MCP - Agents First Docs | Agents First

Start

Overview
Quickstart
Concepts

App Foundations

Workspaces
Access portal

External Agents

External Agents
OAuth and auth.md
MCP
OpenAPI
Result contracts

Internal Agents

Internal Agents
Sales Command

CRM

Contacts
Inbox

Client Delivery

Project rooms

Automations

Lead intake

Billing And Access

Subscriptions

Troubleshooting

Login
Domains

Browse sections

Start

Overview Quickstart Concepts

App Foundations

Workspaces Access portal

External Agents

External Agents OAuth and auth.md MCP OpenAPI Result contracts

Internal Agents

Internal Agents Sales Command

CRM

Is this helpful?

AI tools

Client Delivery

Project rooms

Automations

Lead intake

Billing And Access

Subscriptions

Troubleshooting

Docs/External Agents

Public documentation

MCP

Model Context Protocol setup and support.

Is this helpful?

AI tools

CLI-first quickstart (canonical)

The live, unauthenticated copy-paste quickstart for Claude Code, OpenAI Codex, Claude web/Desktop, and ChatGPT is:

https://app.agentsfirst.ai/auth/connect.md

That page is rendered from the same connect copy as the in-portal guide (connect-copy.ts) — prefer it over any hand-copied snippet. OAuth discovery and the full lifecycle (discover → register → authorize → use → revoke) live at auth.md.

Connect Claude

In Claude web, desktop, or mobile, open Settings → Connectors (or ), add a custom connector, and paste the MCP URL. Then select , sign in to Agents First, and approve the requested access once.

Canonical page: agentsfirst.ai/docs/external-agents/mcp

Customize → Connectors

Connect

For Claude Code:

claude mcp add --transport http agentsfirst https://app.agentsfirst.ai/api/external-agent/mcp

Run /mcp in Claude Code and choose Authenticate to finish the browser OAuth flow.

Connect ChatGPT

In ChatGPT, enable Developer Mode in Settings → Apps → Advanced settings. Then open Apps & Connectors → Create, paste the MCP URL, choose OAuth authentication, and complete the Agents First sign-in and approval flow.

Connect Codex

codex mcp add agentsfirst --url https://app.agentsfirst.ai/api/external-agent/mcp
codex mcp login agentsfirst

The browser authorization flow opens during login. Use /mcp in Codex to confirm the connection after approval.

Manual /.codex/config.toml fallback

If codex mcp add is unavailable, add this block to ~/.codex/config.toml (never put a bearer token in this file):

[mcp_servers.agentsfirst]
enabled = true
url = "https://app.agentsfirst.ai/api/external-agent/mcp"
oauth_resource = "https://app.agentsfirst.ai/api/external-agent/mcp"

Older Codex builds: if the connection does not start the OAuth login, add experimental_use_rmcp_client = true to ~/.codex/config.toml (newer builds enable the OAuth-capable MCP client automatically).

Resources and semantic discovery

After initialize, use the existing MCP resources capability (no extra capability key) for:

| Method | What you get | | --- | --- | | resources/list | Compact indexes: ontology, processes, approved-business-context, plus principal-visible playbook URIs | | resources/templates/list | URI templates for exact concept / logical data-asset / process / playbook rows | | resources/read | One resource body; inaccessible and nonexistent ids share one enumeration-safe miss (no URI echo). Malformed URIs fail shape-only before catalog/tenant reads |

Stable resource URIs:

agentsfirst://ontology/index
agentsfirst://ontology/concepts/{conceptId}
agentsfirst://ontology/data-assets/{assetId} (logical assets only — never physical table refs)
agentsfirst://processes/index
agentsfirst://processes/{processId}
agentsfirst://business-context/approved (IDs/dispositions/authority only — never live CRM copies or raw source content)
agentsfirst://playbooks/{playbookId}

Operation discovery stays on the meta-tools (search_operations / describe_operation / execute_operation). Search accepts optional conceptId / processId filters and keeps compact byte-capped rows; describe returns bounded semantic fields and logical assets only. Every result is already filtered to the current OAuth principal, protected resource, and scopes.

Prefer get_playbook or playbook resources for recipes; call check_send_readiness before proposing any send job.

Troubleshooting

Use the exact /api/external-agent/mcp URL. Do not substitute the retired marketing MCP endpoint. If authorization fails, reconnect from your client and review the grant in the Agents First access portal.

See also OAuth and auth.md and Result contracts.