Skip to main content
Glama
deenesjakoruzh
jnot807

Juicebox MCP

by jnot807
OverviewSchemaSourceRelated ServersDiscussionsAdmin
JavaScript
Local

Juicebox MCP

A local MCP server that reads your Juicebox sourcing data into Claude — saved searches and their scored results — using your own logged-in Juicebox session.

Runs entirely on your machine. Your session never leaves it, and every call is made as you, on your own seat.

Reads cost no export credits. Everything the read tools return comes from the same free surface the search-results page already renders. One tool writes, and says so: jb_run_search creates a real saved search in your workspace.

Install

Option A — Desktop Extension (easiest)

Download juicebox-mcp.mcpb from Releases, then double-click it, or drag it into Claude Desktop → Settings → Extensions.

There is no API key to paste. After installing, do the one-time browser steps below.

Option B — from source

git clone https://github.com/jnot807/juicebox-mcp.git
cd juicebox-mcp
npm install          # also downloads the Chromium build (see note)
npm run login        # a real browser opens — sign in to Juicebox yourself
npm run check        # proves the session works headless

Then register it with Claude Code:

claude mcp add -s user juicebox -- node "$(pwd)/server.js"

-s user makes it available in every session; without it the registration is scoped to whatever directory you happened to run it from.

The one-time browser download

This drives a real Chromium, and that binary is not part of node_modules — it is a one-time download of roughly 500MB into a shared cache (~/Library/Caches/ms-playwright on macOS).

npm install fetches it automatically via a postinstall step. Desktop Extension users need to run it once by hand, because an extension bundles node_modules but not that cache:

npx patchright install chromium

If it is missing, the server tells you so in plain language rather than throwing a stack trace about a missing executable.

Signing in

Authentication is a real sign-in, not a key. npm run login opens a browser window; sign in to Juicebox as you normally would. The session is then stored in session/ (gitignored, chmod 600) and reused headlessly.

Sign in again whenever npm run check starts failing — sessions expire.

Related MCP server: JobSpy MCP Server

Tools

Tool

What it does

jb_list_searches(projectId?)

Saved searches on a project (id + name).

jb_get_results(searchId, limit?, minMatchRate?)

A search's ranked candidates — name, LinkedIn URL, title, company, location, matchRate, per-criterion verdicts, and dated experience[] + education read off the rendered cards. Up to ~500 per call.

jb_count(queryInput, searchId?)

Size a filter set without running a search — the tuning primitive. queryInput is a PATCH over a harvested template; check noEffect in the response.

jb_run_search(prompt, need?)

WRITES. Creates and runs a new search from a natural-language prompt, then returns its candidates. Leaves a saved search visible to your whole workspace — confirm before using it.

experience[] is the only way to see past employers: the API payload carries just the current one, so alumni of a target company are invisible without it.

Which project it reads by default

Nothing is hardcoded. At sign-in, a probe loads /projects, which redirects into a project your seat can see, and that id is saved as defaultProjectId in session/session-meta.json.

It is written once and then left alone. The redirect follows whichever project the app most recently had open, so trusting it on every run would make a tool call with no projectId read a different project than it did yesterday.

Resolution order:

  1. JUICEBOX_PROJECT_ID (env — this is what the Desktop Extension's optional "default project" field sets)

  2. JUICEBOX_VALIDATOR_PROJECT (env — also pins the auth check to that project)

  3. defaultProjectId in session/session-meta.json, set by discovery

Every tool also takes an explicit projectId, which always wins.

Juicebox project ids are ~20-character keys like c5PheL2fANnX6uBQVUdo — the /project/<id>/ part of a URL. If you pass a UUID, the server rejects it with an explanation rather than silently navigating to a project that does not exist.

Two rules the tools carry

  • verdictFound: false → unknown, never a negative. "No evidence found" and "evidence says no" are different verdicts. Collapsing them scores a candidate down for a criterion nobody could actually check.

  • Broad skill terms dilute ranking. Skills are OR-weighted; a population-wide term like "Account Management" on a customer-success search inflates the pool by roughly 3.4×. Drop the generic terms and promote the one hard requirement to a skill filter.

Running scripts while the server is up

You can't share the browser profile: session/profile/ is single-writer, and the MCP server holds it whenever it is running. A second process trying to open it fails the auth check — which reports itself as "session expired" and sends you round in circles re-logging-in.

For diagnostics, build a fresh context from the checkpoint instead. No lock, same session:

const { chromium } = require('patchright');
const browser = await chromium.launch({ headless: true });
const context = await browser.newContext({ storageState: 'session/storage-state.json' });

How it works, and the traps

The results page is server-rendered on first load, so /api/profiles/results only fires on interaction. The client nudges the pager to make the app issue its own request, then captures the response — which carries the whole ranked set, not just the visible page.

