ‘active’ mean here?”	lexicon_*	The pack’s terms: list, read, preview, create, edit, merge, confirm, and remove.
“Is this pack ready to rely on?”	pack_health, pack_health_metric, pack_run_cases	Readiness against target, one metric over time, and a test run case by case.
“Where does calibration stand?”	calibration_*	Calibration status and changes, the decisions it asks of a person, grades, and test questions to retire or re-key.
“Brainstorm objectives for this pack.”	brainstorm, target_questions_*	The questions the pack’s calibration tests first.

Permissions

Which tools work depends on your access to the pack, not on how you connected:

• Any org member can list packs and ask the ones they can read.
• teach needs contribute access on the pack. task and brainstorm need steer access.
• Changing the Lexicon needs manage_pack. Calibration tools need manage_sources or steer access, depending on the call; each reference section says which.
• Anyone who can read a pack can list its schedules. Creating, firing, pausing, resuming, or deleting one needs steer access.

A forbidden response, or not found on a pack you can otherwise read, means you don’t have that permission. Ask the pack’s owner for access.

One organization per connection

A connection is bound to one edisyl organization. To use another org, disconnect and connect again.

Not on MCP

These are done in the CLI:

• Connecting and importing data sources, editing PACK.md, and publishing a pack.
• Creating packs and changing them, and managing members.
• Resuming calibration. calibration_pause is on MCP, but edisyl calibrate resume is not and needs an organization admin, so a team working only through chat should know who can resume.
• Approving a large compile, sweeping the test questions, and starting a test run now.
• Editing or archiving an objective.
• Finding Lexicon conflicts.
• Making figures searchable and downloading a figure.
• Reading every row an answer rests on with ask-rows.

SQL, raw knowledge retrieval, and the semantic catalog are not exposed at all: ask covers them and returns its own evidence.

This is not the header Connect menu

The Connect menu at the top of this site adds a different server, one that lets an AI tool search these docs. It does not connect to a pack. Use the URL on this page for that.