MCP Integration
Model Context Protocol (MCP) lets AI assistants like Claude, Cursor, Windsurf, and VS Code (with GitHub Copilot) use Truesight as a tool provider. Instead of writing API calls yourself, your AI assistant can manage datasets, build and deploy evaluations, run live scoring, review results, and perform error analysis, all through natural language.
Overview
Truesight exposes a curated set of tools through MCP. When connected, your AI assistant can:
- List, inspect, create, and upload datasets
- Configure dataset inputs and judgment criteria
- Create and deploy live evaluation endpoints
- Run live evaluations and score inputs on demand
- View evaluation results
- Browse, flag, and judge review queue items; promote reviewed data back to datasets
- Perform error analysis: suggest error notes, consolidate categories, and apply mappings
MCP connections use the same Platform API Keys as the REST API. The scopes you assign to the key determine which tools the assistant can use.
Quick setup
Step 1: Create a platform API key
- Go to Settings in Truesight
- Click Create Key in the Truesight API Keys section
- Enter a name (e.g., "Claude", "Cursor", or "VS Code")
- Select the scopes you want the assistant to have access to (or select all)
- Click Create
Step 2: Copy the MCP config
After creating the key, the key reveal dialog shows a collapsible Use with MCP section. Expand it and click Copy to copy the MCP configuration JSON to your clipboard.
The config looks like this:
{
"mcpServers": {
"truesight": {
"url": "https://api.truesight.goodeyelabs.com/mcp/",
"headers": {
"Authorization": "Bearer YOUR_API_KEY_HERE"
}
}
}
}
Step 3: Add to your MCP client
Paste the configuration into your MCP client's config file. See the client-specific instructions below.
Client setup
VS Code (GitHub Copilot)
Click a button to install, then replace YOUR_API_KEY_HERE with your platform API key. Requires VS Code 1.99+ with GitHub Copilot enabled.
Alternatively, add it manually:
- Open VS Code and press Cmd/Ctrl + Shift + P to open the Command Palette
- Run MCP: Add Server
- Select HTTP as the server type
- Enter the URL:
https://api.truesight.goodeyelabs.com/mcp/ - Give it the name
truesight
Or add it directly to your .vscode/settings.json:
{
"mcp": {
"servers": {
"truesight": {
"url": "https://api.truesight.goodeyelabs.com/mcp/",
"headers": {
"Authorization": "Bearer YOUR_API_KEY_HERE"
}
}
}
}
}
Replace YOUR_API_KEY_HERE with your actual platform API key.
Cursor
Click the button to install with one click, then replace YOUR_API_KEY_HERE with your platform API key in Cursor's MCP settings (~/.cursor/mcp.json).
Alternatively, add it manually to your project's .cursor/mcp.json:
{
"mcpServers": {
"truesight": {
"url": "https://api.truesight.goodeyelabs.com/mcp/",
"headers": {
"Authorization": "Bearer YOUR_API_KEY_HERE"
}
}
}
}
Replace YOUR_API_KEY_HERE with your actual platform API key and restart Cursor.
Claude Code
Add the server with one command, replacing YOUR_API_KEY_HERE with your platform API key:
claude mcp add --transport http truesight \
https://api.truesight.goodeyelabs.com/mcp/ \
--header "Authorization: Bearer YOUR_API_KEY_HERE"
This adds the server to your user-level config (available in all projects). To scope it to a specific project instead, add --scope project before the server name — this writes the config to .mcp.json in your project root.
Claude Desktop
- Open Claude
- Go to Settings (gear icon) > Developer > Edit Config
- This opens your
claude_desktop_config.jsonfile. Add the Truesight server to themcpServersobject:
{
"mcpServers": {
"truesight": {
"url": "https://api.truesight.goodeyelabs.com/mcp/",
"headers": {
"Authorization": "Bearer YOUR_API_KEY_HERE"
}
}
}
}
- Replace
YOUR_API_KEY_HEREwith your actual platform API key - Save the file and restart Claude
- You should see Truesight listed as an available tool provider
Windsurf
- Open Windsurf and press Cmd/Ctrl + , to open Settings
- Search for MCP or navigate to Cascade > MCP Servers > View raw config
- Add the Truesight server to the
mcpServersobject:
{
"mcpServers": {
"truesight": {
"serverUrl": "https://api.truesight.goodeyelabs.com/mcp/",
"headers": {
"Authorization": "Bearer YOUR_API_KEY_HERE"
},
"disabled": false
}
}
}
- Replace
YOUR_API_KEY_HEREwith your actual platform API key - Save and click Refresh (or restart Windsurf)
Workflow prompts
The Truesight MCP exposes two built-in prompts for clients that support the MCP Prompts primitive (such as Claude Desktop). These provide step-by-step workflow guidance without requiring any additional setup.
| Prompt | Description |
|---|---|
truesight_mcp_guide | Orchestration playbook covering scoring inputs, error analysis, and the review-and-promote loop. Includes a decision tree to help choose the right workflow. |
create_evaluation_workflow | Step-by-step guide for scoping and deploying a new live evaluation from scratch. |
For clients that don't support MCP Prompts, use the companion agent skills instead. See Agent Skills.
