MCPJam Review (2026): Pricing, Features & Honest Verdict

MCPJam is an open-source testing and evals platform for MCP servers that shows how ChatGPT, Claude and Cursor would use your tools before your users do. Verdict: every feature is on the free plan, but the credit system takes a calculator to follow. Best for: teams shipping MCP servers. Price: free, then $29 a month. Rating: 8/10

What is MCPJam?

We found MCPJam on the Product Hunt board for 17 September 2026, ranked seventh for the day under the tagline “The testing & evaluations platform for MCP servers”. If you have built a Model Context Protocol server, you already know the problem it goes after. Your server can pass every protocol check and still fail in the wild, because ChatGPT, Claude and Cursor each connect, authorize and call tools a little differently. MCPJam runs your server through those clients in a test bench, scores what happens and can block a pull request when the score drops.

The company is MCPJam Inc., and the product has been public for a while. Its GitHub repository was created in May 2025 and had 2,222 stars, 288 forks and a release, version 3.8.1, dated 18 September 2026 when we looked. The npm package recorded 21,603 downloads in the 30 days to 18 September. You can start it locally with one npx command, open a web app, or install a desktop build for macOS or Windows. Nothing in this review was tested hands-on. It is built from the site, the docs, the pricing data, the licence files and the public registries.

What Are MCPJam’s Key Features?

An inspector that shows several clients side by side

The Inspector is the free, local heart of the product. You connect a server, send a prompt, and watch it land in ChatGPT, Claude and Cursor columns at once, each with Chat, Trace and Raw views so you can see which tool was picked and what arguments it got. The homepage promises testing “in 16+ major AI clients”. Playground prompts cost no platform credits, only model usage, and you can plug in your own LLM keys. For day-to-day debugging this replaces a lot of manual copying between apps.

An OAuth debugger that stops at the broken step

Auth is where many remote MCP servers fall over, and MCPJam gives it a dedicated tool. The OAuth debugger walks the handshake step by step, runs spec conformance checks and, in the site’s words, lets you “See the exact step where auth breaks”. It also traces the newer cross-app access flows that enterprise identity setups use. It costs zero credits on every plan. If your server works with an API key but dies the moment a real client tries to authorize, this is the feature to open first.

Evals and a CI gate on every pull request

Evals turn one-off checks into a regression suite. You write cases, MCPJam runs them against real models several times, and it scores tool choice, arguments and goal completion instead of matching exact strings. That matters because model output is never identical twice. The same suite runs from the CLI, the SDK or a GitHub Action, and the pitch is simple: “Run the same suite on every pull request. A failing check blocks the merge.” Each run records pass rate, latency, tokens and tool calls, so you can see a slow drift as well as a hard failure.

Swarms and User Testing for whole journeys

Swarms are synthetic users. You describe who uses your server, MCPJam drafts personas and goals, and then “Agent personas run them across every client. Every session is captured and scored.” User Testing is the human version: you share a sandboxed chat that mirrors an AI host, invite testers, and collect per-turn ratings and replayable sessions. Both are included on the Free plan, which surprised us. Swarm prompts are the most expensive thing in the product at 5 platform credits each, so the cost shows up in usage, not in a locked feature.

How Much Does MCPJam Cost?

MCPJam starts at $0 with no card. Free includes every feature, unlimited seats, 30 days of trace history, 200 credits a day and “a one-time grant of 500 free eval iterations that can be used indefinitely”. Pro is $29 a month, or $24 a month billed annually at $288, with 5,000 credits a month, unlimited trace history, rollover up to 10,000 credits and email support. Team is $249 a month, or $199 billed annually at $2,388, with 50,000 credits, basic role-based access and priority support. Enterprise is custom and adds SSO, audit logs, advanced roles, an uptime SLA and a signed DPA.

The monthly figures come from the pricing page’s own data, because the page opens on the annual view. The toggle says “Up to 20% off”. We checked: Team saves 20.1% and Pro saves 17.2%, so the claim holds for one plan and is close on the other. No plan charges per seat, and that is rare in developer tooling. Braintrust, a general eval platform, is $0 or $249 a month with unlimited users. LangSmith is free for one seat and then $39 per seat a month. Postman is $19 per user a month on its Team plan, billed annually.

The terms of use contain something we rarely see. You have thirty days from the start of a subscription, or from any renewal, to cancel and get that fee refunded. After that window, fees are non-refundable. For a tool you might buy to get through one launch, a 30-day refund on a $29 plan removes most of the risk of trying it.

How Far Do the Credits Actually Go?

This is the part that needs care. A credit is the billing unit, and the FAQ publishes the platform rates: Playground and the OAuth debugger are free, Evals and User Testing cost 3 credits a prompt, and Swarm costs 5. If you use MCPJam’s hosted models, model usage is billed in credits on top. The pricing calculator estimates that at 1 extra credit per prompt step, and it prices on-demand insight requests at 20 credits plus 1.

Now run a real suite through that. Say 10 cases, 3 iterations each, across 3 client and model pairs, with 2 prompt steps per case. That is 180 prompt steps, or about 720 credits a run on hosted models. Pro’s 5,000 credits cover roughly 6 full runs a month. Team’s 50,000 cover about 69. So “gate every PR” on an active repository is a Team plan job, not a Pro one, unless your suite is small or you only run it on release branches. That example is our arithmetic, not the vendor’s, so put your own numbers into the calculator before you pick.

