AI Stack Migration + Validation: A New Revenue Stream for AI Agencies

Published August 29, 2026By ABD Legacy LLC
ai agent migration service ai migration consulting pricing post-import validation ChatGPT Work import Codex import

What is an AI agent migration service?

An AI agent migration service moves a client's existing agent setup — instructions, skills, plugins, project folders, memories, MCP configs, hooks, slash commands, and subagents — from Claude Code, Claude Cowork, or Cursor into ChatGPT Work or Codex, then validates that the imported configuration actually works in the new runtime. The import itself is a free OpenAI feature [1]; the service is the migration project around it: mapping what the client has, executing the import, reviewing what OpenAI's own docs say to review, fixing what breaks, and documenting the result. Since the import path is one-way into OpenAI — no reverse export is documented [1][2] — that documentation and the pre-migration due diligence are part of the deliverable.

On August 11, 2026, OpenAI shipped the desktop-and-CLI wave of "import from another agent," and the AI agency world got a new line item it didn't have the week before. Julian Goldie framed the feature as the end of a familiar pain: "OpenAI just attacked one of the most annoying problems with AI tools: Starting over. ChatGPT Work + Codex can now import working setups from tools like Claude Code and Cursor. Not just chats." [3]

The coverage that followed treated import as a self-serve convenience — a button that saves a rebuild. That framing misses the business half of the story: import is not fire-and-forget, and the work that comes after it is a priced, repeatable agency service. This page is the pricing playbook for that service: what the work actually is, what to charge, and where it slots into the models you already sell. For the switching-cost economics behind this same feature, see the sister guide on AI agent switching costs.

What the import feature actually moved (and what it didn't)

The August 11 wave made the ChatGPT desktop app able to import from Claude Code, Claude Cowork, or Cursor, and the Codex CLI able to import from Claude Code or Cursor via /import [1][2]. The importable items are spelled out in OpenAI's docs:

Two details matter for the sales conversation. First, import is non-destructive: "Importing doesn't change or delete your existing agent setup" [1], and the desktop flow has an optional automatic-updates setting that keeps imported work in sync with the original agent [1][2]. Second, credentials are not on the list. No API keys, tokens, or provider credentials are documented as importable [1] — a client that assumes "my whole setup moves" is wrong about the most sensitive part of the stack.

Why "import" is not fire-and-forget: the validation checklist

Here is the sentence that makes this a service: "A successful import means the configuration was mapped; it does not mean every permission, credential, hook, or integration has been validated for the new runtime." That is not our editorializing — it is the documented reality of the feature [1]. OpenAI's own import docs list exactly what to review after importing:

Every one of those bullets is a task a human has to do, check, and re-test. None of them happen inside the import button. That is billable validation work — and it is the exact scope agencies should package as a service rather than absorb as a favor.

What an AI stack migration + validation engagement includes

A clean engagement is five phases. The table below is a service menu with illustrative hour ranges for a small-to-mid stack (a solo founder or a small team's worth of agents); treat the hours as a sizing assumption to adjust per client, not a published standard.

Phase What happens Illustrative hours
1. Discovery & mapping Inventory the stack: agents, skills, plugins, MCP servers, hooks, slash commands, memories, recent chats. Identify credentials that will not transfer and flag them for re-setup. 2–6
2. Import execution Run Settings > Import (desktop) or /import (CLI). Map AGENTS.md and settings.json → config.toml, preserve project folders. Non-destructive: source setup stays intact. 1–3
3. Validation pass Work the documented review list: permissions and tool restrictions, MCP custom auth, hooks, plugin re-authorization, placeholder-based prompt templates. Re-test each imported item in the new runtime. 3–12
4. Documentation & handoff Write the mapping document, the one-way caveat, and the owner's manual; note what to re-check if automatic sync updates an imported item. 2–6
5. Post-migration support Short support window (typically 30 days) for breakage discovered after handoff. 4–8

The validation pass is the phase clients don't expect and agencies shouldn't skip — it is also the phase with the most leverage, because it is where a broken MCP auth or a silently-failing hook gets caught before it burns a client week.

The one-way direction is a due-diligence deliverable

The import path runs into ChatGPT Work and Codex. OpenAI documents no reverse export route [1][2] — we are stating the absence, not promising one. For your client, portability is asymmetric: leaving Claude Code, Claude Cowork, or Cursor for OpenAI is comparatively cheap, while the reverse move still means manual reconstruction.

That asymmetry is not a footnote to warn about — it is a deliverable. A client standardizing on OpenAI deserves a document that says, plainly: here is what moved, here is what maps back to the old tooling, and here is what a reverse migration would cost. That due-diligence artifact is planning value you can charge for, and it protects you if the client later asks why the door only swings one way. Do not promise clients a lock-in-free future — the honest framing is asymmetric portability, and the client pays for the honest version.

What to charge: pricing benchmarks for migration + validation

There is no published market benchmark for AI agent migration pricing yet — this is a new line, and most coverage still treats import as a self-serve convenience, so agencies are setting the price rather than matching one. That is the opportunity, and it comes with a discipline: don't invent a market that doesn't exist. Price it like the configuration work it is, and label your ranges as assumptions.