Available tools
Each tool requires a specific scope on your API key. If the key doesn't have the required scope, the tool will return an error.
| Tool | Description | Required scope |
|---|---|---|
list_datasets | List all datasets you have access to | datasets:read |
get_dataset | Get details for a specific dataset | datasets:read |
get_dataset_rows | Get rows from a dataset with pagination | datasets:read |
create_dataset | Create a new empty dataset with specified columns | datasets:write |
upload_dataset | Create a dataset and populate it with rows in one step | datasets:write |
configure_dataset_inputs | Set which column(s) contain the AI input to evaluate | datasets:write |
configure_judgment_columns | Configure how the evaluation judges AI outputs | datasets:write |
update_dataset_row | Update one or more column values in a single dataset row | datasets:write |
list_evaluations | List all evaluations you have access to | evaluations:read |
get_evaluation | Get evaluation details including config | evaluations:read |
deploy_live_evaluation | Deploy an existing evaluation as a live scoring endpoint | live-evaluations:write |
create_and_deploy_evaluation | Create and immediately deploy an evaluation in one step | evaluations:write, live-evaluations:write |
run_eval | Score inputs against a deployed live evaluation | live-evaluations:execute |
list_live_evaluations | List deployed live evaluations | live-evaluations:read |
list_results | List evaluation run results | results:read |
get_result | Get detailed results for an evaluation run | results:read |
list_review_items | List flagged review items | review:read |
flag_review_item | Flag an evaluation run's results for human review | review:write |
judge_review_item | Submit a human judgment for a single review item | review:write |
add_reviewed_items_to_dataset | Add all judged review items as new rows in the linked dataset | review:write |
suggest_error_notes | Get AI-suggested error notes and category for a trace | error-analysis:execute |
consolidate_error_categories | Get AI suggestions for merging similar error categories | error-analysis:execute |
apply_category_mappings | Apply error category consolidation mappings to all matching rows | datasets:write |
Pagination
All list tools accept optional page and page_size parameters. Responses include total, page, page_size, and total_pages fields alongside the items array. Defaults vary by tool:
| Tool | Default page_size | Range |
|---|---|---|
list_datasets | 25 | 1-500 |
get_dataset_rows | 25 | 5-10,000 |
list_evaluations | 25 | 1-500 |
list_live_evaluations | 25 | 1-500 |
list_results | 10 | 1-500 |
list_review_items | 50 | 1-500 |
Idempotency
The following mutating tools accept an optional idempotency_key parameter for safe retries: create_dataset, upload_dataset, deploy_live_evaluation, create_and_deploy_evaluation, and add_reviewed_items_to_dataset. If you provide the same idempotency key within a 10-minute window, the server returns the cached response instead of repeating the operation.
Agent skills
For clients that support agent skills (Claude Code, Cursor, and others), companion skills are available that give your AI assistant step-by-step workflow guidance for all Truesight MCP workflows.
Install both skills with:
# truesight-mcp skill: covers scoring inputs, error analysis, and the review loop
curl -fsSL https://raw.githubusercontent.com/Goodeye-Labs/truesight-mcp-skills/main/skills/truesight-mcp/SKILL.md \
-o .claude/skills/truesight-mcp/SKILL.md --create-dirs
# create-evaluation skill: covers scoping and deploying new live evaluations
curl -fsSL https://raw.githubusercontent.com/Goodeye-Labs/truesight-mcp-skills/main/skills/create-evaluation/SKILL.md \
-o .claude/skills/create-evaluation/SKILL.md --create-dirs
Or install globally (available in all projects):
curl -fsSL https://raw.githubusercontent.com/Goodeye-Labs/truesight-mcp-skills/main/skills/truesight-mcp/SKILL.md \
-o ~/.claude/skills/truesight-mcp/SKILL.md --create-dirs
curl -fsSL https://raw.githubusercontent.com/Goodeye-Labs/truesight-mcp-skills/main/skills/create-evaluation/SKILL.md \
-o ~/.claude/skills/create-evaluation/SKILL.md --create-dirs
The skills are hosted at github.com/Goodeye-Labs/truesight-mcp-skills.
Example interactions
Once connected, you can ask your AI assistant things like:
- "List my Truesight datasets"
- "Show me the rows in dataset ds_abc123"
- "Create a dataset from this CSV data and set up an evaluation for response quality"
- "Run an evaluation on this text using my customer_support live evaluation"
- "What are the latest evaluation results?"
- "Flag evaluation run run_abc123 for review"
- "Analyze the errors in my dataset and suggest categories"
- "Consolidate the similar error categories in my dataset"
The assistant will use the appropriate Truesight MCP tools to fulfill your request.
Troubleshooting
"Authentication required" or 401 errors
- Verify your API key is correct and hasn't been revoked
- Check that the key hasn't expired
- Make sure the
Authorizationheader is formatted correctly:Bearer ts_pat_...
"Missing required scope" errors
- The tool requires a scope your key doesn't have
- Create a new key with the needed scopes, or revoke and recreate with additional scopes
Tools not appearing
- Restart your MCP client after adding the configuration
- Verify the MCP endpoint URL is correct:
https://api.truesight.goodeyelabs.com/mcp/ - In Windsurf, make sure the server is not set to
"disabled": true - In VS Code, verify GitHub Copilot agent mode is enabled (requires VS Code 1.99+)
- Check your MCP client's logs for connection errors
"Subscription inactive" or 402 errors
- Your organization's subscription may need attention
- Check your billing status in the Truesight web app
Security notes
- MCP connections use the same authentication and access control as the REST API
- Your AI assistant can only access resources you have permission to view or modify
- Scope restrictions apply: the assistant can only use tools covered by the key's scopes
- Revoke the key from Settings to immediately disconnect all MCP clients using it
