Sequentum

Sequentum MCP

[License: MIT]

The Sequentum MCP Server connects your AI coding assistant to Sequentum using the Model Context Protocol (MCP), giving your AI tools the ability to create web scraping agents, run management, scheduling, analytics, and more. Sequentum hosts and manages a remote MCP server with OAuth authentication, so there's nothing to install.

Tool Reference | Prompts Reference | Resources Reference | Troubleshooting | Changelog

Key Features

• Agent management: Build, list, search, and get detailed information about your web scraping agents.
• Run automation: Start, stop, and monitor agent executions with real-time status tracking.
• Schedule management: Create and manage automated schedules using cron expressions.
• Analytics & diagnostics: Get run statistics, error analysis, and suggested fixes for failures.
• Space organization: Manage agent workspaces and run batch operations across spaces.

Disclaimers

sequentum-mcp exposes your Sequentum account data to MCP clients, allowing them to view, run, and manage your web scraping agents. Keep your credentials secure and avoid sharing sensitive information that you don't want accessible to MCP clients.

Getting Started

Add the Sequentum MCP server to your client with this configuration:

{  "mcpServers": {    "sequentum": {      "url": "https://mcp.sequentum.com/mcp"    }  }}

Most clients support the OAuth configuration. Claude.ai and Claude Desktop are simpler still — Sequentum MCP is listed in Claude's connector directory, so there is nothing to enter by hand; see Claude.ai and Claude Desktop below. For other clients, when you first connect, you'll be prompted to:

1. Log in with your Sequentum account
2. Accept the OAuth authorization
3. Grant access to the necessary permissions

Once authenticated, all tools become available in your client. For client-specific setup details, see Set Up Your Client below.

Set Up Your Client

Select your client below for specific setup instructions. All clients use the remote OAuth server at https://mcp.sequentum.com/mcp unless noted otherwise.

Cursor

Go to Cursor > Settings > Cursor Settings > MCP and follow the prompts to add the Sequentum MCP server. Cursor 1.0+ includes native OAuth and Streamable HTTP support.

You can also add the server manually by editing your mcp.json file using the configuration above.

Claude.ai and Claude Desktop

Sequentum MCP is listed in Claude's connector directory, so there is no URL to enter by hand and no custom connector to create.

1. Open Settings and go to Connectors.
2. Browse the Directory, select Connectors, and find Sequentum MCP.
3. Click Connect and sign in with your Sequentum account.

Team and Enterprise workspaces: connector availability is governed at org level, so an admin may need to enable Sequentum MCP before members can connect from their own accounts.

Enabling per conversation: once connected, enable Sequentum in an individual conversation via the + button near the message composer, then select Connectors.

ChatGPT

Note: While the Sequentum app is pending directory approval, you can connect via Developer Mode. Apps & Connectors → Developer Mode is currently available on Plus, Pro, Business, Enterprise, and Education plans (Education is web-only). On Business / Enterprise / Education accounts, only workspace owners and admins can access Advanced settings — regular members will not see the option. See OpenAI's Developer Mode documentation for current eligibility.

1. In ChatGPT, go to Settings > Apps & Connectors > Advanced settings and enable Developer mode.
2. Navigate to Settings > Apps & Connectors and click Create app (it appears once Developer mode is enabled).
3. Enter the connector name Sequentum and URL: https://mcp.sequentum.com/mcp
4. Click Create. You'll be prompted to sign in with your Sequentum account via OAuth.

Once connected, enable Sequentum in a conversation via the + button near the message composer, then select your connector from the list.

Claude Code

Connect Sequentum MCP once from the connector directory (see Claude.ai and Claude Desktop above) and it is available in Claude Code on the same account.

To add it directly instead — useful per-project or in scripts:

claude mcp add --transport http sequentum https://mcp.sequentum.com/mcp

Then launch Claude Code with claude and type /mcp to trigger the OAuth browser flow.

