Run your company's work straight from ChatGPT and Claude
Samurai App ships an MCP server and a CLI. Paste one URL and your AI agent reads and writes conversations, knowledge, databases, and SIROE wiki with the same permissions as the screen.
MCP server: paste one URL#
Paste this URL into the connector settings of Claude, ChatGPT, or another client. A browser opens the first time; sign in to Samurai App and allow access.
The URL for the connectorhttps://app.samuraiax.com/mcp/<tenant-id>
- 21
One tool per resource
The operation is chosen with action. It does not mirror every endpoint one to one, so the tool list stays short.
- 2
Protocol generations
Both 2026-07-28 and the 2025 generation are supported, so any client works.
- 0
Secret keys on the client
Sign-in is OAuth with PKCE. The only thing the client holds is the URL.
Names in the client's tool list
samuraiapp_conversationsamuraiapp_approvalsamuraiapp_knowledgesamuraiapp_databasesamuraiapp_siroesamuraiapp_calendarsamuraiapp_meetingsamuraiapp_memorysamuraiapp_skillsamuraiapp_botsamuraiapp_appsamuraiapp_filesamuraiapp_triggersamuraiapp_workflowsamuraiapp_approversamuraiapp_tenantsamuraiapp_settingsamuraiapp_secretsamuraiapp_connectionsamuraiapp_auditsamuraiapp_activity
Endpoints not in the list are reachable through samuraiapp_request. Authorization takes the same path as every other tool, so it widens the surface but not the permissions.
CLI: from a terminal or an agent's shell#
The samuraiapp command installs in one line and needs no administrator rights. It signs in with the same URL as MCP.
Install and sign in
# Mac / Linux
curl -fsSL https://app.samuraiax.com/install.sh | sh
# Windows (PowerShell)
irm https://app.samuraiax.com/install.ps1 | iex
samuraiapp login https://app.samuraiax.com/mcp/<tenant-id>
# There is only one shape
samuraiapp <resource> <action> [--key value]
Works when an agent calls it
Run from anything other than a terminal, it skips the interactive screen and prints the command list. Scripts, CI, and AI agents' shells use it as is.
Deletes ask first
Deleting asks for confirmation. In a pipe or CI it refuses unless --yes is given, so nothing disappears silently.
--help returns the real actions
samuraiapp <resource> --help returns the actual actions and arguments, so nothing has to be guessed.
Connect your own machine
Keep your machine connected with tunnel, and the conversation's AI can work inside the folders you allow.
Skills: hand the AI the manual#
So that AI agents use it without guessing, Samurai App ships its usage guide in a form the AI reads.
A skill for Claude Code
Download samuraiapp-cli.zip from Connect your PC on the screen and extract it under ~/.claude/skills/. It arrives with your host and tenant ID already filled in.
Skills bundled with the product
The AI inside Samurai App also carries working patterns as bundled skills, one set per role. Send a request to a conversation from an outside AI and it runs with those patterns.
Choose the AI that answers your conversations#
A Samurai App conversation can be handed to Claude Code on your own PC, or run on models through an API key your company already pays for.
Hand it to Claude Code on Claude Pro / Max
Register Claude Code with tunnel, and Your local Claude Code appears under the AI that runs it for a new conversation. The official claude command runs on your PC, and Samurai App never receives your Claude credentials.
Bring your own key (BYOK)
Register an API key for OpenRouter, OpenAI, Gemini, an OpenAI-compatible endpoint, Azure OpenAI, or Amazon Bedrock. Keys can be personal or shared across the tenant, and a registered key's value can never be read back.
Same permissions as the screen, same records#
Using it from an outside AI changes none of the safeguards of the screen.
Same authorization as the screen
Tools and commands go through the same authorization as the screen. What you cannot see on screen, an outside AI cannot see either.
People press approve
When an agent stops for approval, there is no way to pass it automatically over MCP. An answer given over MCP is recorded as the token owner's action.
Secret values stay unreadable
Secrets can be stored and used, but there is no way to read the values back.
Network restrictions apply
Outside the networks an administrator allows, signing in works but nothing can be read.
Leaving takes effect at once
The tenant is named by the URL, not baked into the token. Leaving, a role change, or a suspension applies immediately.
Human decisions stay on screen
Network restrictions, single sign-on settings, registering SaaS credentials, and removing members cannot be done from an outside AI.
Frequently asked#
Does it work with ChatGPT?
Yes. Paste the MCP URL into ChatGPT's connector settings, sign in to Samurai App, and allow access. Claude, Claude Code, and other MCP clients use the same URL.
Do I need to issue an API key?
No. Sign-in is OAuth with PKCE, and the client only receives the URL.
What can it do?
It handles 21 resources, including sending to conversations, searching knowledge, reading and writing database rows, editing SIROE wiki pages and canvases, files, skills, and triggers. A request sent to a conversation returns once the agent's reply is final.
Can I use more than one tenant?
Yes. Add one connector per tenant. Only the tenant ID in the URL differs.
Is this different from connecting Samurai App to outside MCP servers?
It is the opposite direction. This page covers Samurai App acting as the server. Connecting Samurai App to outside services is under Connections in the feature list.
Can my Claude subscription run Samurai App conversations?
Yes, with Claude Pro / Max. The official Claude Code on your PC answers the conversation. It works for your own private conversations, and only Samurai App tools are available in them.
Everyday work comes together, connects, and becomes the company's strength.