6 Best Ramp Glass Alternatives: AI That Writes Specs in 2026

Ramp's Glass agent drew a lot of attention at Lenny and Friends Summit in September 2026. It is an AI agent that turns real customer and product data into specs that coding agents can build from. The catch is that Glass is internal to Ramp, so nobody outside Ramp can use it. Product teams who want the same result need to assemble it themselves. This guide ranks the six best Ramp Glass alternatives, starting with BuildBetter for the evidence layer. Each tool is scored against a rubric drawn from what Glass connects to and what it produces, and the guide ends with a practical stack for writing evidence-based PRDs, agent-ready requirements and prototypes in 2026.

What Is Ramp Glass, and Why Can't You Just Buy It?

Ramp Glass is an internal AI agent that helps Ramp product managers decide what to build. Instead of starting from a blank prompt, it connects AI to the company's real data, strategy and codebase. Geoff Charles, Chief Product Officer at Ramp, presented it in his talk “The limiting factor — how to design an AI software factory for speed” at Lenny and Friends Summit in September 2026.

The thesis of the talk fits in one sentence: AI does not remove the bottleneck, it moves it. Charles argued that PMs should invest in “the factory” as well as the product. He closed with a direct invitation: “I want you to copy us.”

Glass is step 2 of Ramp's five-step AI software factory: Identify, Define, Build, Coordinate, Improve. It owns Define. For the full factory, see our breakdown of Ramp's AI product factory and internal tools.

Glass exists because “What do you want to build?” is the wrong starting prompt. A generic AI chat box knows nothing about your customers, metrics or product, so it produces specs that read well and have nothing behind them. Glass replaces the blank prompt with connected context. According to the talk, it connects to:

• Snowflake for quantitative data
• User research for qualitative data
• Product strategy
• The spec format
• The codebase
• The design system
• Product principles

Glass also works as a “tech lead.” A PM can ask it “is this possible?” or “would this break something?”, and Glass can build working prototypes inside the real Ramp product.

Glass is an internal Ramp tool. It is not for sale. Ramp has not released it publicly. Anyone who wants Glass has to assemble or build an equivalent.

The “Next Contract”: The Three-Part Handoff Glass Produces

The handoff engineers want from PMs, which Charles called “the next contract,” has three parts.

1. Evidence: qualitative and quantitative proof that the problem is real. That means customer quotes plus usage or revenue data.
2. Requirements usable by coding agents: structured, unambiguous specs that an AI coding agent can execute against.
3. A prototype for inspiration: a working example that shows intent. It is not the final implementation.

A long spec alone is not the contract. A prototype alone is not the contract. The contract is evidence, agent-ready requirements and a prototype together.

The timing explains why this matters. Per the talk, Ramp's background coding agent Inspect builds 75% of Ramp's PRs. Once agents write most of the code, the input spec becomes the constraint. Coding agents execute literally, so a vague requirement turns into wrong code quickly.

Outside research supports the “bottleneck moves” argument. A July 2025 METR randomized controlled trial found that experienced open-source developers took 19% longer to finish tasks with AI tools, even though they believed AI had sped them up by about 20%. The Stack Overflow Developer Survey 2025 found that 84% of developers use or plan to use AI tools. More of them distrust AI output accuracy (about 46%) than trust it (about 33%). Google Cloud's 2025 DORA report describes AI as an amplifier of whatever system it enters. Weak inputs produce weak outputs, only faster.

Treat the prototype as intent. Charles described it as something engineers “can actually get inspired by.” It shows engineers and agents the UX direction, and it should not become production code by default.

No single commercial tool produces all three parts today. The evaluation below scores each tool on which parts of the contract it covers.

How We Evaluated Ramp Glass Alternatives: A 6-Point Rubric

Each Ramp Glass alternative is scored on six criteria taken directly from what Glass connects to and what it produces.

1. Connected to quant data: Can it read from a warehouse like Snowflake, or from product analytics?
2. Connected to qual research: Can it read customer calls, support tickets, Slack, surveys and interview notes?
3. Knows product strategy and your spec format: Can it follow your PRD template, principles and strategy docs?
4. Knows the codebase: Can it judge feasibility and what might break?
5. Can prototype: Can it produce a working prototype, ideally inside your real product?
6. Produces the three-part handoff: Does it deliver evidence, agent-ready requirements and a prototype?

Methodology note: Tools are assessed on documented capability categories. Where a feature varies by plan or changes quickly, we say “check current capabilities” instead of making a specific claim. This category moves monthly, so verify integrations and pricing before you buy.