VS Code / GitHub Copilot

Open the Command Palette with Ctrl+Shift+P (or Cmd+Shift+P on macOS) and select MCP: Add Server. Enter the Sequentum MCP server URL:

https://mcp.sequentum.com/mcp

Windsurf

Configure via the Configure MCP option in Cascade (Cmd+L or Ctrl+L). Add the Sequentum MCP server URL:

https://mcp.sequentum.com/mcp

Other Clients

The Sequentum MCP Server follows standard MCP protocols and works with any client that supports:

• OAuth authentication (recommended)
• Streamable HTTP with automatic SSE fallback

Use the server URL https://mcp.sequentum.com/mcp in your client's MCP configuration.

The server supports Client ID Metadata Documents (CIMD) as the preferred client identification method, with Dynamic Client Registration (RFC 7591) as a fallback. MCP clients that support CIMD (such as Cursor) can use their own URL as a client_id without any prior registration.

Local stdio mode (removed)

Running the server locally over stdio with SEQUENTUM_API_KEY (npx sequentum-mcp) was removed in 3.0.0. Connect to https://mcp.sequentum.com/mcp instead — see Set Up Your Client. Sequentum API keys still work for the REST API.

Example Usage

Once connected, try these prompts to start using Sequentum context in your AI assistant:

What agents ran yesterday?Run agent <agent name> now.Is agent <agent name> still running?What agents are scheduled to run today?Download the extracted data from agent <agent name>.How many records were found the last time <agent name> was run?What is my current balance?Which agents cost the most this month?Schedule agent <agent name> to run every Monday at 9am.Look at the run log for <agent name> run at 9:22am. What caused the agent to fail?Show me the cost breakdown for agent <agent name> in January.What were the most expensive runs for agent <agent name>?How much did I spend on server time vs exports last week?

Available Tools

The Sequentum MCP Server provides 43 tools across 9 categories for interacting with the Sequentum platform. See the Tool Reference for detailed documentation.

• Agent Management (5 tools)
  • list_agents
  • get_agent
  • search_agents
  • get_agent_search_count
  • get_personal_agent_count
• Run Management (7 tools)
  • get_agent_runs
  • get_agent_run_summary
  • get_run_status
  • start_agent
  • stop_agent
  • kill_agent
  • delete_run
• File Management (2 tools)
  • get_run_files
  • get_file_download_url
• Version Management (2 tools)
  • get_agent_versions
  • restore_agent_version
• Schedule Management (8 tools)
  • list_agent_schedules
  • get_agent_schedule
  • create_agent_schedule
  • update_agent_schedule
  • enable_agent_schedule
  • disable_agent_schedule
  • delete_agent_schedule
  • get_scheduled_runs
• Billing & Credits (6 tools)
  • get_credits_balance
  • get_spending_summary
  • get_credit_history
  • get_agents_usage
  • get_agent_cost_breakdown
  • get_agent_runs_cost
• Space Management (6 tools)
  • list_spaces
  • get_space
  • get_space_agents
  • get_space_agent_count
  • search_space_by_name
  • run_space_agents
• Analytics & Diagnostics (4 tools)
  • get_runs_summary
  • get_records_summary
  • get_run_diagnostics
  • get_latest_failure
• Agent Builder (3 tools)
  • start_agent_build
  • get_agent_build_status
  • stop_agent_build

Structured output

36 of the 43 tools declare an outputSchema and return structuredContent next to the usual text block (MCP 2026-07-28). The text block is unchanged, so clients that ignore structured content see no difference. The seven acknowledgement-only tools (stop_agent, kill_agent, delete_run, restore_agent_version, delete_agent_schedule, enable_agent_schedule, disable_agent_schedule) return text only.

• start_agent returns { "run": … } in async mode and { "data": … } in sync mode.
• Tools that return a list (list_spaces, get_space_agents, list_agent_schedules, get_scheduled_runs, get_agent_versions, get_run_files) have an array schema. Clients on a 2025 protocol revision receive it wrapped as { "result": [...] }, as the SDK projects it.
• Status-like fields are numeric codes; their meanings are in each schema's descriptions.

