Aident AI
How to Connect Claude Code and Codex to Real-World Tools
Claude Code and Codex become much more useful when they can do more than edit files. With Aident Loadout, the same agent can discover an integration when a task requires it, inspect the exact input contract, connect the account through Vault, and execute the action without putting the provider credential in the prompt.
That means you can ask for an outcome—research a market, create a product video, triage an incident, update a CRM, or publish a report—without installing and maintaining a separate MCP server for every service.
The practical pattern is:
Search for the capability that matches the job.
Inspect its schema and cost before execution.
Connect only the required integration.
Run read actions first.
Review any public, destructive, or paid action before approving it.
This guide shows how to use that pattern in Claude Code or Codex and which integration combinations produce real work instead of a long list of logos.
What Integrations Can You Use?
The live Loadout tool directory is the source of truth because the catalog changes as integrations and actions are added or updated. It currently covers hundreds of services and several execution types:
Communication: Gmail, Outlook, Slack, Microsoft Teams, Lark, WhatsApp, and SendGrid
Engineering: GitHub, GitLab, Sentry, PagerDuty, PostHog, Supabase, CircleCI, and Cloudflare
Project management: Linear, Jira, Asana, Monday, Basecamp, and Wrike
Sales and support: Salesforce, HubSpot, Pipedrive, Attio, Zendesk, Freshdesk, and Intercom
Research and data: Parallel, SerpAPI, Ahrefs, Reddit, YouTube, Hacker News, and product analytics tools
Media: Fal, HeyGen, Gemini, Gladia, Cloudinary, Instagram, X, and YouTube
Content and commerce: Contentful, Shopify, Mailchimp, Square, and other publishing, marketing, and payment systems
Several recently active capabilities make the catalog especially useful outside a code repository. Fal can generate short videos from a text prompt; HeyGen can create avatar videos and speech; Gladia can transcribe audio or video; Hacker News can be searched directly; and the Lark CLI integration can create documents, spreadsheets, charts, and messages.
Do not try to enable the whole catalog at once. Search for the job in front of you and connect the smallest useful set.
Set Up Loadout
If Loadout is not installed, give your agent the canonical setup guide:
Read https://loadout.aident.ai/SETUP.md and set up Aident Loadout.
Then verify the CLI and your Vault status:
Search in plain language. You do not need to know a provider's action name:
Search results return canonical names such as direct:hacker_news_tools:search or composio:heygen_tools:heygen_v2_video_generate. Copy that exact name when you inspect or execute the action:
The schema tells you the required fields, whether the action reads or writes data, and whether pricing is fixed or variable. Treat that schema as the contract; do not guess field names from a provider's marketing page.
Workflow 1: Turn Market Signals Into a Brief
A useful research workflow needs independent evidence, dates, and links—not one search result rewritten by a model.
Use several integrations for different parts of the evidence:
Search Hacker News, Reddit, X, LinkedIn, and YouTube for recurring questions.
Use SerpAPI or Ahrefs to check the current search results and competing pages.
Use Parallel for deeper web research when the question spans many sources.
Use PostHog or another product analytics integration to check whether the topic relates to actual product behavior.
Save the evidence to Lark, Google Drive, Notion, or a project tracker.
The Hacker News search action is read-only and makes a good first test:
Expected result: a dated list of matching stories with links and engagement fields. If the result is noisy, narrow the wording or switch the type to comment; do not treat the total number of indexed items as demand.
This workflow can produce a launch brief, competitor memo, customer-problem report, or weekly trend digest. The value comes from the evidence trail: a reviewer can open the sources and challenge the conclusion.
Workflow 2: Create a Product Video From a Brief
A coding agent can now coordinate a media pipeline without pretending video generation is a coding task.
A practical sequence is:
Research the audience and message with Reddit, YouTube, X, and web search.
Draft a short script with Gemini or another text model.
Generate a visual clip with Fal, or an avatar-led explainer with HeyGen.
Transcribe the result with Gladia to check the spoken copy.
Store the approved asset in Cloudinary or Drive.
Publish through YouTube, Instagram, or X only after human review.
Start by finding and inspecting the generation action:
The Fal action exposes duration, resolution, and aspect ratio. It also has variable pricing, so a call may return a quote that requires explicit approval before generation. That is the correct stopping point for an autonomous agent: review the prompt and cost, then approve or revise.
Expected result after an approved generation is a returned asset ID or video URL. Reuse that output; do not rerun the generation because the first result is still processing or because the URL was not copied into the next step.