Two details are worth knowing. First, the Free plan’s 200 credits a day add up to 6,000 over a 30-day month, which is more than Pro’s 5,000. Free cannot roll credits over or buy top-ups, and its hosted models lock at the daily cap, but a patient solo developer gets a lot for nothing. Second, top-ups cost more than included credits. Pro works out to 0.58 cents a credit monthly, and extra credits are $10 per 1,000, or 1 cent each. Team is about 0.5 cents included and 0.9 cents extra. Top-up credits never expire, which softens it.

Bringing your own key is the escape hatch, with one open question. The FAQ says that when credits run out “Free hosted models lock” and anything run on your own key keeps working. The calculator says “With your own key, you pay your model provider directly. MCPJam platform fees still apply.” We could not tell from the public pages whether an eval run on your own key still stops when the 3-credit platform fee has nothing to draw on. Ask before you wire it into CI. Owners can set a budget ceiling with a warning at 80%, which is a welcome guard.

How Open is the Open Source, and What Should Enterprises Check?

The homepage FAQ says “The Inspector, CLI, SDK, local evals, and conformance checks are open source and free.” That is mostly true, and the npm package is marked Apache-2.0. The repository’s licence file is more precise. Everything is Apache 2.0 except the server’s services directory and its evals route, which fall under “The MCPJam Enterprise Edition (EE) license”. That licence says the code “may only be used in production” with a valid Enterprise Edition subscription. GitHub’s API cannot classify the mixed file and reports the licence as NOASSERTION. It is a common open-core split, and you can read and patch the code. But if you planned to self-host evals in production for free, read that file first.

Understand what “every client” means as well. MCPJam does not drive the real ChatGPT or Claude apps. The FAQ says “MCPJam maintains each client’s current behavior so your tests reflect what users experience today”, and the feature page says it tests the chain “the way each host will”. So a pass is a strong signal from a maintained model of each host, not a guarantee from the host itself. The vendor publishes a monthly changelog of client updates, which suggests it takes the upkeep seriously. We would still do a final manual check in the real client before a big launch.

For buyers with a security review, three facts. The footer says “SOC 2 in progress”, so there is no report yet. SSO, audit logs, the uptime SLA and the DPA are all Enterprise only, while basic roles start at Team. And the site states that “MCPJam is pre-production and never sees live traffic”, which limits the exposure: you are sending test prompts and your server’s test responses, not customer data, as long as you keep it that way. The homepage claim of “106,000+ developers and 280+ enterprises” is not something we could check. The public numbers we could check are the ones above.

Who is MCPJam Best For?

Use MCPJam if you are building an MCP server or a ChatGPT app and you are tired of finding out from users that a tool call broke in one client. The free plan is enough for a solo developer to debug OAuth, compare clients and keep a small eval suite. Use Team if your server is a product surface with paying customers on it and you want a merge gate that a whole engineering group shares, because unlimited seats on a flat $199 to $249 is cheap next to per-seat tools.

Skip MCPJam if you are evaluating an agent you built yourself and not a server that other clients call. The vendor says so plainly of Datadog, Braintrust and LangSmith: “Those products evaluate an agent you control. MCPJam tests how external AI clients use your MCP server.” Skip the paid plans for now if your company requires a finished SOC 2 report or SSO below an enterprise contract. And if all you need is to poke a server and read raw protocol messages, the official MCP Inspector does that for free.

Best MCPJam Alternatives

MCP Inspector is the official debugging tool from the Model Context Protocol project. It is free and open source, with 10,913 GitHub stars when we checked, and it covers connecting to a server, listing tools and calling them by hand. It has no model in the loop, no evals and no cross-client view. Start here if you are learning the protocol. Move to MCPJam when the question becomes how a model uses your tools.

Postman is where many API teams already live, and it can send MCP requests. Pricing is Free, Solo at $9 a month and Team at $19 per user a month, both billed annually. It is a general API client, so you get collections, monitors and documentation, and you give up OAuth flow tracing for MCP and scored evals. For a team of ten, Postman Team costs $190 a month against MCPJam Pro’s flat $29.

Braintrust is the pick if your real need is evaluating your own agent or prompts. Starter is $0 with 1 GB of processed data and 10,000 scores, and Pro is $249 a month with 5 GB and 50,000 scores, both with unlimited users. It is strong on datasets, experiments and production traces. It does not model how third-party AI clients will treat your MCP server, which is the one thing MCPJam is built for.

Final Verdict: Is MCPJam Worth It?

MCPJam picked a narrow problem and built the right tools for it. The OAuth debugger and the side-by-side inspector cost nothing and would justify the install on their own. Evals, Swarms and User Testing are all on the free plan too, seats are unlimited everywhere, the prices behind the annual toggle are there to be found, and a 30-day refund is written into the terms. The repository is active, with a release two days before we looked. That is a lot of goodwill for a tool in a young category.

What keeps it at 8 is the accounting and the fine print. Credits carry platform fees, model fees, daily caps, monthly caps, rollover limits and top-up rates, and whether platform fees bite on your own key is not answered in public. The “open source” line skips over an enterprise licence on the evals code. Client coverage is a maintained model, not the real apps. SOC 2 is still in progress. None of that should stop a developer from running the npx command today. It should make an engineering manager spend ten minutes with the calculator, and one email to sales, before promising the team a gate on every pull request.