Voyager Agent MCP server

@dir-ai/voyager-agent

The one universal Voyager. A thin, model-independent orchestrator that turns the modular Voyager senses — voyager-repo (code), voyager-net (hosts), and the cognitive contract that binds them — into a single agent. You name a goal; Voyager activates the senses it needs, runs a live mission, links the evidence into one graph, and concludes.

It is composition, not a monolith. The agent never absorbs the senses — it calls the published organs and adapts what they see into the shared CognitiveClaim substrate. The reasoning model is a swappable seam (Brain); a rule-based brain ships so it runs with no LLM at all.

Sensing is autonomous. Acting is not. Every sense here is read-only. Any

remediation a mission surfaces is carried as a consent-gated, withheld

action and is never applied by this orchestrator.

Install

npm i -g @dir-ai/voyager-agent

Use

# Orient in a repo and flag risk (voyager-repo)
voyager-agent mission "audit this project" --repo .

# Audit ONE host you own (voyager-net — fail-closed, needs --authorized)
voyager-agent mission "check my server's exposure" --host example.com --authorized

# Observe a live page (voyager-browser — read-only, static HTML)
voyager-agent mission "check my landing page" --url https://example.com

# All three senses in one mission; vet 10 dependencies; machine-readable
voyager-agent mission "full audit" --repo . --host example.com --authorized --url https://example.com --check-deps 10 --json

Senses fan out in parallel and are resilient: a sense that errors (or even throws) becomes a low-confidence, flagged observation — it never crashes the mission or discards a sibling sense's good result.

The mission then re-plans: it picks the single most informative next probe (by information gain) and executes it, bounded by maxRounds. How a probe is executed is an injectable dispatch seam (the model decides); the safe default only ever re-reads (deepens repo dependency-vetting) and never acts. A verify claim closes the mission so state().satisfied is meaningful.

Voyager — one agent, one entry.
goal: "orient in this repo and flag risk"

  👁 observe  [repo]  (70% · moderate)
     @dir-ai/voyager-agent@0.1.0 — The one universal Voyager…
     ↳ [low] no-lockfile: no lockfile — installed versions are not pinned/reproducible
  🧠 infer  [memory]  (70% · moderate)
     repo observed; 0 high/critical signal(s). No high-severity issue surfaced.

next probe: [repo] open src/index.ts

As an MCP server

One tool, run_mission — describe a goal, get back the full claim graph.

voyager-agent mcp            # stdio

{
  "command": "voyager-agent",
  "args": ["mcp"]
}

As a library

import { runMission, DeterministicBrain, type Brain } from '@dir-ai/voyager-agent'

const { mission } = await runMission(
  'audit this repo and my host',
  { repoPath: '.', host: 'example.com', authorized: true /* your own host */ },
  Date.now(),
)

for (const claim of mission.allClaims()) console.log(claim.operation, claim.sense, claim.verdict)
console.log(mission.state().bestNextProbe)   // the most informative next step
console.log(mission.state().contradictions)  // where two senses disagree

Bring your own model

The Brain interface is the whole point — it's where an LLM plugs in without touching the senses, the memory, or the mission machinery. A ready-made LlmBrain ships: give it one model-agnostic complete() function and it plans and synthesizes through the model. Wire Claude, a local server, or any OpenAI-compatible endpoint — Voyager never depends on a specific SDK.

import { runMission, LlmBrain } from '@dir-ai/voyager-agent'

Shortened. The full README is on GitHub.