MCP serverbuilt on TypeSafe Jev

The Jev MCP server for coding agents.

A fast decision layer for Claude Code, Codex, Cursor, and nine more agents. agent-fastpath answers the small questions agents keep stopping on (is CI ready to ship, which files handle auth, is this command safe) with typed, calibrated results, and hands control back when it is not sure.

claude mcp add agent-fastpath -s user -e TYPESAFE_API_KEY=<key> -- npx -y agent-fastpath start

Other agents → npm

98.8%

Less host context in file triage

< 1 ms

Rules decide clear cases

~350 ms

Typical Jev decision

3 states

Accept • Review • Escalate

triage 35 source files for "SSRF protection"

measured

Agent reads the files itself

Files read by the agent: 35

Host context used: 35,256 tokens

Agent calls fastpath_triage

Host context used: 408 tokens (98.8% less)

Latency: 2.9 s

Top hit: network_policy.ts (correct)

Run on the agent-fastpath repository, September 2026. Method and full results

How it works

Rules first, then Jev, then a confidence gate

Clear-cut cases are decided by code without an API call. Everything else goes to TypeSafe Jev in one request, and a gate turns low-confidence answers into review or escalate instead of a guess. Select a stage to see what it does.

What it guarantees

From the source

Examples

Real responses

Each example is a recorded response from agent-fastpath 0.2.0. The semantic ones were answered by TypeSafe Jev, so your confidence values will differ slightly.

Input the agent sent

Result

Which stage decided

Response

Benchmark

Estimate the context you save

Triage reads files on the server, so your agent only sees ranked results with short snippets. Move the sliders to estimate the saving for your repository.

Files to triage

Average file size

Triage calls per day

Agent reads every file

Agent reads the triage result

Host tokens avoided per call

Host tokens avoided per day

Less host context

Estimate. It assumes the triage response stays near the measured 408 tokens (top 10 results with snippets). Measured result: 35 files, 35,256 tokens reduced to 408. Triage sends redacted excerpts of up to 4,000 characters per file to TypeSafe for scoring, so the saving is in your agent's context, not in data sent. How it was measured

Why a System One layer

Coding agents need fast reflexes, not only deep reasoning

Your agent is good at slow, deliberate work. It is wasteful at the quick judgments in between. agent-fastpath takes those.

Many agent steps are small judgments: did the tests pass, which files matter, is this command safe, is this request contradictory. When the agent answers them itself it reads whole files and logs into its context, then reasons in free text. agent-fastpath answers them in one call: rules for the clear cases, TypeSafe Jev for the rest, and a typed result the agent can act on.

System 2 (your agent) vs. System 1 (agent-fastpath)

	Your coding agent	agent-fastpath
Good at	Planning, reasoning, writing code	Ship gates, triage, risk and ambiguity checks
Speed	Seconds per model turn	< 1 ms (rules) to ~350 ms (Jev)
Reads	Whole files into its context	Files on the server; returns ranked snippets
Answers with	Free text	Typed decision, probabilities, and a status

Questions

Why not let the agent read the files itself?

It works, but every file it reads stays in its context for the rest of the session and competes with your instructions. In the benchmark, triage cut the agent's context for the same task from 35,256 tokens to 408.

Why not grep or a script?

agent-fastpath does use exact rules first, and they answer in under a millisecond. Rules cannot judge meaning, like whether a requirement contradicts itself or whether a test run proves the change. Those go to Jev.

What happens when it is not sure?

The confidence gate decides. Below 0.55 confidence the result is escalate: decide yourself or use a larger model. When the top two answers are within 0.15 of each other the result is review. Contradictory or vague requests return ask_user, risky commands return review with ask_user, and failing builds return blocked with fix_and_retry. You can tune the thresholds per call.

Does my code leave the machine?

Rules run locally. Semantic questions send the redacted state to TypeSafe. Triage reads up to 64 KB per file inside the directories you allow and sends redacted excerpts of up to 4,000 characters per file for scoring. Traces stay in memory (the last 500) and are not written to disk. Do not triage files that must never leave the machine.

Do I need a TypeSafe API key?

For semantic questions, yes. Without a key the server still runs: rules still decide clear cases, and semantic questions return escalate instead of guessing.

Compared with a plain Jev connector

More than a pipe to Jev

A basic Jev MCP server passes your agent's questions to Jev and returns the answer. agent-fastpath adds what an agent needs to act on that answer safely.

	Plain Jev connector	agent-fastpath
Questions	The agent writes question schemas	Presets: ship_gate, risk, severity, ambiguity, relevance, verify_claim, and more
Clear cases	Every call goes to the API	Rules decide failing builds, force pushes, and DROP TABLE in under 1 ms, with no API call
Uncertain answers	Raw probabilities; the agent decides	A gate returns accept, review, or escalate, with a recommended action
Files	The agent reads them first	Server-side triage; files stay out of the agent's context
Web pages	Not covered	Headless browser with SSRF blocking and an irreversible-action gate
Audit	Varies	Secret redaction and an evidence trace for every decision

Security

Safe by default

Tool inputs, logs, and web pages can be hostile. You set the limits; tool callers can tighten them but never loosen them.

SSRF blocking

The browser cannot reach localhost, private networks, or cloud metadata endpoints such as 169.254.169.254. Every connection, including redirects and page scripts, goes through a checking proxy.

Irreversible-action gate

Elements labelled like delete, pay, checkout, publish, deploy, or merge are flagged. The browser will not click them unless the call sets allowIrreversible: true.

Bounded file reads

Triage reads only inside the directories you allow, after resolving symlinks, and at most 64 KB per file.

Secret redaction

TypeSafe, Anthropic, OpenAI, GitHub, AWS, and Slack keys, private keys, bearer tokens, and password fields are redacted before semantic requests and traces.

Details in the security model. Report vulnerabilities privately through GitHub security advisories.

Setup

Connect your coding agent

Pick your agent and copy the config. You can also print it with npx agent-fastpath install-config <agent>.

Node.js 20.12+Runs through npx; nothing else to install.

TypeSafe API keyNeeded for semantic questions. Rules work without it.

Chromium (optional)npx playwright install chromium for the browser tool.