The platform for shared agent context. Build tools that read or write context injected into agent fleets. Connect data sources, sync, monitor, scan, improve, compress, benchmark, and more.
Alignbase is the system of record for agent context. Teams write, version, tag, publish, and audit the context their agents should use. This repo is the open-source platform layer around that system of record: a place to build tools that work with Alignbase context through agent-facing APIs.
Agents are starting to work in groups. A company might have coding agents, browser agents, support agents, data agents, workflow agents, and custom agents running at the same time. They all need the same current context, but most tools still treat context as something copied into one session at a time.
Alignbase treats context as shared infrastructure. Alignbase repos are the source of truth for shared context: documents, tags, permissions, publication state, OAuth authorization, and audit history. This repo is for the ecosystem around that source of truth.
That means a tool should be able to:
- Read the context an approved agent would receive.
- Write better context back into Alignbase when it has permission.
- Sync outside systems into context documents.
- Check context quality, freshness, size, and drift.
- Help agent fleets use the right context without each tool needing its own private memory system.
This repo is early. The intended shape is:
- A core plugin framework for tools that connect to Alignbase.
- First-party plugins that cover common jobs like sync, scan, compression, benchmarking, and monitoring.
- Contributor guides for third-party plugins.
- Shared helpers for auth, permissions, context reads, context writes, logs, and test fixtures.
- Examples that show how to build a useful context tool without knowing the internals of the Alignbase repos.
The first plugins should stay small and concrete. A good plugin does one job well, uses the Alignbase permission model, and leaves an audit trail when it writes context.
Alignbase exposes agent access through MCP at /mcp. Agents authorize through
OAuth with PKCE. Each approval maps to a specific agent authorization, so reads
and writes can be tied back to the agent that made them.
Core operations include:
get_current_context, which returns the published context routed to the calling agent.- Listing context documents and reading a specific document.
- Creating, editing, and publishing context when the agent has the matching permission.
- Creating tags when the agent has permission to do so.
Tags are both routing and access control. A plugin should not treat tags as labels only. Tags decide which agents receive which context and which agents can write back to which parts of the context repository.
Some useful starting points:
- Onboarding: turn repo docs, runbooks, and setup notes into draft context for a new team or project.
- Sync: keep context in step with a system like GitHub, Linear, Notion, Slack, Google Drive, or an internal service catalog.
- Scan: find stale context, missing owners, broken links, unclear instructions, and policy conflicts.
- Improve: suggest shorter context, clearer routing, better tags, and safer publication state.
- Compress: reduce token use while preserving the operational facts agents need.
- Monitor: track context freshness, read volume, write volume, agent coverage, and failed authorization attempts.
- Benchmark: compare agent results with and without specific context bundles.
- Alignbase remains the source of truth. Plugins can read from or write to Alignbase, but they should not become separate context repositories.
- Permissions come first. A plugin should request the narrowest permissions it needs and should fail clearly when it lacks access.
- Published context is special.
get_current_contextreflects what agents receive by default. Draft and published state should stay distinct. - Context windows are scarce. Good tools help teams route, trim, and test context instead of dumping more text into every session.
- Every write should explain itself. If a plugin changes context, a human should be able to see what changed, why, and which agent or authorization did it.
- Open-source plugins should be portable. A contributor should be able to build against documented APIs without reading the Alignbase repos.