Browser / quickest first proof
Open a working tool, load its explicit fictional example and inspect the full output. No assistant account or installation is needed.
Choose a browser tool →Portable assistant methods · Evaluation release 0.6.0
Bring the methods
to your assistant.
Eight focused skills and four review utilities support the same eight review workflows. One local MCP server provides thirteen tools for those workflows, fictional quote data, connected cases, case comparison, policy libraries and pilot scorecards. Earlier focused and native downloads remain available below. Public assistant directory acceptance and model activation are separate checks.
Browser access is available at llmsquared.com/tools without an assistant account.
Choose your route
Start where you already work.
Claude Desktop / prepared local bundle
Review the unsigned Evidence Desk 0.6.0 MCPB and its included installation instructions. Use your approved extension process and verify a fictional tool call.
Current Claude Desktop bundle →Gemini CLI / prepared extension
Review the current 0.6.0 extension archive and included instructions. A compatible local Node runtime and approved client setup are prerequisites.
Current Gemini CLI extension →Other local MCP / portable commands
Inspect the Evidence Desk source and supplied stdio installation instructions, or select a separate analyst CLI below. A download does not establish model use.
Current local source package →Grok web and remote connectors require a separately reachable service. This website does not expose a production remote MCP endpoint. Public listing, installation and model activation remain separate validation steps.
One local connection · Evaluation release 0.6.0
Eight workflows.
One assistant package.
Evidence Desk supplies thirteen deterministic tools: eight review workflows, a fictional quote example, a connected case runner, a case comparison utility, a policy library review and a pilot scorecard. Twelve focused and utility skills cover the eight workflows and the four review utilities. Comparison reruns both explicit case drafts, retaining source-use and evidence gaps in each result.
Follow INSTALL.md in the source ZIP. Use a compatible Node runtime, install the pinned dependencies, test fictional inputs and confirm the tools in your chosen client. Installing skills alone does not connect the server. The static website package does not create a public remote MCP endpoint.
Protocol initialization, tool listing and supplied-data tests are separate from client installation, model activation, directory review and paid delivery. Source rights, human confirmations and control evidence remain declarations. Tool annotations do not enforce authorization. An assistant host may send supplied inputs and results to its model provider under that provider’s terms.
The earlier focused packages and previous native bundles remain available for an explicitly chosen smaller or earlier installation. Local developer execution and public directory publication follow separate requirements.
Current prepared formats · Evaluation release 0.6.0
Use the format your client expects.
These unsigned bundles contain the current source release and pinned runtime dependencies. The Claude Desktop archive follows MCPB 0.3 format; the Gemini CLI archive includes its extension manifest and twelve root skills. Inspect the package and follow your client’s approved local extension process. A compatible Node runtime is required.
Archive integrity, manifest checks and extracted-server protocol tests do not establish installation, model activation or directory acceptance. Grok web Custom MCP connectors and the xAI remote MCP API require a reachable server. Grok Build documents local MCP and skill routes. These downloads supply a local stdio service; actual connection, host discovery and model use require separate checks.
Previous native releases — retained archives
Previous 0.5.0 native bundles
The previous local edition adds guided policy-library evidence.
Previous 0.4.0 native bundles
The previous release adds case comparison to the eight workflows and connected case utility.
Previous 0.3.0 native bundles
The previous release supports eight workflows and connected cases. It does not include case comparison.
New / Separate portable analyst kits
Bring the two new methods
to your local workflow.
Each source package includes its own focused skill, deterministic dependency-free Node CLI, the same browser engine, fictional input and review instructions. These are separate from Evidence Desk 0.6.0 and do not add tools to its thirteen-tool server.
PowerOn Portfolio Analyst 1.0.0
Literal authorized source inventory and complete reviewer handoff. No compilation or core access.
Download portable PowerOn analyst →Agent Access Analyst 1.0.0
Exact-scope grant evaluation and explicit event metrics in a local simulation. No actual identity verification.
