At TrailblazerDX 2026 on April 15-16, Salesforce shipped three interlocking pieces: Headless 360, exposing the entire platform as APIs, MCP tools, and CLI commands; Agentforce Vibes 2.0, a cloud-hosted VS Code IDE defaulting to Claude Sonnet 4.5; and AgentExchange, a consolidated marketplace Salesforce reports at 13,000+ total listings backed by a $50M Builders Initiative1. For teams building on LangGraph, CrewAI, or AutoGen against Salesforce data, the calculus just changed.
What Salesforce Shipped
The Salesforce developer blog describes Headless 360 as two distinct MCP server variants. The DX MCP Server runs locally and covers developer tooling: metadata operations, Apex testing, and LWC. The Hosted MCP Server runs in cloud infrastructure behind OAuth 2.0 and covers runtime data and automation: sObject API, invocable actions, and flows. Salesforce claims 60+ new MCP tools and 30+ preconfigured coding skills2 under this umbrella, spanning platform data, metadata, workflows, business logic, permissions, and approvals.
Day-one external agent support lists four targets: Claude Code, Cursor, Codex, and Windsurf, connecting via Salesforce Hosted MCP Servers3. The selection is instructive. These are coding agents living inside developer environments, not orchestration frameworks. Salesforce is positioning as the orchestration plane that tools route to, not the IDE those developers work in.
The MCP Surface Area
Salesforce’s 60+ tool count is a platform-wide aggregate. No secondary source reviewed for this article breaks it out by Sales Cloud, Service Cloud, or Industry Clouds; that granularity isn’t in anything Salesforce published at TDX. Teams evaluating coverage for a specific vertical (Financial Services Cloud, Health Cloud) will need to audit the actual tool registry rather than treating the headline count as a per-cloud guarantee.
The DX/Hosted split does clarify the architectural intent. Metadata operations staying local makes sense given schema drift and org-specific configuration. Live sObject access going through OAuth 2.0 on a hosted server is the expected architecture for any multi-tenant trust boundary. What the split doesn’t tell you is how the 60+ tools distribute across that boundary: how many cover sObject operations versus flows versus Apex-adjacent developer tooling.
Salesforce also claims up to 40% reduction in CI/CD cycle times via the DevOps Center MCP. Both SalesforceBen and ApexHours flag this as a vendor-reported figure with no disclosed benchmark methodology.
Vibes 2.0 and the Multi-Model Question
Agentforce Vibes 2.0 is a browser-based, cloud-hosted VS Code fork. Per SalesforceTrail’s TDX coverage, Claude Sonnet 4.5 is the default coding model, with GPT-5 and Salesforce’s own models listed as switchable alternatives.
What none of the public coverage clarifies: whether Claude Sonnet 4.5 and GPT-5 in Vibes are hitting Anthropic and OpenAI APIs directly or running through Salesforce’s Einstein Trust Layer as proxied variants. That distinction carries compliance weight: data residency, token logging, model provider agreement scope. Treat multi-model support as a confirmed product feature and the hosting architecture as an open question until Salesforce publishes a Trust Layer policy update that explicitly covers Vibes 2.0 traffic.
One conflation to avoid: Google Gemini is GA as a rendering surface in the Agentforce Experience Layer, meaning Agentforce agents can surface through Gemini as an interface alongside Slack, Microsoft Teams, WhatsApp, and ChatGPT. Gemini is not a coding model in Vibes 2.0. The Vibes coding lineup is Claude Sonnet 4.5, GPT-5, and Salesforce-native models.
The free Developer Edition tier ships with 110 requests per month and 1.5M tokens per month4, with monthly resets only through May 31, 2026. After that date, the allocation converts to a single one-time grant with no further refreshes. [Vernon Keenan at SalesforceDevops.net described the free tier as “a sprint of experimentation, not a permanent free tier. The clock is ticking by design”5. That’s accurate as a structural description, whatever the intent.]
AgentExchange and the $50M Builders Initiative
AgentExchange consolidates what were previously separate marketplaces. Salesforce reports 10,000+ AppExchange apps, 2,600+ Slack Marketplace apps, and 1,000+ Agentforce agents, tools, and MCP servers, arriving at a 13,000+ total listings figure at launch. The AppExchange count varies by context: in the Slackbot integration context, ApexHours’ TDX recap cites 6,000+ AppExchange apps; in the AgentExchange marketplace context, the count rises to 10,000+. Both figures come from Salesforce; the discrepancy likely reflects different counting scopes.
Named launch partners include Google, Docusign, and Notion. Customer metrics surfaced at the keynote: Notion reports a sales cycle reduced from four months to three weeks; Docusign reports 200+ Q4 2025 offers processed and 60% faster time to signature; Engine reports agents in production in 12 days. These are customer-reported outcomes presented at a vendor conference.
The $50M Builders Initiative (also referenced in some Salesforce materials as the Builders Fund) targets pro-code developers who, in Cloudgaia’s framing, “are never going to click through a Setup menu”. The explicit goal is pulling Cursor, Windsurf, and Claude Code users into the Salesforce ecosystem, not converting them into Trailblazer-style declarative builders.
