NEW!!!

← NOT THE ACTUAL CARD

YOU COULD WIN THIS

If you have this MCP properly enabled in your AI, and it hallucinates, you could win this
$50.00 outback gift card.

NO PURCHASE NECESSARY. One (1) $50.00 Outback Steakhouse gift card to the first verified person whose AI, with this MCP properly enabled and actually used on that turn (transcript of MCP tool calls required; installed-but-not-called is a DQ), hallucinates a fabricated case, citation, or quotation offered as real; model must have been first released in calendar year 2026 or later (fine-tunes/quantizations/LoRAs/distillations of a pre-2026 base do not qualify), must exceed 24 billion parameters (MoE scored on advertised total, not “active”; lab won’t say = DQ), temperature at or below 0.2 consistent with reasonable professional standards for legal research and in no event above 0.4 (“creative”/“max”/temp-1.0 presets automatic DQ; if the product hides the dial the vendor’s default non-creative setting is deemed reasonable); typos, truncated names, “I can’t find that,” jailbreaks, and “ignore previous instructions” do not count; show the model name, the date, the connection, and the fake cite; text 904-383-7448; U.S. mail only; one winner; one card; sponsor’s determination is final including on whether your model is really 24B and whether your screenshot is a bit; server outages, MCP overall failure, denial of service, overuse, too many visitors, or too many requests is not a way to win; I am the final judge and arbiter but I will admit defeat; not affiliated with Outback Steakhouse or Bloomin’ Brands, Inc.; Bloomin’ Onion not included; if this banner is still up the card is still in the desk; void where prohibited and void if you used a 7B.

Full official rules.

days without
free steak

Benchmark Information

Gemma 4 26B on the same 150 legal research questions: with no tools 68.7% of its citations did not exist (178 of 259); with Syfert's MCP, Sep 22: 0.7% (2 of 289); Sep 23: 2.2% (5 of 230); With MCP + self-check · Sep 23: 0.0% (0 of 203); Oct 9 · 1.4.0: 0.0% (0 of 190); With MCP + self-check · Oct 9 · 1.4.0: 0.0% (0 of 209). Tool pass rates on 100 real briefs, Sep 28 to Oct 9: check_citation 88.4% to 98.9%; find_case 69.9% to 97.5%; get_statute 49.1% to 97.8%; get_treatment 97.6% to 99.9%; check_brief 96.5% to 99.9%; search_cases 80.2% to 90.3%. — Same open model, same 150 questions: on its own, Gemma 26B made up 68.7% of its citations. In the latest run shown (With MCP + self-check · Oct 9 · 1.4.0), it made up 0 of 209 (148 of 150 questions answered; 2 exceeded the model’s context window). And the steak above is still in the desk after 32 days.

Same open model, same 150 questions: on its own, Gemma 26B made up 68.7% of its citations. In the latest run shown (With MCP + self-check · Oct 9 · 1.4.0), it made up 0 of 209 (148 of 150 questions answered; 2 exceeded the model’s context window). And the steak above is still in the desk after 32 days.

Gemma 4 26B, fabricated citations, the same 150 questions in every run (citebench)
Run	Fabricated	Invented cites	All cites	Cited the expected case	Questions answered
No tools	68.7%	178	259	3.3%	150 of 150
With Syfert MCP · Sep 22	0.7%	2	289	72.0%	150 of 150
With Syfert MCP · Sep 23	2.2%	5	230	89.3%	150 of 150
With MCP + self-check · Sep 23	0.0%	0	203	88.7%	150 of 150
With Syfert MCP · Oct 9 · 1.4.0	0.0%	0	190	91.2%	147 of 150 (3 exceeded the model’s context window)
With MCP + self-check · Oct 9 · 1.4.0	0.0%	0	209	92.6%	148 of 150 (2 exceeded the model’s context window)

