Skip to main content

Overview

Jevlin is SuperSearch for coding agents. SuperSearch asks every search provider at once, says which ones found each result, and gives you the receipt. Jevlin asks eight providers in one call and answers with one list. Search is billed at each provider's own price with no markup. Available in Claude Code, Cursor, Codex, opencode, Pi and Hermes.

Every search is observable on the platform. A trace shows one search with a receipt per provider: what each returned, how long it took, what it cost. A trajectory shows one thread, an agent's session, as the path it took, turn by turn: what the agent was working on when it searched, what came back, and what it concluded. Each search carries the context that makes this possible, hashed and scrubbed before it leaves the machine. See Trajectory observability.

Setup registers you with the search platform and gets a key, with no browser. A new account starts with $10 of credit; see Pricing and credit and How an agent gets an API key.

Get started​

Install it, then run setup. Setup registers your installation, writes a skill and hooks into each coding agent it finds, and prints a link to claim your agent. The quickstart walks through it.

npm install -g @jevlinsearch/jevlin
jevlin setup

Install it globally, never with npx. Setup writes the binary's path into your agents' skills and hooks, and npm throws away an npx cache or a project's node_modules. An npm install is updated with npm install -g @jevlinsearch/jevlin@latest.

What you can do​

Here are some of the ways you can use Jevlin:

Search the web from your agent

The skill makes Jevlin the default search in Claude Code, Cursor, Codex, Pi and Hermes, and the rules block you paste into opencode's AGENTS.md sends its public-web searches here too, so an ordinary question that needs the web is enough. Name it to be explicit:

/jevlin what changed in Vite 7's rollup migration

See Search from your agent.

Ask eight providers at once

A balanced search asks four independent search engines and a max search asks all eight (Exa, Brave, Parallel, Serper, TinyFish, Perplexity, Ceramic and Seltz), and merges what they return into one list deduplicated by page. Each entry says which providers found it, so agreement between independent indexes is evidence the agent can read.

{"version":1,"query":"postgres 18 uuidv7 default column performance","tier":"balanced"}

See Providers and tiers.

Inspect every search as a trace

Each search appears on the platform with a receipt per provider: whether it answered, how long it took, how many results it returned and what it cost. Open one when an answer looks thin and see which arm let it down.

See Trajectory observability.

Follow a thread as a trajectory

Searches carry hashed session, turn and call ids and the assistant text that led to them, so the platform shows a thread, one agent session, as the path the agent took: what it was working on when it searched, what came back, and what it concluded, turn by turn, across compactions and subagents.

See Trajectory observability.

Keep queries and secrets off the wire

The query travels as JSON on stdin, never on the command line. The context that travels with a search is scrubbed of anything shaped like a credential, a key block, an email address or your account name before it leaves, and JEVLIN_TRACE=off sends none of it.

export JEVLIN_TRACE=off

See What leaves the machine.

Tune a search per request

Besides query and tier, a request can carry objective, search_queries, recency, domain_filter, max_results, merged_limit, max_chars, page_chars, cache and providers. Every field is in the stdin request.

{"version":1,"query":"array flatten method","domain_filter":["developer.mozilla.org"],"max_results":5}

See Request options.

Report what a turn concluded

With turn_end on, the agent's end-of-turn hook also sends what you asked, the final answer, and each tool called by name, duration and outcome, never a tool's input or output, and only for turns that searched.

[search]
turn_end = true

See The turn end.

Script it from a shell

The same binary answers a person or a script. --stdin is the versioned contract; status, doctor, connect and limits take -json.

jevlin search --stdin <<'JSON'
{"version":1,"query":"what is proof of authority consensus"}
JSON
jevlin status -json

See Search from a shell.

Use Jevlin everywhere​

One installation serves every agent on the machine: the same key, the same tier, the same sessions directory. Setup finds the agents on PATH and writes what each one needs.

What I want to doBest option
Search from Claude CodeClaude Code: a skill, hooks and allow rules in settings.json
Search from CursorCursor: a skill and hooks, one lineage file per conversation
Search from CodexCodex: a skill, hooks and a permission profile that opens only the search hosts
Search from opencodeopencode: an in-process plugin and a rules block for AGENTS.md
Search from PiPi: a skill and an auto-discovered extension
Search from HermesHermes: a skill and one hook in config.yaml
Use it from an agent that is not listedAny other agent: paste the rules block
Call it from a script, CI or my own shellSearch from a shell
Make the agent's own search the default again/jevlin off in the agent, or jevlin agents prefer off; see Search from your agent
Ask for a wider search on a hard question/jevlin tier balanced; see Choosing a tier
See what my agents searched, and whyplatform.jevlin.ai, Traces and Trajectories; see Trajectory observability

Next steps​