Coding agents for Salesforce: Claude Code, Codex, Copilot

In 2026 every major coding agent can do Salesforce work, from Claude Code to OpenAI Codex and GitHub Copilot. The reason sits with Salesforce: its skills and hosted MCP servers work with any compatible agent. Which tool your team or your partner uses therefore matters less for code quality than the rules file, the review and the pipeline. For a manufacturer, four questions drive the choice: where code and data are processed, what the contract says, how usage is billed and who controls the rules. There is no independent Apex benchmark that names a winning model.

Which coding agents can work with Salesforce?

A coding agent is an AI tool that reads files, runs commands, starts tests and prepares changes as a commit or pull request on its own. Salesforce’s own tool, Agentforce Vibes, is covered in chapter 2. This chapter is about the external agents your developers or your implementation partner often already have on their laptops.

Tool	Vendor	Salesforce integration	Billing
Claude Code	Anthropic	official salesforce-development plugin with about 40 skills and three MCP servers	seat with allowance (Team, Enterprise) or tokens via API and cloud providers
OpenAI Codex	OpenAI	sf-skills library, named compatible by Salesforce	credits per token, included in ChatGPT plans
GitHub Copilot	GitHub	MCP client in VS Code, Agent HQ with Claude and Codex agents	seat with AI credits at $0.01
Cursor	Anysphere	named compatible by Salesforce for sf-skills and hosted MCP	check prices with the vendor before use
Agentforce Vibes (for comparison)	Salesforce	native, skills preinstalled	Flex Credits or unmetered licence

For developers

What the Claude Code plugin brings

Salesforce calls the Claude Code plugin "our first official plugin, built for Claude Code" (opens in a new tab). It bundles about 40 of the more than 100 skills in the open sf-skills library and sets up the MCP servers salesforce-api-context, salesforce-metadata-experts and salesforce-lsp. For OpenCode, Codex and Cursor, Salesforce names the same skills library as installable; I found no official plugin for those agents in the Salesforce documentation. According to the Summer '26 guide (opens in a new tab), the hosted MCP servers connect to "any MCP-compatible AI client, such as Claude, ChatGPT, Cursor". GitHub moves Copilot one level up with Agent HQ (opens in a new tab): since February 2026, Claude and Codex agents run there next to GitHub's own agent, at launch as a public preview for Copilot Pro+ and Enterprise.

In practice, Claude Code ships the most ready-made Salesforce setup today. With the other agents, your team configures skills and MCP servers itself. The full comparison by contract, EU data residency, cost model and governance is in Agentforce Vibes, Claude Code or Copilot. If anyone still runs the local Salesforce DX MCP Server, the required switch is in chapter 3.

Why the tool matters less than rules and pipeline

At TDX 2026 Salesforce announced that everything on the platform becomes an API, an MCP tool or a CLI command, and according to SalesforceBen (opens in a new tab) it named Claude Code, Cursor, Codex and Windsurf explicitly. An agent’s Salesforce knowledge therefore comes from skills, MCP servers and your project rules more than from the tool itself.

The rules live in a rules file in the repository, such as CLAUDE.md or AGENTS.md. It describes naming conventions, forbidden patterns and the definition of done. According to the GitHub changelog (opens in a new tab), the Copilot coding agent reads AGENTS.md, CLAUDE.md and GEMINI.md alongside GitHub’s own formats. One well-kept file in the client’s repository therefore works across several tools. Why that file should belong to the client is covered in the AI rules file in your Salesforce repository, and the general principle in why clients should own their DevOps tools.

No benchmark for Apex

I found no independent, public benchmark that measures Claude, GPT and Gemini on Apex, LWC or Flow. Salesforce claimed better accuracy (opens in a new tab) for its earlier code model xGen-Code but published no scores, only an internal survey with about 125 minutes saved per week. What does exist is security data: Veracode’s 2026 GenAI Code Security Report (opens in a new tab) measures a 56 percent security pass rate for AI-generated code, and only 30 percent for Java. Apex is syntactically close to Java, so I would take that number seriously.