One principle from the research shapes how to read the scores: connect evidence before code. A codebase-aware agent without customer evidence can build the wrong thing, and build it well. That is why the ranking weights criteria 1, 2 and 6 heavily. An AI spec writer for product managers is only as useful as the evidence it can cite.

The 6 Best Ramp Glass Alternatives in 2026

Here is the ranked list, followed by a detailed look at each tool.

1. BuildBetter — best for the evidence half of the contract
2. ChatPRD — best for fast, template-driven PRD drafting
3. Zeda.io — best for feedback-to-roadmap product discovery
4. Notion AI — best if your strategy and specs already live in Notion
5. Claude with Projects and connectors — best configurable middle ground
6. Build your own on the Claude Agent SDK — closest to what Ramp did

Options 1–4 are products you can buy. Option 5 is a general-purpose assistant you configure. Option 6 is the path closest to what Ramp actually built. Each entry uses the same structure: what it does, where it fits the Glass pattern, pricing model, best for, and one honest limitation.

1. BuildBetter — Best for the Evidence Half of the Contract

What it does: BuildBetter is an AI product-intelligence platform for B2B product teams. It brings internal customer voice (call recordings, Slack) and external feedback (support tickets, surveys, product feedback) into one place through 100+ integrations, including Zoom, Slack, Jira, Salesforce, Zendesk, HubSpot and Intercom.

Where it fits Glass: BuildBetter covers Glass's qual-research connection and the evidence part of the handoff more completely than any other option here. BuildBetter Documents generates PRDs and requirements from real customer calls, tickets and Slack threads. Source quotes are attached, so every requirement traces back to a customer. Signals analyzes each piece of feedback on its own, with severity, sentiment and business impact, and applies your taxonomy instead of relying on keyword matching. Tickets creates Linear and Jira issues with the full context carried over.

Differentiators:

• It captures source data directly, through a bot recorder, no-bot local recording and mobile, instead of only sitting on top of data you already have.
• It produces finished artifacts such as PRDs, tickets and loop-closure emails to customers, where many tools stop at theme dashboards.
• It exposes evidence to other AI tools through the BuildBetter MCP Server, so Claude, Cursor or a custom agent can pull cited customer signals.

BuildBetter answers the first question of the contract, which is whether the problem is real and who said so, and it attaches the quotes.

Pricing model: Usage-based with unlimited seats. Check current pricing for specifics.

Security: SOC 2 Type II and HIPAA-ready. This matters when an agent reads customer calls.

Best for: B2B product teams whose customer evidence is scattered across calls, Slack and support tools, and who want requirements grounded in that evidence.

Honest limitation: BuildBetter is not a prototyping tool and does not read your codebase. Pair it with a coding agent for the prototype part of the contract, and with your BI or warehouse layer for deep quant analysis. Teams that mainly need dedicated research-repository workflows, such as manual tagging and highlight reels, or enterprise survey distribution at scale, may be better served by specialist tools.

2. ChatPRD — Best for Fast, Template-Driven PRD Drafting

What it does: ChatPRD is an AI assistant built for writing and critiquing product requirements documents. It is designed around PM workflows and document templates.

Where it fits Glass: It is strongest on criterion 3, spec format. It turns an idea into a well-structured PRD quickly, which supports the requirements part of the contract. If your team struggles with inconsistent PRDs, an AI PRD generator that enforces structure is a real improvement.

Integrations: It connects to common document and work-management tools. Check current integrations before relying on it for warehouse or codebase access. We make no claims about either.

Pricing model: Check current pricing. Self-serve tiers are typical in this category.

Best for: Individual PMs and small teams who need better PRDs faster and already keep their evidence somewhere else.

Honest limitation: It starts closer to the “what do you want to build?” prompt Charles warned against. Without a connected evidence source, the PRD is only as good as what you paste in.

3. Zeda.io — Best for Feedback-to-Roadmap Product Discovery

Status: Zeda.io has announced it is shutting down in 2026. Evaluate it only for short-term use.

What it does: Zeda.io is a product discovery platform. It aggregates customer feedback, uses AI to surface themes and insights, and connects them to roadmap and prioritization workflows.

Where it fits Glass: It partially covers qual research (criterion 2) and helps link evidence to what gets prioritized. That makes it an upstream input to the Define step, closer to Ramp's Identify stage than to Glass itself.