MCP tools, expectation pass rate on 100 real briefs (mcpbench, same grader for all three runs)
Tool	Sep 28 (server 1.1.0)	Oct 4 (server 1.2.0)	Oct 9 (server 1.3.0)
`check_citation`	88.4% (2,890/3,270)	98.8% (3,552/3,596)	98.9% (3,555/3,595)
`find_case`	69.9% (1,305/1,867)	97.4% (2,095/2,150)	97.5% (2,099/2,152)
`get_statute`	49.1% (264/538)	97.8% (535/547)	97.8% (535/547)
`get_treatment`	97.6% (2,200/2,254)	99.9% (2,276/2,278)	99.9% (2,278/2,280)
`check_brief`	96.5% (3,370/3,493)	99.9% (3,499/3,502)	99.9% (3,499/3,502)
`search_cases`	80.2% (239/298)	81.5% (243/298)	90.3% (269/298)

Syfert Legal Research

Is your law license worth a $50.00 Outback Steakhouse giftcard?

An AI can write you a handsome brief in ninety seconds. It can also invent a case that never existed, dress it in a confident citation, and let you file it. Lawyers have been sanctioned for exactly that. One URL connects your AI chatbot or TUI coding terminal to this site's research stack, and from then on every citation your robot writes has to survive contact with ten million real cases before it reaches your desk.

Search everything.

Full-text search of 10.7 million opinions, every state and every federal court, ranked by authority.

Catch the fakes.

Every citation gets checked against the corpus: does it exist, and is it the case your robot says it is.

Bluebook out of the box.

Citations come back formatted, court and year resolved, each case name linked to the full opinion on this site.

Know what a case stands for.

The propositions courts actually cite a case for, each flagged on its own. A case can be gold on one point and dead on another.

Audit the whole brief.

Every cite, every quoted passage, every claimed case name, one pass, before you file. Attack mode does the same thing to the other side's brief.

Statutes, everywhere.

All fifty states plus the U.S. Code, the C.F.R., the federal rules, and the sentencing guidelines.

What this is: Syfert Legal Research is a legal research service and MCP connector for lawyers. It lets an AI assistant such as Claude, ChatGPT, Cursor, Codex or the open-source MikeOSS search 10.7 million U.S. court opinions and every state's statutes, verify citations, check whether a case is still good law, and audit a brief, all against syfert.com's own library. Signing in with Google or Microsoft only identifies your account; see the privacy policy.

Create a free account ▸ See the full toolset ▸ Install instructions ▸ Go Pro ▸

Try it with no account.

Point any MCP client at https://mcp.syfert.com/mcp with no credential and it answers. 20 calls a day per address.

Free account.

10,000 searches a month. Unlimited citation checks, treatment, brief audits and statutes. No bot checks on the site. No card, and free stays free.

Pro, $30 a month or $300 a year.

The dedicated index this practice keeps warm on its own hardware, priority ahead of free traffic, no caps, and semantic search when it ships.

It works with any MCP-capable AI: a chatbot like Claude or ChatGPT on the web, desktop, or phone, TUI coding terminals like Claude Code, and open-source legal AI platforms like MikeOSS. No software to install. You paste one URL into your assistant's settings and your robot becomes a research assistant with library privileges.

Install

Your token is in your welcome email, and this page fills it in for you while you are signed in. Otherwise, wherever a snippet says YOUR-TOKEN, paste your token. The same credential works everywhere at once: laptop, office, phone. Setting it up in a second place does not break the first.

Header form: the endpoint plus an Authorization header

https://mcp.syfert.com/mcp

Authorization: Bearer YOUR-TOKEN

For anything that can set a header: Claude Code, Codex, Cursor, VS Code, Gemini CLI, MikeOSS. It keeps the credential out of URLs, logs and screenshots.

URL form: the credential in the address

https://mcp.syfert.com/YOUR-TOKEN/mcp

For claude.ai custom connectors and ChatGPT developer mode, which cannot set a header.

