Agentic-KB
Agentic Engineering Knowledge Base
Persistent knowledge and context infrastructure for agent systems
Agentic-KB is a persistent, cross-referenced engineering knowledge system for agentic AI, autonomous software delivery, agent memory, evaluations, orchestration, and AI engineering patterns.
It contains 1,000+ compiled articles and exposes the knowledge through a Wikipedia-style web UI, CLI, graph/search interfaces, and an MCP server.
The core idea is simple:
Useful agent memory should be durable, inspectable, attributable, and continuously maintained—not trapped in one chat window or rebuilt from raw context on every run.
Related MCP server: sourcebook
Quickstart
Requires Node 24.x.
git clone https://github.com/jaydubya818/Agentic-KB.git
cd Agentic-KB
npm installBrowse the knowledge base in a browser
cd web && npm install && npm run devOpen http://localhost:3002 for Wikipedia-style search, article rendering, backlinks, and graph navigation.
Query it from the terminal
node cli/kb.js search "multi-agent orchestration"
node cli/kb.js query "What is the best pattern for a supervisor-worker system?"
node cli/kb.js read concepts/tool-use
node cli/kb.js list conceptsThe CLI and web server both use http://localhost:3002 by default. Set
KB_API_URL only when the API is running at a different address.
CLI command reference
This command list mirrors node cli/kb.js help and is checked for drift by the
test suite.
Command | Purpose |
| Show the complete command reference. |
| Search compiled knowledge. |
| Ask a synthesized question through the API. |
| Read one compiled page. |
| List pages in a wiki section. |
| Show pending raw sources. |
| Compile pending knowledge. |
| Run the API-backed knowledge lint. |
| Rebuild |
| Convert and stage a local file. |
| Ingest a YouTube transcript. |
| Ingest a Twitter/X archive. |
| Print a Hermes, Pi, or universal bootstrap. |
| Print the Hermes or Pi acceptance contract. |
| Promote a bus learning. |
| Validate the local environment. |
| List or print a personal agent bootstrap. |
| Preview redaction rules against a file. |
| Show API cost totals. |
| Run the local health checks. |
Repository and bus command | Purpose |
| List tracked repositories. |
| Show repository metadata. |
| Sync one repository. |
| Sync all active repositories. |
| Search imported repository docs. |
| Show repository sync status. |
| List imported repository docs. |
| Show repository progress. |
| Close or preview a repository task. |
| List repository bus items. |
| Publish a repository bus item. |
| Transition a bus item. |
| List repository rewrite artifacts. |
| List canonical repository docs. |
| Read a canonical repository doc. |
Agent-runtime command | Purpose |
| List agent contracts. |
| Show one agent contract. |
| Assemble bounded agent context. |
| Start an agent task. |
| Show the active task. |
| Show recent runtime status. |
| Append durable task state. |
| Verify task-state integrity. |
| Repair recoverable task-state drift. |
| Abandon an active task. |
| Close or preview an agent task. |
| Show recent runtime traces. |
| Preview close-task writes. |
| Scaffold an agent contract. |
| Verify the audit-log hash chain. |
Network commands use KB_API_URL (default http://localhost:3002) and
KB_API_TIMEOUT_MS. Private scopes require PRIVATE_PIN. Repository sync uses
GITHUB_PAT unless --token is supplied. Query and compile operations require
ANTHROPIC_API_KEY.
Expose it to an agent runtime over MCP
node mcp/server.jsPoint any MCP client at that process to get bounded, policy-checked knowledge tools instead of raw filesystem access. See mcp/README.md.
Run the tests
npm testBeyond RAG
Agentic-KB does not treat the knowledge base as a pile of documents behind semantic search.
Raw sources move through an explicit compilation and maintenance process into a persistent wiki:
Raw sources
↓
Ingestion / normalization
↓
Compilation
↓
Cross-referenced knowledge
↓
Lint / graph / contradiction checks
↓
Queryable wiki + CLI + MCP
↓
Agent and human workflowsThe compile step is deliberate, incremental, logged, and auditable. Retrieval remains useful, but the durable asset is maintained knowledge rather than transient context assembly.