What This Means for LangGraph, CrewAI, and AutoGen
The historical value-add of Salesforce wrappers inside orchestration frameworks (clean SOQL adapters, schema introspection, retry logic around API limits) was always a compensating mechanism for the fact that Salesforce’s own tooling wasn’t agent-accessible. Sixty-plus first-party MCP tools covering sObjects, flows, metadata, and Apex testing compresses that gap.
The framework role doesn’t disappear; it shifts. The question before TDX was which orchestrator wraps SOQL best. The question after TDX is which orchestrator routes effectively between Salesforce-hosted MCP and external tools: a vector database, a non-Salesforce SaaS API, a model provider sitting outside the Einstein Trust Layer. LangGraph, CrewAI, and AutoGen teams building on Salesforce data now need to evaluate whether their current architecture adds routing intelligence or merely adds indirection.
The four day-one external agents (Claude Code, Cursor, Codex, Windsurf) are coding agents, not orchestration frameworks. LangGraph, CrewAI, and AutoGen are not listed as native Hosted MCP consumers in any TDX material reviewed here. Whether that’s a roadmap gap or a deliberate positioning decision isn’t clear from public sources. Salesforce may be capturing the coding agent surface and leaving orchestration framework integration to standard MCP protocol.
The Builder Gap
Keenan’s TDX notebook frames the conference’s central tension as the Builder Gap: Headless 360 and Vibes 2.0 target a “Claude-pilled” pro-code developer audience, while the traditional Trailblazer base, declarative admins who build with clicks, flows, and configuration, received no visible bridge program at TDX 2026.
The gap is structural. Salesforce’s agentic surface area requires Apex literacy, MCP tooling familiarity, and OAuth 2.0 configuration comfort. The free tier is time-boxed in a way that compresses evaluation cycles for a developer who can ship a proof of concept in two weeks, but doesn’t serve an admin who needs a longer ramp. If the admin base can’t participate in the agentic architecture, adoption will cluster around SI partners and pro-code shops, and the long tail of smaller Salesforce implementations will lag significantly.
Keenan also reports Salesforce’s explicit concession that fully autonomous multi-agent orchestration “isn’t enterprise-ready.” Agent Script is the current answer: deterministic guardrails over probabilistic systems, with human-in-the-loop fallback when confidence thresholds aren’t met. Keenan terms this “guided determinism.” It’s an architecturally honest position, even if it sits in tension with a keynote themed around the “Agentic Enterprise.”
Open Questions Practitioners Should Track
Multi-model hosting in Vibes 2.0. Whether Claude Sonnet 4.5 and GPT-5 traffic is direct provider API or Trust Layer-mediated has data residency and cost implications. Salesforce has not published a definitive architecture document covering this as of April 28, 2026.
Free-tier sustainability after May 31, 2026. The monthly refresh cutoff means developer adoption curves compress into a six-week window. Whether Salesforce extends the free tier, converts it to a paid SKU, or lets the one-time allocation stand without a path forward matters for any team building evaluation infrastructure around the Developer Edition.
Sales Cloud vs Service Cloud vs Industry Clouds tool inventory. Salesforce’s 60+ tool count is a platform aggregate. Per-cloud MCP tool availability is not documented in any secondary source reviewed here. Teams planning integrations in vertical clouds need this breakdown before committing to architecture changes.
Framework MCP consumption. Whether LangGraph, CrewAI, and AutoGen can natively consume Salesforce Hosted MCP Servers without a custom adapter layer is not confirmed in TDX materials. S2 Labs’ TDX recap lists four day-one targets (Claude Code, Cursor, Codex, Windsurf), and orchestration frameworks are absent from that list. Until a framework ships explicit Salesforce Hosted MCP support, the wrapper-collapse argument applies to the data layer, not the framework-to-platform bridge.
Frequently Asked Questions
Which Agentforce Experience Layer surfaces are GA versus still in beta?
Slack, Microsoft Teams, WhatsApp, ChatGPT, the Anthropic Claude app, Google Gemini, and iOS/Android are all GA. Web/React rendering remains in open beta, so teams building browser-embedded agent interfaces should expect breaking changes before that surface graduates.
Is Slack becoming the primary deployment surface for Agentforce agents?
Salesforce reports 300% growth in custom AI agents on Slack since January 2026, and Slackbot now orchestrates across 2,600+ Slack apps and 6,000+ AppExchange apps. Combined with AXL Slack being GA while Web/React sits in open beta, Slack is the lowest-friction channel for near-term Agentforce deployments.
What is a realistic time-to-market benchmark for shipping on Agentforce?
Among TDX customer references, Engine reached production agents in 12 days and MeshMesh landed its first Fortune 500 client six weeks after launch. Both worked directly with Salesforce as early partners, but they set a rough ceiling: teams that can’t match a 2-6 week sprint cadence will see the free tier’s May 31 monthly-refresh cutoff expire before evaluation completes.
Does the DevOps Center 40% CI/CD claim apply to mixed-stack pipelines?
The figure covers Salesforce-specific deployment cycles — metadata pushes, Apex test execution, and DevOps Center change management. It was not measured against mixed-stack pipelines that include non-Salesforce services, and no benchmark methodology was disclosed. Teams running polyglot CI/CD should treat it as a Salesforce-scope metric, not a pipeline-wide expectation.