Claude Code (and any MCP-capable terminal)

claude mcp add --transport http syfert https://mcp.syfert.com/mcp --header "Authorization: Bearer YOUR-TOKEN"

Then /mcp inside Claude Code to confirm it connected.

claude.ai and the Claude desktop app

Settings → Connectors → Add custom connector, name it Syfert Legal Research, paste the URL form, leave everything else blank. Or use a prefilled link:

https://claude.ai/customize/connectors?modal=add-custom-connector&connectorName=Syfert%20Legal%20Research&connectorUrl=https%3A%2F%2Fmcp.syfert.com%2FYOUR-TOKEN%2Fmcp

Substitute your token in the encoded URL at the end. The URL form is the way here: claude.ai's custom-connector dialog has no field for a request header. Then start a new chat and switch the connector on in the tools menu.

ChatGPT (developer mode)

Settings → Apps → Advanced → turn on Developer mode, then Apps → Create, paste the server URL, and set authentication to None. The URL is the credential:

https://mcp.syfert.com/YOUR-TOKEN/mcp

Developer mode needs ChatGPT Pro, Team or Enterprise; the Plus plan cannot add connectors at all. You also have to switch it on in each new chat.

Codex CLI

export SYFERT_MCP_TOKEN=YOUR-TOKEN codex mcp add syfert --url https://mcp.syfert.com/mcp --bearer-token-env-var SYFERT_MCP_TOKEN

Codex refuses a literal token in config.toml, which is why it reads the environment variable instead.

Cursor

In ~/.cursor/mcp.json (global) or .cursor/mcp.json (one project):

{ "mcpServers": { "syfert": { "url": "https://mcp.syfert.com/mcp", "headers": { "Authorization": "Bearer YOUR-TOKEN" } } } }

VS Code / GitHub Copilot

In .vscode/mcp.json. The inputs block makes VS Code prompt for the token once and keep it out of the file you commit:

{ "inputs": [ { "id": "syfert-token", "type": "promptString", "description": "Syfert MCP token", "password": true } ], "servers": { "syfert": { "type": "http", "url": "https://mcp.syfert.com/mcp", "headers": { "Authorization": "Bearer ${input:syfert-token}" } } } }

Gemini CLI

In ~/.gemini/settings.json. Note the key is httpUrl, not url:

{ "mcpServers": { "syfert": { "httpUrl": "https://mcp.syfert.com/mcp", "headers": { "Authorization": "Bearer YOUR-TOKEN" } } } }

MikeOSS

MikeOSS is an open-source (AGPL) legal AI platform for document review, drafting and research, available hosted or self-hosted (source on GitHub). In Mike: Settings → Connectors → + Custom. Label it Syfert Legal Research and fill in the fields below. Leave Custom headers empty.

URL endpoint:

https://mcp.syfert.com/mcp

Bearer token (just the token; Mike adds the word Bearer itself):

YOUR-TOKEN

Click Connect. Mike should report that it found the Syfert tools. Type the URL exactly as shown, with no trailing slash: Mike does not follow redirects. Mike runs on your own Anthropic or Google API key; all Syfert tools arrive switched on, and you can turn individual ones off in the connector's details. More in Mike's connector guide.

Anything else that speaks MCP over streamable HTTP works the same way: give it the endpoint, and the token either as a bearer header or in the path. Stuck? Call 904-383-7448 and we will do it together.

Your research stays yours

After reviewing everything this server could keep, the only log it keeps of your legal research within the MCP is whether a tool call succeeded or failed: which tool ran, how long it took, and whether it errored. It does not record what you searched: not your search terms, not the parties or citations you checked, not the opinions you read. A running record of a lawyer's research is a dangerous thing for anyone to hold, so this machine doesn't hold it.

One thing is stored only if your AI chooses to send it: when it calls suggest_treatment to say a case's red or green flag is wrong, the case, the flag it proposes, its one-line reason, the authority it names and the date are kept for human review. Nothing about you, your session or your research goes with it, and the suggestion never changes a flag by itself.

export_document hands your AI a readable copy of a case or statute to save on your own machine, if it chooses to; which document was exported is not recorded.

Your account

Appearance

Create a free account

You’re in.