What it includes
1,000+ agentic-engineering articles
concepts, patterns, frameworks, entities, recipes, and evaluations
persistent operational memory
cross-referenced wiki links and backlinks
graph-oriented navigation and maintenance
CLI query and maintenance workflows
MCP access for agent runtimes
source citations and contradiction markers
incremental compilation state
ingestion ledgers and durable receipts
private/public knowledge boundaries
linting, stale-content detection, and graph-maintenance checks
agent-driven capture and maintenance workflows
Why this matters for AI-native engineering
As agent systems become more autonomous, context engineering becomes infrastructure.
A durable knowledge layer can help agents and operators answer:
What do we already know about this system?
Which source supports this claim?
Is the knowledge current or stale?
Does another source contradict it?
Which concepts and systems are related?
What was learned from previous execution?
Which knowledge is safe to expose to a given agent?
What should become durable memory versus temporary context?
The objective is not unlimited memory. It is useful, governed, high-signal context.
Interfaces
Web
Wikipedia-style browsing, search, article rendering, backlinks, graph-oriented navigation, and maintenance workflows.
CLI
Command-line access for ingestion, compilation, querying, verification, and maintenance.
MCP
Agent-facing tools expose bounded knowledge operations so external agent runtimes can query the KB without treating the filesystem as an unrestricted authority surface.
Knowledge lifecycle
Agentic-KB distinguishes raw input from compiled knowledge and private/canonical state.
Important design principles include:
Raw content is untrusted input.
Compilation is an explicit state transition.
Sources and citations should survive synthesis.
Contradictions should be visible rather than silently resolved.
Writes should be atomic and recoverable.
Private knowledge must not leak through reports, indexes, or git.
Agent access should be policy-bounded.
Maintenance should be continuously testable.
Reliability and security work
The repository includes extensive correctness and maintenance coverage around areas such as:
atomic writes
SSE/event-stream failure handling
graph and backlink correctness
private-layer exclusions
PIN-gated operations
webhook authentication
MCP error propagation
citation preservation
contradiction signaling
ingestion idempotency
file-descriptor safety
supply-chain pinning and install-script restrictions
The latest maintenance cycle reports 503 passing tests.
Relationship to autonomous software delivery
Agentic-KB is the knowledge/context layer in a broader autonomous-engineering architecture.
Mission Control governs intent, WorkOrders, execution, verification, evidence, and delivery decisions.
Agentic Pi Harness explores governed worker execution and knowledge-access boundaries.
Agentic-KB provides durable knowledge those systems can query without turning transient model context into the system of record.
Mission / WorkOrder
↓
Agent runtime / harness
↓
bounded context request
↓
Agentic-KB
↓
source-backed knowledge
↓
execution + evidenceTechnical themes
context engineering
agent memory
knowledge graphs
MCP
retrieval and synthesis
provenance and citations
contradiction detection
incremental compilation
durable state
privacy boundaries
operational memory
agent-access policy
knowledge maintenance automation
Status
Active and continuously maintained. The project combines a large compiled knowledge corpus with working web, CLI, MCP, graph, ingestion, linting, and maintenance paths. Current development emphasizes correctness, privacy boundaries, durable operations, and making the knowledge layer safer and more useful for autonomous agent systems.
License
MIT — see LICENSE.
Available Tools
37 toolsagent_abandon_taskB
Mark an active task as abandoned. Sets status in working-memory and clears the active-task pointer.
Inputs: task_id (required), agent_id (required), reason.
agent_active_taskB
Return the current active task metadata for an agent, or null if no task is active.
Inputs: agent_id (required).
agent_append_task_stateB
Append a timestamped state entry to the active working-memory file. Requires an active task.
Inputs: entry (required), task_id (required), agent_id (required).
agent_dry_run_close_taskA
Dry-run a close-task operation — returns the full write plan (allowed/rejected ops, bus publications, file writes) without executing anything. Useful for validating a payload before committing.