Download portable Access analyst →Node installation is a prerequisite for CLI use. Assistant-host installation, model invocation and directory acceptance require separate validation.
Connected review · No account required
Related reviews.
One useful handoff.
Review Desk brings all eight workflows into one case. Queue explicit inputs, see which jobs need corrections, inspect the full reports and export a portable case or printable handoff. Every export is checked against a fresh run of the current engines.
A prepared report can still contain missing evidence or blocked source use. The desk does not retrieve documents, approve a decision or grant access. Your entries stay in browser memory; exported cases contain all supplied inputs and results.
Guided policy evidence / No account required
Bring the source files.
Leave with a reviewable handoff.
Collect permitted UTF-8 policy text, record owners and versions, and declare the controlling sources. Add review questions and exact phrases. Inspect source gaps, line excerpts and follow-ups before preparing a complete handoff.
1. Prepare the collection
Select permitted .txt or .md files and record source metadata. New source declarations begin false. Conflicting controlling versions stay visible until you correct them.
2. Inspect the evidence
Review literal matching lines with source references and version details. See missing terms and source issues. Found phrases require a person to interpret them.
3. Keep the correction inspectable
Save the complete library and handoff. Export a bounded Review Desk input draft, then use Review Comparison to inspect later declared changes.
This browser utility processes explicit source text in memory. It does not answer policy questions, convert PDF or Word files, retrieve links or approve decisions. Save your files before closing.
Pilot evidence / No account required
Measure the whole task.
Keep the quality checks visible.
Compare a matched sample of an existing workflow with the assisted workflow. Include preparation, operation, review, corrections, re-entry and handoff. Missing observations remain unknown, and incomplete or rejected samples stay outside favorable comparisons.
Capture complete effort
Use all six phases for each path. Keep actual zero separate from an unmeasured value. Labor rates are optional; an unknown rate prevents a cost comparison.
Explain exclusions
Pair the same units and record supplied quality and reviewer declarations. A faster incomplete or rejected sample cannot support the eligible total.
Prepare the debrief
Save a complete input-bound report and printable handoff. Effort and modeled labor-cost differences describe only the eligible supplied sample.
This utility does not authenticate reviewers, verify acceptance or prove payroll reduction, cash savings, causality or buying demand. Your references remain text and are not retrieved.
Local case utility · No account required
See what changed.
Keep both reviews inspectable.
Compare two explicitly supplied Review Desk drafts for the same case. Both versions run through the current eight workflow engines. Inspect exact declared values, added and removed jobs, workflow and label changes, and both complete reports.
A changed input or engine result describes supplied values. It does not establish that an evidence gap was addressed, that risk decreased or that a decision should proceed. Downloads preserve the complete current comparison after a fresh parity check.
Sixth workflow · Working preview · Version 0.1.0
Invoice Evidence.
Compare explicitly entered invoice lines with approved fixed USD rates and quantities. Rate differences, arithmetic errors, unmatched scope and source gaps remain separate. Missing or unsupported evidence withholds complete totals.
The invoice downloads are separate from the five decision skills and six-tool server above. The invoice server exposes one tool: reconcile_invoice_terms. Follow its included INSTALL.md and confirm a fictional call in your chosen client.
The browser preview accepts typed fields and structured CSV or saved JSON. It does not extract documents, interpret legal terms, verify source permission, authorize payments or establish refunds or achieved savings. Taxes, credits, discounts, tiered rates, minimums, proration and foreign currency are outside this version. Quantity caps cover only the supplied invoice.
Local MCP initialization, listing and calls passed using the official client, along with clean package installation. Actual model activation, public remote access and directory acceptance remain unverified.
Focused workflow · Working preview
Vendor Evidence Review
Compare a buyer’s checklist with declared evidence types, dates, versions and reviewed mappings. Export an owned action queue.
