AI assistants
The hands for the AI you already use.
Claude, ChatGPT, or Copilot decides. StitchOps executes inside your systems, with credentials vaulted and evidence on every run.
It decides. StitchOps executes.
The model chooses from typed tools. The agent owns the credentials and the coordinates, and stops for confirmation on anything destructive.
Your assistant
- Claude
- ChatGPT
- Copilot
Picks from typed tools. Never sees a credential.
MCP
The execution layer
- ·validates against the registry
- ·resolves secrets on the worker
- ·gates destructive steps
- ·logs every step
Deterministic. Same inputs, same outputs.
outbound only
TLS 1.2+
Your systems
- Vendor portals
- Desktop apps
- Terminals
- On-prem AD
Reached from a worker VM inside your network.
Tools
Every workflow is a tool.
Read and write tools are split so you grant them separately. Destructive tools stop for confirmation. Ask in plain language; get the evidence back.
I found a saved workflow for this. Running it on worker-01.
tool call
workflow_offboard_closed_apps
{ "email": "j.doe@example.com" }
Confirmation required
Step 5 deactivates the user in vendor-portal. This cannot be undone by the workflow.
completed in 5.2s
3 portals · 6 steps · evidence: step_05.png, step_11.png, step_17.png
Done. The user is removed from all three portals. Evidence is attached to run_01j8k.
Connect
Authorize once.
OAuth with dynamic client registration. No API keys in prompts. Any MCP client works; Claude, ChatGPT, and Copilot are the ones we test on every release.
Setup guide →Connect StitchOps
MCP server for your AI assistant
- Server
- https://mcp.stitchops.io
- Auth
- OAuth 2.1 · dynamic client registration
- Tenant
- acme-dealers
- Scopes
- readwrite
14 tools discovered
- workflow_searchread
- workflow_run_statusread
- workflow_authorwrite
- workflow_offboard_closed_appswrite · confirmation
- flex_data_queryread
- identity_user_lookupread
write tools stop for confirmation on destructive steps
What it can and cannot do.
The same allowlist applies whether a person or a model asks.
The model can
- Find and run saved workflows with typed inputs
- Check run status and read redacted results
- Author and validate new workflows
- Query flex data tables and connected identity systems
The model cannot
- See a credential value
- Run shell commands
- Skip a confirmation gate on a destructive tool
- Cross a tenant
Connect your assistant to a real workflow.
We set up one wedge workflow in your tenant and show it running from Claude, ChatGPT, or Copilot.