Inputs: agent_id (required), gotcha, project, rewrites, +4 more.
agent_repair_stateB
Attempt safe repair of an agent task lifecycle state. Rebuilds or clears the active-task pointer when it can do so safely.
Inputs: agent_id (required).
agent_start_taskA
Start a new task for an agent. Creates a working-memory file and sets active-task.md pointer. Returns taskId and paths.
Inputs: agent_id (required), project, task_id, description.
agent_statusB
Return agent lifecycle status: active task, verification issues, close policy, and recent runtime traces.
Inputs: agent_id (required), limit.
agent_traceA
Return recent runtime traces (context loads and close-task writes) for an agent from logs/agent-runtime.log. Useful for debugging context budget issues and rejected writes.
Inputs: agent_id (required), type, limit.
agent_verify_stateB
Verify task lifecycle consistency for an agent. Detects broken active-task pointers and orphan active working-memory files.
Inputs: agent_id (required).
append_repo_progressC
Append a progress entry to wiki/repos//progress.md.
Inputs: repo (required), entry (required), agent_id.
close_agent_taskA
Transactional end-of-task writeback for an agent: appends task log, updates hot, writes gotchas, publishes discoveries/escalations, creates rewrites. Atomic — any forbidden write aborts the whole commit.
Inputs: agent_id (required), gotcha, project, rewrites, +4 more.
close_repo_taskB
Transactional end-of-task writeback for a repo-scoped agent workflow. Appends progress, updates hot memory, writes gotchas, publishes discoveries/escalations, and creates rewrites atomically.
Inputs: repo (required), agent_id (required), gotcha, project, +5 more.
compile_wikiB
Process uncompiled raw documents and compile them into structured wiki pages using Claude. Use mode="incremental" for new docs only (default) or mode="full" to recompile everything.
Inputs: pin, mode.
dry_run_close_repo_taskA
Dry-run a repo close-task operation and return the full write plan without executing any writes.
Inputs: repo (required), agent_id (required), gotcha, project, +5 more.
get_repo_homeC
Get the home page and overview for a tracked repo.
Inputs: repo (required).
lint_wikiA
Run a health check on the wiki: detects contradictions between pages, orphaned pages with no links, stale content, and knowledge gaps. Writes a lint-report.md to the wiki.
Inputs: pin.
list_agent_bus_itemsC
List items in a bus channel, optionally filtered by status.
Inputs: channel (required), status.
list_agentsB
List all agent contracts in the vault.
Inputs: no named parameters.
list_articlesB
List articles in a specific wiki section.
Inputs: section (required).
list_repo_bus_itemsC
List bus items for a repo channel (discovery, escalation, standards, handoffs).
Inputs: repo (required), channel (required), limit, status.
list_reposB
List all tracked repos with their sync status and doc counts.
Inputs: status.
load_agent_contextC
Load a scoped context bundle for an agent, respecting its contract (tier, budget, allowed reads).
Inputs: agent_id (required), project.
load_repo_contextC
Load prioritized context bundle for an agent working on a specific repo.
Inputs: repo (required), agent_id, budget_bytes.
merge_rewriteA
Merge an approved rewrite artifact into the canonical project document. Snapshots previous canonical, writes new with provenance, transitions rewrite to merged. Requires rewrite to be in approved state and approver to meet minimum tier.
Inputs: approver (required), rewrite_path (required), canonical_path (required), force, +3 more.
promote_learningC
Promote a bus item to a target knowledge location with provenance; marks source as promoted.
Inputs: id (required), channel (required), approver (required), target.
promote_repo_learningC
Promote a repo bus item to a canonical or learned location with provenance.
Inputs: id (required), repo (required), channel (required), approver (required), +1 more.
publish_bus_itemC
Publish an item to a bus channel (discovery, escalation, standards, handoffs).
Inputs: body (required), from (required), channel (required), to, +2 more.
publish_repo_discoveryC
Publish a discovery bus item to wiki/repos//bus/discovery/.