The cleanest model is hours × agency rate. The site's cost of hiring an AI agency guide already documents the rate context: AI engineers in North America typically bill $150–$350/hour, with senior engineers at the top of that band. A small-to-mid migration at 12–35 illustrative hours lands at roughly:

Those figures are arithmetic on illustrative hours, not published prices — quote the range you can defend for the specific stack, and say so. For scope clarity, most agencies price migration as a fixed-fee project (the engagement above is well-bounded) and keep post-migration support as either a short included window or a small retainer add-on. If the migration is the door into a larger build — the client importing their old setup because they are standardizing on ChatGPT Work — the migration fee can be bundled into the setup fee rather than billed separately; the calculator's setup-fee model starts custom-agent builds at $5,000–$25,000 and handles that bundling.

Where migration pricing fits in your pricing models

Migration + validation is a natural project-fee line, but it interacts with the models you already sell in three ways:

  1. As a lead generator: a $2K–$12K migration engagement is a low-risk entry that hands you the client's stack inventory — the best possible brief for the larger automation work that follows.
  2. As a retainer entry point: post-migration support (monitoring hooks, re-validating after auto-sync, credential refreshes) is recurring work. See monthly retainer vs. project pricing for how to frame the recurring leg without duplicating the one-time scope.
  3. As a pricing-model case study: the same engagement can be quoted fixed-fee, time-and-materials, or folded into an outcome-based retainer — the pricing models explainer walks through when each wins.

One warning that belongs in every quote: because import is one-way, the migration is a decision point, not a trial. If the client is unsure, sell a validation-only pilot (map the stack + run the review checklist against a test import) before the full migration — it is cheaper, reversible, and it is the same discovery work the full engagement needs anyway. The broader budgeting frame lives in the AI automation pricing guide.

Price it with the calculator

The calculator's Client ROI Justification output now includes a migration + validation line: when the client has an existing Claude Code, Claude Cowork, or Cursor stack, the quote shows the reduced migration effort as a client benefit and flags the one-way direction as a planning note, so the migration hours you quote are visible against the savings the client gets. Run the engagement through the calculator and the migration line shows up in the ROI story you hand the client.

Price a migration + validation engagement with the calculator

Open the Calculator →

Related reading: cost of hiring an AI agency and monthly retainer vs. project pricing.

How to sell it

Do not sell "import" — the client can click that button themselves. Sell what the button doesn't do:

The market gap is real and it is current: as of this writing, the search results for agent migration are vendor documentation and tool listings, and nobody positions post-import validation as a priced agency service [3]. That is a pricing vacuum, and vacuums in service categories don't stay empty long.

Frequently asked questions

What is an AI agent migration service?

An AI agent migration service moves a client's existing agent setup — instructions, skills, plugins, project folders, memories, MCP configs, hooks, slash commands, and subagents — from Claude Code, Claude Cowork, or Cursor into ChatGPT Work or Codex, then validates that the imported configuration actually works in the new runtime. The import itself is a free OpenAI feature; the service is the migration project around it: mapping what the client has, executing the import, reviewing what OpenAI's own docs say to review, fixing what breaks, and documenting the result.

Does the import transfer API keys and provider credentials?

No. Provider credentials and API keys are not on the documented import list — do not assume they transfer. MCP servers using custom auth, headers, env vars, or custom transports may require sign-in again, and imported plugins or marketplace connections may need re-authorization. Credential re-setup is part of the validation pass, not a side effect of the import.

Is the ChatGPT Work / Codex import one-way?

The import path runs into ChatGPT Work and Codex; no reverse export route is documented by OpenAI. That means a client can move from Claude Code, Claude Cowork, or Cursor into OpenAI comparatively easily, but the reverse move still means manual reconstruction. Portability is asymmetric, not a guarantee — which is exactly why documenting the mapping before migrating is part of the service.

What do you review after importing an agent setup?

OpenAI's import documentation explicitly lists what to review: tool restrictions and permissions in imported skills and agents; MCP servers using custom auth, headers, env vars, or transports (they may require sign-in again); hooks, which may behave differently after import; plugins and marketplaces, which may need manual follow-up; and prompt templates that depend on arguments, shell interpolation, or file-path placeholders, which may break.

What should an AI agent migration service cost?

There is no published market benchmark yet — this is a new service line, so price it as configuration work: hours × your agency rate. For a small-to-mid stack, a 12–35 hour engagement (discovery and mapping, import execution, validation, documentation, and a short support window) at typical agency engineering rates of $150–$350/hour comes out to roughly $1,800–$12,000 as an illustrative planning range. Treat the range as an assumption to size per client, not a published price.

Does importing change or delete the client's existing setup?

No. OpenAI's docs state that importing doesn't change or delete the existing agent setup. The source stack stays in place, and the optional automatic-updates setting can keep imported work in sync with the original agent while both exist. That makes migration low-risk for the client — and it is part of the sales story, not the whole story, because the post-import validation work is still yours to do.

Related reading

Sources

Accuracy note: all import mechanics, the importable-item table, the non-destructive guarantee, the what-to-review list, and the absence of credentials on the import list are taken from OpenAI's official import documentation (learn.chatgpt.com/docs/import) and changelog, verified 2026-08-16 and re-verified 2026-08-29. The one-way direction is stated as an absence of a documented reverse export route, not as a promise. The 12–35 hour engagement and $1,800–$12,000 range are illustrative planning assumptions for a small-to-mid stack, not published benchmarks; the $150–$350/hour rate band is this site's published agency-engineering context (see the cost-of-hiring guide) and varies by market and seniority. OpenAI's docs state no plan requirement for the import feature; plan gating beyond that is unstated — verify current terms before quoting.