Available Prompts

The server includes 9 reusable prompt templates that guide the AI through common multi-step workflows. See the Prompts Reference for detailed documentation.

• Debugging & Diagnostics
  • debug-agent -- Diagnose why an agent is failing
  • compare-runs -- Compare last successful vs failed run
• Health & Monitoring
  • agent-health-check -- Comprehensive health overview for an agent
  • daily-operations-report -- Daily ops report across all agents
  • space-overview -- Overview of all agents in a space
• Execution
  • run-and-monitor -- Start an agent and monitor until completion
  • schedule-agent -- Walk through creating a schedule
• Billing & Costs
  • spending-report -- Spending and credits report
  • cost-analysis -- Analyze costs across agents
• Agent Building
  • build-agent-from-prompt -- Build a new agent from a natural language description
  • inspect-agent-draft -- Inspect a build session and decide whether to save or discard

Available Resources

The server exposes 18 read-only resources (7 static + 11 templates) that AI clients can browse and pull into context. See the Resources Reference for detailed documentation.

• Static Resources (7)
  • sequentum://agents -- First page of all agents
  • sequentum://spaces -- All accessible spaces
  • sequentum://billing/balance -- Current credits balance
  • sequentum://billing/spending -- Monthly spending summary
  • sequentum://billing/agents-usage -- Top agents by cost
  • sequentum://analytics/runs -- Runs in the last 24 hours
  • sequentum://analytics/upcoming-schedules -- Scheduled runs for next 7 days
• Resource Templates (11)
  • sequentum://agents/{agentId} -- Agent detail with configuration
  • sequentum://agents/{agentId}/versions -- Agent version history
  • sequentum://agents/{agentId}/schedules -- Agent scheduled tasks
  • sequentum://agents/{agentId}/cost-breakdown -- Agent cost by usage type
  • sequentum://agents/{agentId}/runs -- Agent run history
  • sequentum://agents/{agentId}/runs/{runId} -- Specific run status
  • sequentum://agents/{agentId}/runs/{runId}/files -- Run output files
  • sequentum://agents/{agentId}/runs/{runId}/diagnostics -- Run error diagnostics
  • sequentum://agents/{agentId}/latest-failure -- Most recent failure diagnostics
  • sequentum://spaces/{spaceId} -- Space details
  • sequentum://spaces/{spaceId}/agents -- Agents in a space

Troubleshooting

Verifying scope enforcement end to end

pnpm run probe --env qa --mode log-only runs the real OAuth flow against QA (browser login and consent, loopback callback, PKCE), then calls the V1 API directly and through this server with tokens of three scope profiles and checks CloudWatch for the Control Center's scope-check lines. See docs/oauth-scope-probe.md.

The server checks each tool's and resource's scopes before dispatch for user tokens and for tokens issued to an agent registered through the Control Center. Client-credentials tokens are exempt, as they are on the Control Center itself.

For more troubleshooting help, see the Troubleshooting Guide.

HTTP Mode Configuration

The hosted server at mcp.sequentum.com runs the Streamable HTTP transport, the only transport as of 3.0.0, behind its own deployment configuration. The following environment variables tune caching, rate limiting, and proxy trust for that transport. They are documented for readers of this source — there is no supported self-hosted deployment, and no container image is published.

HTTP mode is stateless as of 2.0.0. There is no Mcp-Session-Id header and no per-client session state on the server: every request is handled independently by a fresh MCP server instance. GET /mcp no longer opens an SSE stream — it now returns 405 Method Not Allowed (or 401 first if the request is unauthenticated). DELETE /mcp is a no-op that always answers 200, since there is no session left to tear down. MAX_SESSIONS is no longer read.

Rate limiting across replicas

The rate limiter is per-process and keyed on req.ip; it holds no state shared across replicas. With the stateless transport there is no session affinity, so a given client's requests are not pinned to one replica — in a horizontally scaled deployment of N replicas behind a load balancer, the effective cluster-wide ceiling becomes N × MCP_RATE_LIMIT_MAX per window, not MCP_RATE_LIMIT_MAX. To hold a specific global rate, divide MCP_RATE_LIMIT_MAX by your replica count.