Inputs: body (required), from (required), repo (required), project.
publish_repo_escalationC
Publish an escalation bus item to wiki/repos//bus/escalation/.
Inputs: body (required), from (required), repo (required), to.
query_wikiA
Ask a natural language question against the KB using the AI WikiQuery engine. Returns a synthesized answer with citations.
Inputs: question (required), pin, scope.
read_articleB
Read the full content of a wiki article by its slug (e.g. "concepts/tool-use"). Private articles require a pin.
Inputs: slug (required), pin.
read_indexB
Read the wiki master index (index.md) — lists all articles by section.
Inputs: no named parameters.
search_repo_docsC
Full-text search within a specific repo namespace (canonical docs, imported docs, bus).
Inputs: repo (required), query (required), limit.
search_wikiB
Search the Agentic Engineering Knowledge Base for articles matching a query.
Inputs: query (required), pin, limit, scope.
sync_repo_markdownA
Sync a repository from GitHub — fetches markdown docs (.md/.mdx: README/CLAUDE.md plus docs/, specs/, plans/, reports/, architecture/; vendored dirs excluded) and writes them to wiki/repos//repo-docs/, recording sync state in the registry.
Inputs: repo (required), token.
write_repo_task_logC
Append a task log entry for an agent working on a repo.
Inputs: repo (required), entry (required), agent_id (required), task_id.
write_rewrite_artifactC
Create a rewrite artifact in wiki/repos//rewrites//.
Inputs: repo (required), type (required), author (required), content (required), +1 more.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 37 tools
There are parallel general and repo-scoped tools (e.g., close_agent_task vs close_repo_task, publish_bus_item vs publish_repo_discovery/escalation, promote_learning vs promote_repo_learning) that can be confused. Multiple agent state tools (status, verify, repair, trace) also overlap in lifecycle focus, though descriptions help clarify scope.
Names consistently use snake_case and mostly follow verb_noun or noun_verb patterns. Minor deviations exist, such as agent_* prefix tools (agent_status, agent_verify_state) mixed with verb-first names, and dry_run_close_repo_task vs agent_dry_run_close_task ordering.
With 37 tools, the surface is heavy for a knowledge base server. Many parallel repo/agent variants (e.g., close_agent_task, close_repo_task, dry_run_close_repo_task, agent_dry_run_close_task) could likely be consolidated.
The toolset covers core wiki operations (read, search, list, query, compile), full agent lifecycle (start, active, status, verify, repair, append, abandon, close, trace), bus publishing, repo sync/context/task close, and rewrites. Minor gaps include explicit article delete/update and agent contract management, but most workflows are supported.
Maintenance
Related MCP Connectors
A shared team wiki your coding agents read and write — across every repo and every MCP client.
Self-hostable team wiki; agents read & write it via MCP; Atlas turns your repo into a cited wiki.
Hosted markdown project wikis your team's AI assistants read, search, and update over MCP.
Shared project wiki for AI agents: read and write pages, next actions, and activity logs over MCP.
Related MCP Servers
- AlicenseNot gradedqualityCmaintenanceGoverned knowledge base for AI agents via the Model Context Protocol (MCP), enabling agents to search, read, and contribute persisted knowledge with versioning, audit trails, and approval workflows.65 npmMIT
- AlicenseNot gradedqualityBmaintenanceMCP server that gives AI coding agents a git-backed markdown wiki to read and update, enabling search, read, write, verify, ingest, promote, and lint operations on versioned knowledge documents with schema validation, staleness tracking, and contradiction detection.4MIT
- AlicenseNot gradedqualityCmaintenanceA machine-local, agent-writable knowledge base MCP server that lets coding agents autonomously record and retrieve hard-won operational facts, incidents, and corrections as markdown pages in a git repo, with provenance, talk pages, and search.2 npmMIT
- FlicenseAqualityDmaintenanceProvides coding agents with a shared, Markdown-based engineering knowledge base to search, capture, create, and update internal conventions, API details, infrastructure configs, and development setup via lightweight MCP tools.6-