56 %

security pass rate for AI-generated code

30 %

security pass rate for Java

125 min

saved per week according to an internal Salesforce survey, without scores

Salesforce’s own choice is a signal. According to the GA announcement (opens in a new tab), Agentforce Vibes is “powered by Claude and the Mastra agent orchestration engines”, and according to the Claudeforce press release (opens in a new tab) Salesforce plans to make Claude Code and Claude Enterprise available to all of its own developers and knowledge workers. A signal does not replace a measurement on your code. I recommend a small test suite of your own: 10 to 20 tasks from your backlog, for example a trigger handler with a bulk test, an LWC for the dealer portal and a flow change in the warranty process. Measure deployment success, test results and the number of review findings. The checks every piece of AI code must pass are in chapter 8.

Anthropic, OpenAI and Google: what the partnerships mean

Separate two decisions: which model your Agentforce agents use at runtime, for example a service agent answering dealer questions, and which coding agent your developers use. Salesforce’s partnerships affect the two differently.

Vendor	In Agentforce at runtime	In development	My reading
Anthropic	Claude via Amazon Bedrock inside the Salesforce Trust Boundary; a reasoning model for the Atlas Reasoning Engine	default in Agentforce Vibes; Claude Code for Salesforce employees	deepest integration; customers get the plugin and skills, not Claude Code licences
OpenAI	GPT-4o to GPT-5.6, operated by partners in a “shared trust zone”	GPT models selectable in Vibes; Codex with sf-skills	broad model choice, EU data residency for API projects
Google	Gemini 2.5 to 3.6 via Vertex AI; Gemini Enterprise GA in the Reasoning Engine according to SalesforceBen	the Vibes announcement names only Claude and GPT	relevant for Agentforce, hardly for the code path

Sources: Supported Models (opens in a new tab), Anthropic, October 2025 (opens in a new tab), Claudeforce (opens in a new tab), SalesforceBen on Dreamforce 2026 (opens in a new tab).

Where are code, metadata and test data processed?

A manufacturer’s repository holds more than people think: pricing logic, discount approvals, dealer terms and the interface definitions to SAP. The prompts are even more sensitive. Anyone who hands an agent error logs from a full-copy sandbox sends customer and dealer data to the model. That is why the question of where processing happens comes before the choice of tool.

Route	Where processing happens	Open point
Anthropic first-party API (also Claude Code via the Console)	inference “us” or “global”, workspace storage US only	no EU option
Claude via Amazon Bedrock or Google Cloud	regional endpoints, according to Anthropic at a 10 % premium over global	which models in which EU region, check with the cloud provider
Claude Code on Team or Enterprise seats	no EU option found in the sources checked	clarify contractually with Anthropic
OpenAI API with an EU project	in-region processing, requests and responses not stored	OpenAI does not say there whether Codex is covered
ChatGPT Enterprise with an EU workspace	storage in Europe, optional in-region inference since January 2026	same question for Codex
GitHub Copilot	GitHub Enterprise Cloud with data residency exists	ask GitHub where Copilot model calls run

Sources: Anthropic data residency (opens in a new tab), Anthropic pricing (opens in a new tab), OpenAI data residency in Europe (opens in a new tab), GitHub changelog, September 2026 (opens in a new tab).

German labour law adds another step. According to CMS (opens in a new tab), AI systems are generally capable of evaluating performance or behaviour data, so introducing them usually requires a works agreement, and using them without one is not permitted. How to bring in the works council and GDPR early is covered in AI coding tools, works council and GDPR.

What do coding agents cost per developer?

All major vendors now bill by tokens. The seat price buys an allowance, not unlimited use.