Capabilities: Feedback aggregation, AI insight summaries and roadmapping. Check current spec-generation and integration features, since these change often.

Pricing model: Check current pricing.

Best for: Teams that want feedback, discovery and roadmap in one product-management workspace.

Honest limitation: It does not read your codebase or build prototypes. Verify quant warehouse connectivity before you depend on it for the quantitative half of your evidence.

4. Notion AI — Best if Your Strategy and Specs Already Live in Notion

What it does: Notion AI is built into the Notion workspace. It writes, summarizes and searches across workspace content and connected apps.

Where it fits Glass: It is strong on criterion 3. If your product strategy, principles and PRD template already live in Notion, the AI works in your spec format by default. The talk noted that Ramp's own roadmap, specs and customer calls live partly in Notion, where they feed Ramp's Gadget agent.

Connected search: It can search across some connected tools. Check which connectors your plan includes.

Pricing model: Check current pricing, since AI features are tied to specific plans.

Best for: Notion-native teams who want spec drafting grounded in existing internal docs.

Honest limitation: It is a knowledge-workspace AI. It does not analyze customer evidence or act as a codebase-aware tech lead. Quant data, raw call analysis and prototyping all fall outside what it does well.

5. Claude with Projects and Connectors — Best Configurable Middle Ground

What it does: Anthropic's Claude offers Projects, which hold persistent context such as strategy docs, principles and spec templates. Connectors let it reach external tools and data sources.

Where it fits Glass: This is the most flexible buyable option against the full rubric. You load your strategy and spec format into a Project and connect your data sources. Depending on setup, Claude can reason about code and produce prototype-style artifacts.

Coverage depends on configuration. Claude does not come connected to your warehouse or codebase. It reaches only what you expose. Connecting an evidence source such as the BuildBetter MCP Server gives it cited customer signals without raw transcript dumps.

Context-limit caution: Per the talk, a 1M-token context window holds less than 0.5% of Ramp's recorded call transcripts. A rough estimate shows why. Speech runs about 130–150 words per minute, so an hour-long call is about 11,000–12,000 tokens. That means 1M tokens covers only about 80–90 hours of calls. Pasting raw data does not scale. You need retrieval or a pre-processed evidence layer.

Pricing model: Check current pricing for team and enterprise plans.

Best for: Technical PMs who want a Glass-like experience this quarter without starting an engineering project.

Honest limitation: It is a general assistant rather than a finished system. You have to design evidence traceability, consistency across PMs and governance yourself.

6. Build Your Own on the Claude Agent SDK — Closest to What Ramp Did

What it is: An internal PM agent built on Anthropic's Claude Agent SDK and wired to your warehouse, research, strategy docs and codebase. The SDK runs on the same harness that powers Claude Code. Anthropic renamed it from the Claude Code SDK in late September 2025 to reflect uses beyond coding. It offers Python and TypeScript libraries with tool use, context management, subagents, permissions and MCP support.

Attribution: Public write-ups, not the talk, describe Glass as built on Anthropic's Claude Agent SDK, with one Okta SSO login connecting 30+ tools including Salesforce, Snowflake, Slack and Figma.

Where it fits Glass: It can meet every rubric criterion because you define the connections. It is the only option that can fully replicate the three-part handoff inside your own product.

Reference material: Ramp's engineering blog explains how it built its background coding agent Inspect, using Modal sandboxes and the open-source OpenCode agent framework, and it published a spec for replicating it. Read why Ramp built its background agent.

Cost model: Engineering time plus model and infrastructure usage. There is no license fee, but maintenance is an ongoing cost.

Best for: Teams with platform-engineering capacity and a technical PM willing to “build the factory,” one of the three PM futures Charles described.

Honest limitation: You own everything: connectors, permissions, evaluation and upkeep. Charles himself said, “Everything you've seen here is outdated.” Expect to rebuild often.

Practical shortcut: Build the agent layer yourself and buy the evidence layer. That way you are not writing call-capture and feedback pipelines from scratch. Point your agent at an evidence platform over MCP and put your engineering hours into the codebase and prototype connections.

Ramp Glass Alternatives Compared Against the Rubric

No buyable tool covers all six criteria. The table shows where each one is strong and where you will need to add something.

Takeaway: No off-the-shelf tool matches Glass end to end. The practical answer is a stack.

How to Assemble a Glass-Like Stack

A Glass-like stack assigns each part of the three-part handoff to the tool best suited for it.