Three things that will bite anyone editing client.js:

  1. Never use addInitScript. Patchright silently no-ops it as an anti-detection measure — no error, the script just never runs. Use page.on('response').

  2. The API's linkedin_url is encrypted (hex:hex), as are profiles[].url and profileDetails.id. Real URLs come from the rendered cards and are joined on normalised full_name — measured at 100% on a live search.

  3. The list blanks mid-pagination. A null pager reading means "still moving", not "failed". Gating anything on pager-change detection during a transition is how two earlier bugs happened.

When it breaks

This rides Juicebox's internal API. There is no stability contract, and it can change without notice.

  • npm run check fails → session expired: npm run login.

  • The server says Chromium is missing → npx patchright install chromium.

  • jb_get_results returns source: "dom-fallback" → the API capture broke; you lose matchRate and criteria. Check RESULTS_PATH still matches.

  • jb_get_results reports joinedLinkedInUrls: 0 → the card markup changed; revisit harvestCards / rewindToFirstPage.

  • Empty search list → the project page markup changed; see listSavedSearches.

Requirements

  • Node.js 18 or newer

  • A Juicebox account you can sign into

  • ~500MB free disk for the Chromium download

Licence

MIT. Not affiliated with or endorsed by Juicebox.

Available Tools

4 tools
jb_countA

Size a filter set WITHOUT running a search — the tuning primitive. Free and side-effect free, so permute filters and compare before committing to a run. The endpoint is VERIFIED DETERMINISTIC (same body, same count on repeat), which is what makes the comparisons below trustworthy.

queryInput is a PATCH, not a whole body. It is merged over a ~164-key template harvested live from a saved search, because a PARTIAL body does not error — it returns result:0, which reads exactly like "nobody matches your filters". Pass only the keys you want to change. Use searchId to choose which saved search supplies the template; the response echoes baseline (the unpatched count).

ALWAYS CHECK noEffect IN THE RESPONSE. Unrecognised keys and wrong value shapes are dropped SILENTLY, so an unchanged count means the key name is probably wrong, NOT that the filter does not matter. Measured example: coSizes (company headcount) is ignored in all eight shapes tried, including [{name:"51-200"}], ["51-200"], [{title,tag}] and [{min,max}]. Its real shape is still unknown — set the filter in the Juicebox UI and capture the request before trusting it.

SOME FILTERS EXPAND RATHER THAN RESTRICT, which is the opposite of the intuition and was measured on one search (baseline 1413): dropping industries gives 1387 and dropping coTags gives 1247 — REMOVING them SHRINKS the pool, so they act as OR-expanders and cannot be used to narrow onto a category. An unrecognised industry ("mining & metals") returns the same 1387 as an empty list. What actually restricts: skills (removing all four widened 1413 to 4526), yearsOfExp (20 gives 881), selectedLocs, and criteria. Note this REVISES the older claim that broad skill terms widen the pool — specific skills restrict hard, broad ones merely fail to restrict. The practical rule is unchanged: drop skill terms that describe the whole population, and promote the ONE hard requirement to a skill (e.g. "AI Agents", not prose about being AI-first).

ParametersJSON Schema
NameRequiredDescriptionDefault
searchIdNoSaved search to harvest the template from (from jb_list_searches). Defaults to the first saved search on the project. Pick the one whose titles/locations you want as your starting point.
projectIdNoJuicebox project id — a ~20-character key like c5PheL2fANnX6uBQVUdo, taken from the /project/<id>/ part of a Juicebox URL. Optional: omit it to use the default project discovered at sign-in.
queryInputYesPATCH over the harvested template — only the keys you are changing. Verified keys: jobTitles[{title,tag}], pastJobTitles, skills[{title,tag}], industries[{name}], coTags[{title,name,tag}], selectedLocs[{name,country,continent,type,...}], selectedCos, timezones, revenueRanges, fundingStages, yearsOfExp (number), criteria[{text,id}], excludeDncCountries (bool), coSizes (present but IGNORED in every shape tried).

TDQS

A5/5.0
Behavior5/5

Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?

With no annotations provided, the description carries the entire behavioral burden, and it excels. It reveals that the endpoint is deterministic, that partial bodies are silently merged over a template, that unrecognized keys are dropped silently, and that filter behavior can be counterintuitive (some expand rather than restrict). It even provides measured examples showing how to detect silent failures via the noEffect flag. This is exemplary transparency.

Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.

Conciseness5/5

Is the description appropriately sized, front-loaded, and free of redundancy?

The description is long, but every sentence carries critical, non-redundant information. It front-loads the core purpose and safety properties, then layers in warnings and measured data. No filler or repetition exists; each paragraph adds actionable knowledge an agent needs to use the tool correctly.

Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.

Completeness5/5

Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?

Given the tool's complexity (nested objects, no output schema) and the absence of annotations, this description is remarkably complete. It covers expected behavior, failure modes (silent drops, noEffect), practical examples, and even guidance on filter semantics that only become apparent through experimentation. An agent has everything it needs to call this tool correctly and interpret results.

Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.

Parameters5/5

Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?

Although the schema already describes each parameter, the description adds value by clarifying that queryInput is a PATCH over a template, enumerating the verified key shapes (e.g., 'skills[{title,tag}]'), and explicitly flagging coSizes as ignored in every tested shape. It also explains how searchId selects the template and the meaning of the echoed 'baseline.' This goes well beyond the schema definitions.

Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.

Purpose5/5

Does the description clearly state what the tool does and how it differs from similar tools?

The description opens with a clear verb and object: 'Size a filter set WITHOUT running a search — the tuning primitive.' This immediately distinguishes it from running a search and states its role as a sizing/measurement tool. It also explicitly contrasts with the sibling jb_run_search by saying 'before committing to a run,' leaving no ambiguity about what the tool does.

Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.

Usage Guidelines5/5

Does the description explain when to use this tool, when not to, or what alternatives exist?

The description gives explicit when-to-use guidance ('permute filters and compare before committing to a run') and explains the side-effect-free nature that makes this safe for experimentation. It also instructs on selecting the saved search template via searchId and warns about the partial-body PATCH semantics. This is far more than typical usage guidance.

Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.

jb_get_resultsA

Pull a saved search's ranked candidates — up to ~500 in one call. Returns name, LinkedIn URL, title, company, location, a numeric matchRate (0-100, banded) and per-criterion verdicts. Costs NO export credits: this is the same free surface the results page renders.

READING THE RESULTS: at the top of a ranked list every candidate meets every criterion, so criteria-met count does NOT discriminate up there — use matchRate and depth. A criterion with verdictFound=false is reported as result "unknown": that means no evidence was found, NOT that the candidate fails it. Never treat an unknown as a negative.

If the search has never been executed, this errors — open it in the app and press "Run search" once.

ParametersJSON Schema
NameRequiredDescriptionDefault
limitNoMax candidates to return (default 50, max 500).
searchIdYesSaved search id from jb_list_searches.
projectIdNoJuicebox project id — a ~20-character key like c5PheL2fANnX6uBQVUdo, taken from the /project/<id>/ part of a Juicebox URL. Optional: omit it to use the default project discovered at sign-in.
minMatchRateNoOnly return candidates at or above this matchRate.

TDQS

A4.7/5.0
Behavior5/5

Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?

No annotations are provided, so the description carries the full burden. It discloses the free/costless nature, the exact return data, the significance of verdictFound=false ('not a negative'), and the error condition for unexecuted searches. It also notes that matchRate is 'banded' and that criteria-met count doesn't discriminate at the top of a ranked list. This is rich behavioral detail well beyond structured fields, giving the agent a clear understanding of what happens when calling the tool.

Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.

Conciseness5/5

Is the description appropriately sized, front-loaded, and free of redundancy?

The description is long but every sentence earns its place: the first line states the action and output, the cost note is critical, the 'READING THE RESULTS' section is indispensable for correct interpretation, and the final error note is essential. It's well-structured with clear sections (purpose, cost, interpretation, error condition). No fluff or repetition.

Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.

Completeness5/5

Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?

There is no output schema, but the description enumerates the return fields and explains the semantic meaning of matchRate and verdicts (including the crucial 'unknown ≠ negative' nuance). It covers error behavior, limits, and the relationship to search execution. With four parameters fully documented in the schema and the description adding inter-tool references, an agent has everything necessary to correctly invoke and interpret results.

Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.

Parameters4/5

Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?

The input schema covers all parameters with descriptions at 100% coverage, so the baseline is 3. The description adds meaningful context: it tells the agent that searchId comes from jb_list_searches, that projectId is a specific format and can be omitted to use a default project, and that limit has a default of 50 and max of 500. These references connect the tool to its ecosystem and clarify usage, exceeding the schema's raw field meanings.

Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.

Purpose5/5

Does the description clearly state what the tool does and how it differs from similar tools?

The description begins with a specific verb ('Pull') and a specific resource ('a saved search's ranked candidates'), and details the exact fields returned (name, LinkedIn URL, title, company, location, matchRate, per-criterion verdicts). This clearly distinguishes it from siblings like jb_list_searches (which lists searches) and jb_count (counts). The scope is precise and non-tautological.

Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.

Usage Guidelines4/5

Does the description explain when to use this tool, when not to, or what alternatives exist?

