jpx

JMESPath + 400 functions.
A jq alternative you can actually remember.

Josh Rotenberg

JMESPath syntax — the whole language

Access
foo.bar.baz  # nested field
[0]  [-1]  [3:5]  # index, last, slice
[*].name  # list projection
*.values  # object wildcard
[][]  # flatten nested arrays
Multi-select
{n: name, a: age}  # hash
[name, age]  # list
Pipe
[*].score | sort(@) | [-1]
Literals & refs
`42`  `"str"`  `true`  `[1,2]`  # JSON literals
'raw string'  # unquoted string
&field  # expression reference
Filter
[?age > `30`]
[?active && score >= `90`]
[?!archived]
[?contains(name, 'test')]
Functions
length(@)  sort(@)  reverse(@)
sort_by(@, &age)  max_by(@, &score)
join(', ', @)  contains(@, 'x')
keys(@)  values(@)  merge(a, b)
Let expressions (JEP-18)
let $lat = `40.42`,
    $lon = `-3.70` in
[*].{dist: geo_distance_km(
  $lat, $lon, lat, lon)}
Current node
@  # the value being evaluated

Query language for JSON
that reads like what you want

curl -s https://api.github.com/users/octocat \
  | jpx '{
      login: upper(login),
      joined: format_date(parse_date(created_at), `"%B %Y"`),
      repos: public_repos
    }'

{"login": "OCTOCAT", "joined": "January 2011", "repos": 8}

JMESPath is an RFC spec. jpx adds 400+ functions on top — dates, hashing, geo, fuzzy matching, and more.

32 categories. 400+ functions.

string array object math datetime hash encoding regex geo fuzzy network semver validation expression text phonetic url color duration ids

sha256('secret') → "2bb80d5..."

geo_distance_km(40.42, -3.70, 37.77, -122.42) → 9318 Madrid → SF

jaro_winkler('Josh', 'Joshua') → 0.93

format_date(now(), '%B %d, %Y') → "February 10, 2026"

The jpx ecosystem

jpx	CLI with REPL, streaming, table/CSV output
jpx-mcp	MCP server — 29 tools for AI assistants
jpx-core	From-scratch JMESPath engine in Rust
jpx-engine	Introspection, BM25 search, query store
jpx (PyPI)	Python bindings via PyO3

MCP server: 29 tools — AI agents can query JSON, discover functions, and build query libraries without the data ever entering the context window.

Why give this to an agent?

Without jpx: agent reads full JSON into context

60 KB → 21,469 tokens

With jpx MCP: file stays on disk, agent sends expression, gets result

92 + 263 = 355 tokens — 60x reduction

It scales. It's correct.

Input	Without jpx	With jpx	Reduction
85 earthquakes (60 KB)	21K tokens	~200 tokens	~100x
200 earthquakes (143 KB)	51K tokens	~200 tokens	~250x
100 Nobel laureates (367 KB)	131K tokens	~150 tokens	~850x

Result size is constant — the output depends on the query, not the input.

Deterministic: the query engine processes the data, not the LLM.
No hallucinated values. No miscounted arrays. Same result every time.

Tool discovery across servers

# Agent connects to 3 MCP servers — 100+ tools total
# Registers each server's tools with jpx at session start
register_tools({ server: "redisctl", tools: [...65] })
register_tools({ server: "github",   tools: [...20] })
register_tools({ server: "deploy",   tools: [...18] })
 
# Now search across all of them with natural language
query_tools("backup database")
→ redisctl:enterprise_database_backup score: 8.4
→ redisctl:enterprise_database_export score: 5.2

BM25 search — finds "backup" from "snapshot", "export", "persist".
The agent doesn't scan tool lists. It asks jpx.

Get started

# Install
brew install joshrotenberg/brew/jpx
# or: cargo install jpx

# Discover functions
jpx --search "hash"
jpx --describe sha256

# Interactive REPL
jpx --repl

# Reusable query libraries
jpx -Q earthquakes.jpx:mag-stats -f data.json

github.com/joshrotenberg/jpx
joshrotenberg.github.io/jpx