Assemble it in this order:

1. Connect the evidence sources first. Calls, tickets, Slack and surveys go in before anything else, followed by warehouse metrics.
2. Encode your spec format and product principles as reusable context. Put your PRD template, strategy and non-goals where every PM's agent reads the same version.
3. Connect the codebase and design system last. Feasibility checks only help once you know the problem is real.
4. Require every spec to cite evidence before it reaches a coding agent. No quote and no metric means no build.

An agent-ready PRD also needs explicit, testable acceptance criteria, clear scope and non-goals, edge cases, references to relevant components or files, test expectations and links to evidence. A coding agent will not ask follow-up questions, so the spec has to answer them in advance.

Glass draws on user research, and Ramp's Identify step exists to make customer evidence traceable. To build that upstream layer, read how to build a customer insight agent like Ramp and our roundup of Ramp customer insight agent alternatives.

The define step is only as good as the evidence feeding it. Fix the inputs before you automate the spec.

Which Ramp Glass Alternative Should You Choose?

Choose the tool that solves your current bottleneck in the three-part handoff.

• Choose BuildBetter if your biggest gap is evidence: customer calls, Slack threads and tickets that never make it into specs.
• Choose ChatPRD if you already have evidence and need faster, cleaner PRDs.
• Avoid new Zeda.io rollouts: the product is shutting down in 2026.
• Choose Notion AI if your strategy and specs already live in Notion.
• Choose Claude with Projects and connectors if you are a technical PM who wants a configurable Glass-like assistant now.
• Build on the Claude Agent SDK if you have engineering capacity and want the full three-part handoff inside your own product.

Charles made a point in the talk that applies here: constraints force you to pick a dimension. His example was Audi at Le Mans, which won on fuel efficiency and fewer pit stops instead of top speed. Apply the same logic to your spec workflow. If engineers keep asking “why are we building this?”, fix evidence first. If they keep asking “what exactly do you mean?”, fix requirements. If they keep misreading the UX, add prototypes. Solve one part, then move to the next.

FAQ: Ramp Glass Alternatives

Can I buy Ramp Glass?

No. Glass is an internal Ramp tool and is not sold or publicly available. You can replicate the pattern with a stack (evidence platform, spec tool, coding agent) or by building a custom agent on a framework like Anthropic's Claude Agent SDK.

What is Ramp Glass?

Ramp Glass is an internal AI agent for Ramp PMs. It connects to Snowflake, user research, product strategy, the spec format, the codebase, the design system and product principles. It helps PMs define what to build, check feasibility and prototype inside the real product.

What did Ramp build Glass on?

Public write-ups describe Glass as built on Anthropic's Claude Agent SDK, with one Okta SSO login connecting 30+ tools including Salesforce, Snowflake, Slack and Figma.

What should a PM's AI agent connect to first?

Connect customer evidence first: calls, tickets, Slack and research. Add quant data from the warehouse next, then strategy and spec format, and the codebase last. Without evidence, the agent is back to the blank “what do you want to build?” prompt.

What is the “next contract” between PMs and engineers?

Per Geoff Charles, it is a three-part handoff: qual and quant evidence that the problem is real, requirements usable by coding agents, and a prototype for inspiration.

What makes a PRD “agent-ready”?

An agent-ready PRD has explicit, testable acceptance criteria, clear scope and non-goals, edge cases, references to relevant components or files, test expectations and links to evidence. Coding agents execute literally, so any ambiguity turns into wrong code.

Can AI write a PRD from customer data?

Yes. Tools that connect to call recordings, support tickets and feedback can generate requirements with source quotes attached, so each requirement traces back to a real customer. BuildBetter does this across calls, Slack and 100+ integrations.

How does Glass fit with Ramp's other internal tools?

Glass handles the Define step. It sits between the customer insight agent (Identify) and Inspect, Review Buddy and Testo (Build). For the Improve loop, see how Ramp fixes UX issues in 24 hours.

Build the Evidence Layer Your Specs Are Missing

Every Glass-like stack depends on evidence. BuildBetter captures calls and pulls in tickets, Slack threads and surveys. It analyzes each signal for severity and business impact, then generates PRDs and tickets with the customer quotes attached, so your coding agents build what customers asked for. When the work ships, BuildBetter notifies the customers who requested it and closes the loop. Teams at Clay, Brex, WordPress, PostHog and 30,000+ others use it to ground product decisions in real customer evidence.

Make churn optional. Book a demo to see how BuildBetter turns customer conversations into agent-ready specs.