Eva · Tool Engine · MCP
Let an AI agent use your app. It gets your permissions, nothing more.
Evlet has one Tool Engine for every AI entry point. An assistant, an MCP client or an API key acts as the signed-in person, through the same routes and permission checks as the browser.
In shortEvlet is a Python framework with a built-in MCP server and AI assistant. Any tool call from Claude, ChatGPT or another MCP client runs as one signed-in user and passes the same row, field and action permission checks as the browser. AI access is off by default, read-only by default, has no SQL or file access, and every call is logged.
The permission model
This is the main point of the design. An assistant is never a super user.
It acts as the person
Every tool call runs as the signed-in person and is checked again each time it runs. Account, tenant, row, field and action checks are the browser's.
No back doors
Tools never get database, SQL, file system or shell access, and never act as anyone else. Ledger models are not reachable at all.
Off until you turn it on
Eva is off by default, and so is MCP access for each person. An admin turns on Allow MCP or Allow Eva per user in Controls.
Reads first
Remote MCP clients are read-only unless the instance allows writes and the person turns on "Can make changes" for that client.
Approvals for changes that matter
Moving a document, running an action and deleting wait for the person's approval in Evlet. The approval covers one call on one revision of the record.
Everything is logged
The Tool log keeps one row per call for 90 days: who, which client, which tool, the outcome. Field values are not stored in it.
Three ways in, one engine
Eva
The built-in assistant. It needs three switches: config.py, Controls and a per-user setting. The framework works the same with no AI provider configured.
Local MCP
evlet mcp is a stdio connector that needs only the Python standard library. It calls the same authenticated HTTP routes as the browser.
Remote MCP with OAuth
People add https://your-host/_/mcp as a custom connector in their assistant and sign in with their Evlet account. OAuth 2.1 with PKCE.
Set up local MCP
Sign in once from the shell. The command saves a session file, never the password. Then point your MCP client at evlet mcp.
mkdir -p ~/.evlet
chmod 700 ~/.evlet
evlet login --url http://127.0.0.1:8000 --username admin --session ~/.evlet/acme.session.json
evlet describe --session ~/.evlet/acme.session.json{
"command": "/absolute/path/to/.venv/bin/evlet",
"args": ["mcp", "--session", "/absolute/path/to/.evlet/acme.session.json"]
}The tools
A tool is listed only when it can apply, and listing grants nothing. Each call is authorized again by the route it uses.
| Tool | Kind | What it does |
|---|---|---|
discover_models, describe_model | Read | The models this person may see, with fields and schemas |
list_records, get_record | Read | Search, filter, sort and read records the person can read |
aggregate_records, record_history | Read | Exact totals and counts, and a record's activity |
list_reports, run_report | Read | Permitted reports and their rows |
create_record, update_record | Write | Create and change records, with backend validation |
change_state, run_action, call_service | Write | Move a document, run an app action or a service the app marked as a tool |
delete_record | Destructive | Delete a permitted record, after approval in Evlet |
Questions
Can an AI assistant see data the user cannot?
No. A tool call runs as the signed-in person through the same routes as the browser. Row scopes, field rules and model permissions apply, and restricted fields are removed from responses and schemas.
Can an assistant run SQL or read files on the server?
No. Tools have no database handle, no SQL, no Python execution, no file system and no shell access.
Which AI providers does Eva support?
The built-in adapter uses the OpenAI Chat Completions API. You choose the model names and may set a base URL. The adapter uses the Python standard library and adds no dependency.
Does Evlet work without any AI?
Yes. Eva is off by default and the framework works the same with no AI provider configured.
Is there a Python framework with a built-in MCP server?
Yes. Evlet ships a local MCP connector and a remote MCP endpoint with OAuth 2.1 as part of the framework. You do not add a separate MCP library or write tool definitions for your models.
How do I connect Claude or ChatGPT to my own business app?
Run the app on Evlet, turn on Allow MCP for your account, and add the app as an MCP server in your assistant: evlet mcp for a local stdio server, or https://your-host/_/mcp as a remote connector with sign-in.
Which MCP clients work?
The local connector uses stdio and the remote endpoint uses MCP Streamable HTTP with OAuth 2.1, so clients that support either transport can connect. Check your client's guide for where to add a custom server.
Try it on a local site
Create a site, turn on Allow MCP for your account and connect your assistant. The article walks through each step.
Last updated . Checked against Evlet 0.4.3.