TRUST_PROXY and rate-limit evasion

With the default TRUST_PROXY=true, Express trusts every proxy hop and resolves req.ip from the client-supplied, leftmost entry of the X-Forwarded-For header. Because the rate limiter keys on req.ip, a client that simply rotates that header value can evade rate limiting entirely. Running behind a reverse proxy does not fix this by itself: tunnels such as cloudflared and ngrok append to X-Forwarded-For rather than overwrite it, so the attacker-controlled leftmost entry survives unless TRUST_PROXY is configured correctly.

The remediation is to set TRUST_PROXY to the exact number of trusted reverse-proxy hops in front of the server (e.g. 1), or to a comma-separated CIDR/IP allowlist of your trusted proxies, so Express derives req.ip from the correct hop instead of trusting a client-supplied header. This repository does not know your deployment topology, so it deliberately does not choose a safer default for you.

Severity: this is an abuse/DoS-protection bypass, not an authentication or data exposure issue — no account data is exposed, and no tool call succeeds that would otherwise be rejected. It only lets a client avoid being rate-limited.

CORS Origin Allowlist

When the MCP server is accessed from a browser (e.g. the Claude web app or the ChatGPT connector), it checks the Origin header against an allowlist. By default the following origins are permitted:

• https://claude.ai, https://claude.com, and all subdomains (e.g. team.claude.ai)
• https://chatgpt.com, https://platform.openai.com, and all subdomains under chatgpt.com (e.g. connector.chatgpt.com)
• https://glama.ai and all subdomains (for Glama's browser-based MCP inspector)
• https://dashboard.sequentum.com
• https://mcp.sequentum.com
• http://localhost:<port>, http://127.0.0.1:<port>, and http://[::1]:<port> when DEBUG=1

To add your own origins (e.g. an internal dashboard), set the ALLOWED_ORIGINS environment variable to a comma-separated list of exact origins:

ALLOWED_ORIGINS="https://my-dashboard.example.com,https://other.example.com"

These origins are appended to the defaults — Claude, ChatGPT, and Sequentum access is preserved. Wildcards and regular expressions are not supported via the env var; if you need a subdomain wildcard, add a RegExp entry directly in src/server/cors.ts.

Note: Origin matching is case-sensitive and does not include a path or query string. Native MCP clients (Cursor, Claude Desktop, Claude Code) send no Origin header and are not affected by this allowlist.

Development

This repository uses pnpm, not npm. npm install is refused on purpose. You need Node 22 or newer (.nvmrc pins 24, which the Docker image runs). Corepack installs the pnpm version named in package.json. It ships with Node 22 and 24; on Node 25 or newer, install it first with npm install -g corepack:

corepack enablepnpm installpnpm testpnpm run build

pnpm runs no dependency install scripts unless the package is listed under allowBuilds in pnpm-workspace.yaml. If pnpm install reports an ignored build script after a dependency change, add the package there: true if it needs the script to work, false otherwise.

Arguments go straight after the script name, without npm's -- (pnpm run probe --env qa --mode log-only). pnpm passes a -- on to the script.

Privacy Policy

The Sequentum MCP Server accesses your Sequentum account data — including agent metadata, run history, scheduled tasks, billing information, and output files — solely to fulfill the requests you make through your AI assistant. By default, the MCP server acts as an authenticated proxy between your MCP client and the Sequentum API: request data is forwarded to the API and responses are returned to your client without being persisted or shared with third parties.

Operators may enable verbose request logging via the DEBUG=1 environment variable for troubleshooting. In that mode the server redacts Authorization, Cookie, and x-api-key headers, but writes request bodies (which may include tool arguments) to stderr. The hosted server at mcp.sequentum.com does not run with DEBUG=1.

For the full Sequentum privacy policy, see https://www.sequentum.com/privacy-policy.

Links

• Sequentum MCP Server
• Sequentum Dashboard
• Sequentum API Documentation
• Model Context Protocol

License

MIT © Sequentum