vaultbase

Give your agents real context.

VaultBase exposes 27 MCP tools for vault reads, writes, decision traces, intelligence, and admin. Agents get the same permission system as humans — scoped to their role, loaded with org context, and able to write back what they learn.

claude-code / agent briefing
# Agent Briefing — engineering-agent
Identity stack loaded: 3 tiers
org: Acme Corp — "We ship weekly, test everything, prefer Go"
team: Platform — "Own infra, SLO targets, on-call rotation"
user: shane — "Senior eng, focus on observability"
Pending outcomes: 2 decisions awaiting results
Stale references: 1 trace citing revised policy
Knowledge gaps: 3 topics with failed searches
Role: engineering-write — read/write to engineering/**, read shared/**

How it works

27 MCP tools across 5 categories

VaultBase implements the Model Context Protocol with a comprehensive tool surface. Agents can read and search the vault, create and query decision traces, run intelligence reports, manage admin tasks, and generate briefings — all through structured MCP calls.

Vault 8 tools

vault_read, vault_search, vault_list, vault_learn, vault_update_doc, vault_read_context, vault_tag, vault_move

Traces 6 tools

trace_create, trace_find_precedent, trace_update_status, trace_list, trace_get, trace_link

Intelligence 5 tools

intel_knowledge_gaps, intel_stale_refs, intel_pending_outcomes, intel_briefing, intel_consistency_check

Admin 5 tools

admin_users, admin_roles, admin_groups, admin_audit_log, admin_permissions

Briefings 3 tools

briefing_generate, briefing_skill_file, briefing_context_stack

Three-tier identity stack

Every agent briefing loads context at three levels: organization, team, and individual. The org level defines company-wide standards and preferences. The team level adds domain-specific context. The individual level personalizes for the specific user the agent is acting for.

This means an agent working for an engineer on the platform team gets different context than one working for a product manager on the growth team — automatically, without prompt engineering.

Permission-scoped access

Agents inherit the permissions of the user they represent. The same RBAC rules that control human access — folder scoping, tag restrictions, deny rules, group inheritance — apply to every MCP tool call. An agent with an engineering-read role cannot access sales documents.

API keys use a vk_ prefix and are SHA-256 hashed at rest. You can issue multiple keys per user for different agent roles — one for Claude Code, another for a CI pipeline, each with distinct permissions.

Write-back and skill files

Agents aren't read-only. vault_learn() lets an agent add new knowledge to the vault. vault_update_doc() lets it revise existing documents. Both respect permissions and create proper revision history.

The /skill.md endpoint generates a single-file onboarding packet scoped to the agent's role. It bundles identity context, relevant precedent decisions, and the tool surface into one file that any MCP-compatible client can consume.

Proactive agent briefings

Instead of waiting for agents to ask the right questions, VaultBase pushes context to them. A briefing includes: pending decision outcomes that need follow-up, stale references where cited documents have been revised, knowledge gaps identified from failed searches, and recent traces relevant to the agent's domain.

This turns agents from passive responders into proactive participants. They can flag that a decision from last month needs a status update, or that a policy they're citing has been revised since the last trace.

Works with your tools

Claude Code

Plugin + MCP

Cursor

MCP integration

Any MCP Client

Standard protocol

Why it matters

Agents that know your org. The identity stack means agents don't start from zero. They know your company's standards, your team's practices, and your personal focus areas. Every response is grounded in real organizational context.

Security by default. Agents get the same RBAC as humans. No separate permission model to maintain, no risk of agents accessing documents their user can't see. Deny rules propagate to every tool call.

Agents that learn. Write-back means agents can capture knowledge as they work. An agent that resolves an incident can document the resolution. An agent that researches a vendor can save its findings. The vault grows with use.

One protocol, many clients. MCP is an open standard. VaultBase works with Claude Code today and any MCP-compatible client tomorrow. No proprietary SDK, no vendor lock-in on the agent side.

Related

Download VaultBase

Free. Works offline. No account needed.

Open source. AGPL-3.0. Plain markdown files. Export anytime.