Uses the buyer’s declared checklist and explicit reviewed mappings. It does not ingest documents, assess compliance or approve a vendor. No active monitoring, source retrieval or account action is included.
Focused workflow · Working preview
Agent Permission Review
Inspect declared capabilities, resource scope and control evidence. Separate evidence completeness from priority review.
Uses a declared inventory. It does not authenticate identities, discover hidden agents, verify effective permissions, enforce approvals or revoke access. No active monitoring, source retrieval or account action is included.
Start with one real decision
Try the method.
Review the result.
Fictional examples demonstrate the behavior without customer information. Use the free browser tools to see the method before installing an assistant package.
Explore all browser toolsEarlier focused six-tool setup and publication notes
Earlier five-skill package
Guide the review.
Policy sources, workflow value, vendor comparison, quote changes and renewal planning. The skills keep missing evidence visible and ask for review before interpreting a result.
The skills package includes installation notes and a Claude compatibility manifest. A skills package can guide an assistant without providing a calculation engine.
Earlier six-tool server
Calculate and inspect.
Compare vendor terms, load a fictional vendor example, plan renewals, review quote changes, inspect policy sources and model workflow value.
The local server processes supplied inputs. It has no account access, payment capability, automatic file-reading or outbound service calls.
Install the earlier six-tool server.
- Download and extract the local server ZIP into a folder you control.
- Install a supported Node.js version 22.13 or later, then follow the included
INSTALL.mdinstructions. Dependency installation needs internet access. - Use the included client configuration example with the actual path to
assistant/server.mjs. - Confirm the six tools are listed in your MCP client. Try fictional data before supplying permitted real information.
The package was tested with the official MCP client for initialization, tool listing and calls. Installation in your own assistant account is not established by that test. The included INSTALL.md is the authoritative installation guide.
Once installed, the local server does not call an outside model. Your chosen assistant may send inputs to its provider under that provider’s terms. Local processing by the server does not make the whole assistant session local.
| Assistant or client | Available route | What remains |
|---|---|---|
| Browser | Fifteen browser experiences, including focused reviews and connected utilities, at Free tools. No assistant account required. | Save drafts on your device; use an approved route for retained information. |
| Claude Desktop and Claude Code | A compatible local MCP client can launch the supplied stdio server. The skills ZIP includes a Claude manifest. | Install and confirm tools in your own client. Claude directory publication requires review and public repository requirements. Official plugin submission guide ↗ |
| ChatGPT and Codex | The focused skills package is prepared for the current plugin format. Public ChatGPT MCP access needs an HTTPS service reachable by OpenAI. | Directory review, owner identity, metadata and relevant validation. A local stdio server is not a public ChatGPT endpoint. Official submission guide ↗ |
| Grok Build / xAI API | Grok Build documents local stdio MCP. Grok web connectors and the xAI API use a reachable remote service. | Local Grok Build setup is documented in INSTALL.md but has not been installed for this release. This ZIP has no remote hosting. xAI does not support the MCP require_approval option. No consumer Grok directory listing is claimed. Official remote MCP guide ↗ |
| Gemini API | Gemini CLI supports local stdio MCP. The Google Gen AI SDK also documents MCP use. | Follow the included Gemini CLI configuration instructions and verify a call in your client. Installation and Gemini consumer extensions remain unverified. Official MCP guide ↗ |
Keep the publication steps specific.
For a public remote connector, host a separate maintained HTTPS service, define access and retention, test the actual clients and prepare reviewer access. The website ZIP is a static Bluehost site and does not start an MCP service.
For a plugin directory, use the supplied skills package as a review candidate. Complete owner identity and submission requirements, then address the directory’s review findings. No marketplace approval or discoverability is promised here.
OpenAI currently restricts digital-service sales and indirect upsells through published plugins. Any independent website engagement must be reviewed against the rules of the directory involved. Read OpenAI’s current guidelines ↗