Workflow 3: Move From an Incident to a Verified Fix
An incident workflow is valuable when each integration owns one part of the response:
Sentry or PostHog supplies the error and affected path.
PagerDuty supplies incident state and responders.
GitHub supplies the relevant code, issue, and pull request.
CI verifies the change.
Slack or Lark communicates status.
Ask the agent to gather evidence first: error frequency, first and last occurrence, release version, affected users, and reproduction steps. Then let it prepare a minimal patch and test plan. Keep acknowledgment, deployment, public status updates, and incident resolution behind explicit approval.
Expected result: one issue or pull request linked to the original incident, a passing targeted test, and a message that distinguishes verified facts from the suspected root cause.
Common failure: letting the agent create several tickets and messages before it has grouped duplicate errors. Read and deduplicate first; write once.
Workflow 4: Convert Customer Feedback Into Shipped Work
Combine a support inbox with product and engineering systems:
Read recent tickets from Zendesk, Freshdesk, Intercom, Gmail, or Outlook.
Group repeated problems by exact wording, affected feature, and severity.
Check PostHog for the corresponding user path or failure event.
Create one Linear, Jira, or GitHub issue with evidence and acceptance criteria.
Post a concise summary to Slack, Teams, or Lark.
After the fix ships, draft the customer reply for review.
Expected result: a traceable chain from customer language to product evidence to one prioritized issue. Names, emails, tokens, and private message bodies should remain in the connected systems rather than being copied into a public issue or prompt.
Workflow 5: Research and Update a Sales Account
A sales workflow can combine enrichment, communication, and CRM actions:
Apollo or web research finds current company and role evidence.
Gmail or Outlook provides the existing conversation.
Calendar checks available meeting windows.
Salesforce, HubSpot, Pipedrive, or Attio stores the account update.
Sheets or Lark creates a reviewable pipeline report.
Use read actions to assemble the account brief before modifying the CRM. Require a human to approve outbound email and material changes to deal stage, owner, value, or close date.
Expected result: one cited account summary, a draft message, and a structured CRM update whose fields can be reviewed before writing.
Workflow 6: Build a Research-to-Publishing Loop
The same integrations can form a repeatable content loop:
Discover a specific question with social and search data.
Collect primary sources and competing answers.
Draft in Lark, Notion, Drive, or the repository.
Generate or prepare an image or video.
Publish through the target CMS after review.
Measure the exact URL in PostHog, Search Console, or Ahrefs.
Feed evidence-backed lessons into the next topic decision.
The important unit is the canonical URL, not the number of items in the CMS. A published item proves publication; it does not prove traffic, engagement, signups, or revenue.
A Safe Operating Checklist
Use this checklist whenever an agent crosses from code into a business system:
State the outcome. “Create a weekly incident report” is better than “use Slack and Sentry.”
Search before connecting. Confirm that a relevant action exists.
Inspect the schema. Check required inputs, read/write behavior, and pricing.
Connect narrowly. Add only the integration required for the next step.
Read before writing. Gather and deduplicate source data first.
Preview side effects. Show the message, record update, generated prompt, or publish target.
Approve sensitive actions. Keep public posts, payments, deletions, production changes, and paid generation behind confirmation.
Verify the result. Re-read the record or open the public URL after execution.
Keep a receipt. Record what ran, when, against which account, and with what result.
Troubleshooting
“Integration not connected”
Check the capability's requiredIntegrationIds, then request a Vault connection:
Open the returned connection URL yourself. Do not paste provider secrets into the agent conversation.
Search returns the wrong action
Describe the verb, object, and expected output: “list unresolved PagerDuty incidents with service and urgency” is better than “PagerDuty.” Inspect the top few schemas before choosing.
The agent has too many tools
Keep the agent-facing surface small. Let it search the catalog, inspect only the selected schemas, and execute the few actions required for the task. A large static tool list makes selection harder and consumes context before any work starts.
A write action is expensive or destructive
Stop at the preview or quote. Review the exact input, target account, estimated cost, and rollback path. Approval should apply to that action, not to every future action from the same integration.
Why This Approach Works
Claude Code and Codex already support MCP, but access alone is not a workflow. Useful integration work also needs discovery, credential handling, schemas, approvals, execution, and verification.
Loadout keeps those concerns behind one capability layer. The agent can discover a task-specific action instead of loading every provider schema into context, while Vault resolves the connected account at execution time. You keep using the coding agent you prefer, and the business systems remain the systems of record.
Ready to try one outcome instead of installing another pile of connectors? Connect your first integration with Aident Loadout, then ask your agent to search for the exact job you want completed.
