Vercel Eve Features vs Mastra (2026): Who Owns Each Capability

Quick verdict

The Eve vs Mastra pairwise is an authoring-and-gravity decision. This page is the feature ownership matrix behind that cutover.

Eve ships directory skills, file-discovered tools, a first-class sandbox, Markdown cron schedules, channel files, Connect auth, Agent Runs, and eval suites — all assuming Vercel Workflows / Sandbox / Gateway / Connect. Best for teams that want those capabilities as files, not as a product you assemble. Honest limit: memory and RAG are thinner in-box; you attach them as tools and connections.
Mastra ships typed agents, graph workflows with suspend/resume, database-backed memory, RAG primitives, scorers, MCP, Studio, and deployers for Vercel / Cloudflare / Netlify or a Hono server. Best for teams that want the agent runtime in code they own. Honest limit: sandbox, Slack/Teams channels, and Connect-style credential brokerage are your job (or a template), not a first-class directory.

Cipher's cutover rule: score who owns the hard feature, not who lists it on a marketing page. If sandbox + channels + schedules are the hard features, Eve. If memory + RAG + portable evals are the hard features, Mastra. Eve stacks with the Vercel AI SDK; it does not replace it — Eve vs AI SDK · Mastra vs AI SDK.

Last updated: August 2026. Eve is preview/beta. Re-check Eve concepts and Mastra.

Capability ownership matrix

Capability	Vercel Eve	Mastra	Who owns it
Skills / playbooks	Markdown in `skills/`, loaded on demand	Instructions + tools + optional skill packages; code-first	Eve if you want filesystem playbooks; Mastra if skills live in TS
Tools	One file in `tools/`; filename is the name; Zod schema	`createTool()` registered on the agent	Tie — both are typed TS tools
Sandbox	Every agent gets one; Vercel Sandbox in prod; Docker/bash locally	You choose the isolation layer	Eve
Schedules	Markdown cron in `schedules/`; durable without an open session	Recurring schedules exist in the harness; you wire the runner	Eve for cron-as-files; Mastra if the host already has a scheduler
Evals	Framework eval suites; run on deploy and on a schedule; Agent Runs	Built-in scorers + Studio-oriented eval loop	Mastra for in-process scorers; Eve for dashboard-native Agent Runs
Memory & RAG	Session context + skills; deeper RAG as tools/connections	Message history, working memory, semantic recall, RAG primitives	Mastra
Channels	First-class Slack, Discord, Teams, web, cron, GitHub, Linear	HTTP / app; Slack exists as a template, not a channel OS	Eve — see multi-channel agents
Auth / connections	Vercel Connect; short-lived credentials for GitHub, Slack, Linear	Your secrets, your OAuth, your rotation	Eve on Vercel; Mastra when you must own the secret store
Subagents	Declared under `subagents/` plus a built-in agent tool	Multiple agents + workflow graphs	Tie — different shapes
Durability	Vercel Workflows sessions	Native workflows + suspend/resume; host still yours	See durable workflows
Deploy	Happy path: `vercel deploy`	Vercel / Cloudflare / Netlify deployers or a Hono server	Mastra for portability; Eve for Vercel speed

Skills, tools, and sandbox

Eve treats skills as Markdown playbooks the model loads only when relevant. That keeps the always-on instructions.md short. Tools are files. The sandbox is not optional: bash, read, and write hit an isolated filesystem so model-generated code never shares the harness process.

Mastra treats tools as registered TypeScript functions and keeps more of the agent in code. Isolation is a decision you make. That is fine when the agent never runs untrusted shell. It is a gap the first time a tool writes a script and executes it on the app host.

If untrusted code exec is a requirement this quarter, Eve owns the feature. If it is not, do not pick Eve only for the sandbox SKU.

Schedules and evals

Eve schedules are Markdown with a cron front-matter. Daily reports and weekly digests keep running without an open chat session because Workflows hold the work. Evals are suites you run on deploy and on a schedule; Agent Runs in the Vercel dashboard shows sessions, tool calls, timing, and tokens without an extra vendor.

Mastra's eval story is stronger inside the TypeScript process: model-graded, rule-based, and statistical scorers, plus Studio for the loop. Schedules exist, but they are not "drop a cron file and Vercel Cron is the runner." You still pick where the clock lives.

Pick Eve when the schedule is the agent surface (ops digests, Slack stand-ups). Pick Mastra when the eval harness is the product discipline and you refuse a Vercel-only dashboard.

Memory, RAG, and what you still assemble

This is where Mastra wins on features, not gravity. Working memory, semantic recall, and RAG primitives ship in the framework. Eve gives you a durable session and skills. Deeper retrieval is a tool or a connection you write.

Cipher's first-hand read: teams that pick Eve and then spend a month bolting on pgvector have usually picked the wrong framework. Teams that pick Mastra and then spend a month bolting on Slack + a sandbox have also picked the wrong framework. Score the hard feature before you start.

FAQ

Is this the same article as Eve vs Mastra? No. That post decides by authoring model and cloud gravity. This post scores individual capabilities. Read both if the team is split between "we like Markdown agents" and "we need memory and evals in one package."

Does Mastra have Slack? Yes, as a template and as HTTP you own. Eve has a first-class Slack channel with Connect-managed credentials. Different ownership. Details: multi-channel agents.

Can I mix them? Rarely worth it. Pick one agent OS. You can still use the Vercel AI SDK at the UI edge with either.

Who helps score this for an AU/SG team? Cipher Projects — we map the hard feature to a framework before we write the first tool.

Conclusion

Eve owns sandbox, channels, Connect, and cron-as-files. Mastra owns memory, RAG, portable evals, and cloud-portable deploy. Score the feature that will actually slip the date. Then pick the owner.