The description explicitly states that this costs no export credits and is the same surface as the results page, giving context on when it's appropriate. It also provides a prerequisite: 'If the search has never been executed, this errors — open it in the app and press Run search once.' It doesn't explicitly name sibling alternatives, but the distinct purposes of siblings (list, count, run) make the usage boundary clear. The 'READING THE RESULTS' section adds interpretive guidance, though it's more about reading the output than choosing when to use the tool.

Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.

jb_list_searchesA

List the saved searches on a Juicebox project (searchId + name). Saved searches are stable and re-runnable, so this is the entry point for pulling existing sourcing work. Defaults to the configured project.

ParametersJSON Schema
NameRequiredDescriptionDefault
projectIdNoJuicebox project id, or a /project/{id} URL. Optional.

TDQS

A4.4/5.0
Behavior4/5

Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?

With no annotations provided, the description carries the full burden. It discloses that saved searches are stable and re-runnable (reassuring the agent that reuse is safe), and that it defaults to the configured project. This adds useful behavioral context beyond the raw schema. It doesn't mention pagination or error behavior, but for a simple list operation this is adequate.

Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.

Conciseness5/5

Is the description appropriately sized, front-loaded, and free of redundancy?

Two sentences: the first states the purpose and output, the second adds context and default behavior. No wasted words, and the most important information is front-loaded.

Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.

Completeness4/5

Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?

For a one-parameter list tool with no output schema, the description is nearly complete: it explains what is returned, the stability, and the default. It could mention whether an empty list is returned when none exist, but that is minor. It doesn't need to explain return values since it already says 'searchId + name'. Missing details are negligible.

Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.

Parameters4/5

Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?

The schema already describes projectId as optional and as a URL or ID (100% coverage). The description adds value by stating it defaults to the configured project, which is not in the schema. That's meaningful extra guidance for when an agent can omit the parameter.

Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.

Purpose5/5

Does the description clearly state what the tool does and how it differs from similar tools?

The description clearly states the action (list), the resource (saved searches on a Juicebox project), and the output (searchId + name). It also situates the tool as an entry point for pulling existing sourcing work, distinguishing it from siblings like jb_run_search or jb_get_results without needing to inspect them.

Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.

Usage Guidelines4/5

Does the description explain when to use this tool, when not to, or what alternatives exist?

It says this is the entry point for pulling existing sourcing work, implying you should use this before running or fetching results. It also notes the default-project behavior, which helps an agent decide when explicitly passing projectId is unnecessary. It doesn't explicitly say 'don't use this if you want results' or name alternatives, but the use case is implied clearly enough.

Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

  1. 4 tool updatesv1.0.0
    • First observedjb_count
    • First observedjb_get_results
    • First observedjb_list_searches
    • First observedjb_run_search
Deploy Server
One endpoint for every MCP toolRayrun runs every MCP server behind one URL and holds the credentials. Each call is decided per client, tool, and argument, then recorded.ray.runAd

TDQS

A4.6/5.0

Scored across 4 tools

Disambiguation5/5

Each tool has a clearly distinct purpose: list saved searches, fetch ranked results, tune filter counts, and create/run a new search. No two tools overlap in function, making misselection unlikely.

Naming Consistency5/5

All tools follow the consistent pattern of 'jb_' prefix + verb_noun (list_searches, get_results, run_search) with count as a concise verb. This is predictable and uniform.

Tool Count5/5

Four tools is well-scoped for a talent sourcing server. Each tool earns its place and covers the core workflow without redundancy or bloat.

Completeness4/5

The surface covers the essential lifecycle: listing existing searches, retrieving results, tuning filters without execution, and creating new searches. Minor gaps like deleting or updating saved searches exist but are not critical for the primary use case.

Maintenance

ActivityMaintained
ResponsivenessSyncing

Resources

Related MCP Connectors

Related MCP Servers

  • Socials MCP

    Brainrot-Creations
    A
    license
    A
    quality
    C
    maintenance
    Enables Claude to perform actions on X, LinkedIn, and Reddit via a Chrome extension, such as connecting with recruiters, finding threads, and replying to posts.
    36
    8 npm
    1
    MIT

  • JobSpy MCP Server

    chinpeerapat
    A
    license
    Not graded
    quality
    A
    maintenance
    Enables job search and scraping across multiple job boards (LinkedIn, Indeed, Glassdoor, etc.) with advanced filtering, directly from Claude Desktop or other MCP clients.
    5
    MIT

  • job-search-mcp

    jppoamaral
    A
    license
    Not graded
    quality
    C
    maintenance
    Enables searching and evaluating job postings from LinkedIn and freehire.me directly through Claude Desktop. Provides tools to search jobs, fetch full posting details, and assess candidate fit using eligibility scans and a scoring rubric.
    MIT