Tool	Billing	Vendor’s reference figure
Claude Code	tokens; on Team and Enterprise from a seat allowance with five-hour and weekly windows	about $13 per active day, $150 to $250 per month, 90 % of users below $30 per active day
OpenAI Codex	credits per million tokens, since April 2026	$100 to $200 per developer per month; a GPT-5.6 Sol task uses 5 to 40 credits
Copilot Business / Enterprise	$19 / $39 per user per month with 1,900 / 3,900 AI credits, pooled across the enterprise	1 credit = $0.01, consumption by model and tokens
Agentforce Vibes	Flex Credits in units of 2,000 tokens, or unmetered licence	premium model capped at 2,500 requests or 36M tokens per 30 days, then Flex Credits or GPT-5

Sources: Claude Code costs (opens in a new tab), Codex rate card (opens in a new tab), GitHub Copilot plans (opens in a new tab), Vibes usage and billing (opens in a new tab).

If you only look at Copilot’s seat price, you underestimate agentic use. 1,900 credits equal $19 of model usage. A developer who consumes as much as the Claude Code average of $13 a day uses up their share of the pool in a day and a half. The calculation for a whole programme, and whether client or partner should pay, is in what AI coding costs in a Salesforce project. Token prices and caching are covered in depth in chapter 7.

What about the Salesforce-specific DevOps agents?

The DevOps vendors shipped their own agents in 2026. They do not replace the coding agent; they govern what reaches production.

Agent	What it does
Gearset Cam	Takes a ticket through to a pull request and checks beforehand that the change is deployable, according to the announcement of 2 September 2026 (opens in a new tab). In pilots, Cam delivered up to 44 percent of a team’s pull requests, and at one customer 75 percent of those were merged. These are vendor figures.
Copado Agentia Headless	Brings Copado’s agents into the IDE and is generally available according to Copado (opens in a new tab).
Provar MCP	Is an open-source Salesforce CLI plugin that generates tests and validates them against more than 240 rules, according to Provar (opens in a new tab). It runs with Claude, Copilot and Cursor and costs nothing extra with a Provar Automation licence.

Neither Gearset nor Copado names the underlying model in these sources. For a manufacturer with data protection requirements that is a due diligence question: which model, which region, which retention.

Walk-through: choosing tools at a machine builder

Assume a machine builder with SAP S/4HANA, Service Cloud and a dealer portal on Experience Cloud. Three internal developers work with a partner who provides five nearshore developers. All numbers are assumptions. The internal developers have Copilot Business through the existing GitHub licence; the partner wants to use Claude Code with its own accounts.

This is how I would approach it:

Step	Approach
Set the approved list	Copilot stays, Claude Code is added, but through an account the client owns, either Enterprise seats or Bedrock with a regional endpoint. I rule out the partner’s own accounts, because then nobody knows where code and logs end up.
Rules file in the client’s repository	An AGENTS.md as the shared base that Copilot reads, plus a short CLAUDE.md that points to it. The internal architect owns it, not the partner.
Limit where agents work	Agents work only in developer sandboxes without production data. Dealer data from the full-copy sandbox does not go into prompts.
Estimate cost realistically	Eight developers with an assumed 15 active days cost about $1,560 a month at Anthropic’s $13 average. If all eight had Copilot Business, the seat price would be $152, but the pooled 15,200 credits cover only $152 of model usage. With similarly intensive agentic work, both tools end up in the same range.
Adjust the contract	The partner discloses which tools and models it uses, and the cost is either priced into the day rate or passed through. What belongs in the SOW is in partners using AI coding: contract rules.
Start with the test suite	15 tasks from the backlog, both tools, the same gates. The result decides, not individual developers’ preferences.

In my judgement, agreeing with the works council and data protection takes longer than choosing the tool in a setup like this, so plan that first. Steering several partners with different tools is part of my work in partner governance, and the gates belong in DevOps and release management. How Salesforce is shifting its partner model towards agents is the topic of chapter 5.