Glossara Get notified
Self-hosted and private

Shared memory for your agents.

And knowledge management for your teams.

Give your agents one place to read, write and remember, whichever vendor they come from, with the same permissions and the same audit trail your people have.

How it works

People and agents, working together

A colleague in a browser, an assistant over MCP and a script over the API pass the same three checks before they reach anything. Agents propose. People approve. The log records who did what.

How people and agents interact with Glossara People in a browser, agent hosts over MCP, and scripts over REST send requests through an HTTPS front door into the Glossara server. Each request passes a rate limiter, authentication and authorization before it reaches the web app, the MCP server or the REST API. Those surfaces read and write the knowledge store, which lives in your database. Background workers embed pages and type links using a model server you configure. Responses carry cited answers, tool results and the list of changes an agent asked for. A refused request bounces off authorization with its reason, and the server writes each request to the log. People Use the web app in a browser. Signed in with a session. Owner, member or viewer. Agents Claude Code, Cursor and others. Connect over MCP with a token. Limited to the spaces you choose. Scripts and services Call the REST API directly. Act for themselves or for a person. Fetch what changed since last visit. HTTPS front door · TLS Glossara server Rate limiter per address per token per write Limits enforced across all servers. Authentication who is asking? session cookie, or bearer token Access revoked when tokens expire. Authorization may they do this? role in the space token scope read or write Out of scope stays invisible. Web app edit · ask · graph MCP server /mcp · tools REST API what changed? Your knowledge Markdown · typed graph · search index request log · who, what, and why a request was refused · one line each session dana@… GET /spaces/…/pages 200 · 4ms token it-ops-bot POST /blocks 201 · delegated-by priya@… · 11ms token viewer-bot POST /blocks 403 · reason role-lacks:viewer:write · 2ms Database your data, your storage pages, links, log Background workers embeddings, typed links Model server any OpenAI-compatible local or hosted cited answer tool result what changed 403 role-lacks:viewer:write
People, with a session Agent hosts over MCP Scripts over REST, asking for changes Refused, with the reason logged Workers talking to your model server

The MCP server

Assistants such as Claude Code and Cursor connect over MCP and get the same access a person has: search, ask, read and write. A read-only token limits them to reading. Connecting takes one snippet.

Automatic updates

When one agent adds or changes something, the others pick it up on their next visit, without re-reading the whole space. Context stays current across vendors and sessions.

Authentication

People sign in and hold a session. Agents hold a token that a space owner issues, with an expiry date.

Authorization

The server enforces owner, member and viewer roles on each request. A token covers the spaces you choose, for reading or for writing. Anything outside that scope stays invisible to the agent. All actions are logged, including refusals, and why.

An agent asks a question

The agent searches, asks, reads the answer and its citations, then opens the block it needs. In the first scenario the answer comes from other agents' contributions. In the second it comes from your team's own documents.

An agent asks a question A deployment agent searches for the approved identity provider and finds three blocks written by other agents and a person. It asks which provider a new service should use and gets an answer citing all three. It opens block two for the settings, then writes what it did back to the shared notes, where the other agents will see it. Agent transcript MCP tools: search · ask · get_block · write_block ▸ search("approved identity provider") 3 blocks · Identity provider decision · SSO rollout notes · AD retirement ▸ ask("Which identity provider should a new service use?") Okta is the approved provider for new services [1]. The SSO rollout notes give the OIDC settings [2]. Active Directory retires in Q4 [3]. ▸ get_block([2]) "Register the app in Okta. Use OIDC with PKCE. Map group claims to spaces." found it ▸ write_block("SSO rollout notes", "payments-api registered in Okta") written · block [4] · by deploy-bot Each answer carries citations that point at a block, so the agent can open the exact passage. Your knowledge bars: match by meaning and by phrase Identity provider decisionby security-agent · Cursor[1]SSO rollout notesby it-ops-bot · Claude Code[2]SSO rollout notes“Register the app in Okta. Use OIDC with PKCE. Map group claimsto spaces.”✓ block [2][4]newActive Directory retirementby Dana Okafor · person[3]Payments API runbookby deploy-bot · scriptOn-call handbookby Marcus Bell · person written by deploy-bot · other agents see it next visit
Capabilities

What Glossara does

For agents, for your team and for the people who run it.

For your agents

  • Memory shared across vendorsAgents from different vendors read and write the same spaces and learn from each other's notes.
  • Agents as membersAn agent joins a space under its own name, with its own token and its own permissions.
  • MCP serverAssistants such as Claude Code and Cursor connect with one snippet, for reading or for reading and writing.
  • REST APIScripts and services read, search, ask and write.
  • Block-level writesSeveral agents work on one page without overwriting each other.
  • Automatic updatesAn agent picks up what changed since its last visit, so its context stays current.
  • Acting for a personA write records the agent and the person it acted for.
  • Cited answers for agentsAsk returns an answer with citations, or says the pages hold no answer.
  • Command line and SDKsA small command-line tool and client libraries on a published API.

For your team

  • Markdown, byte for byteDocuments stay Markdown, edited one block at a time, with tables and footnotes intact.
  • Spaces and page treesA space for each team or project. Import a folder of Markdown in one step.
  • Search by meaning or by phraseSemantic and keyword search together, with a keyword fallback if the model is offline.
  • Answers with citationsNumbered citations back to the source, and a plain no when your pages do not cover the question.
  • Links with reasonsThe model records whether a page extends, contradicts or exemplifies another, and why.
  • Shared spaces with rolesOwners, members and viewers, invited by link, with an activity feed.
  • Native tables and calloutsTables, callouts and richer lists on the same block model.
  • Linked sourcesA Google Doc or a page in another wiki, indexed and cited beside your own pages.
  • Page history and exportOpen a page's history. Export a whole space as Markdown.
  • Contradictions you can closeWhen your team settles a conflict between two pages, the link records that it is resolved and from when. The graph and answers stop treating it as open, and the history stays.

Operations and security

  • Self-hostedOne server process and one database, on a container platform, a virtual machine or bare metal.
  • Your model, your choiceAny OpenAI-compatible model service, local or hosted. One model for answers, another for link typing.
  • Hashed credentials and scoped tokensSessions and tokens stored as hashes. A token covers chosen spaces, for reading or for writing.
  • Rate limits and an audit trailPer-address, per-token and per-write limits. One log line per request, with the reason for a refusal.
  • Scales horizontallyAdd servers as your usage grows.
Security

Private by construction

Your data stays on your hardware unless you connect an outside service. Glossara checks the identity behind each request, person or agent. These are the defaults.

Credentials stored as hashes

Glossara hashes sessions and agent tokens before storing them. It shows an agent token once, at creation, with an expiry date.

Scopes that hold

A token covers the spaces you choose, for reading or for writing. Outside that scope, a page stays invisible. Inside it, a caller without the right role gets a refusal with the reason, and the log keeps a copy.

Rate limits

Glossara limits requests per address, tightens the limit on sign-in attempts, and gives each agent token its own budget with a separate budget for writes. Behind a load balancer the token budgets live in the database, so the limit stays one limit.

An audit trail

Each request produces one log line: the route, the result, the timing and the identity behind it. Glossara logs no request bodies, headers or query strings.

Your model, or none

Glossara bundles no model and recommends none. Until you connect one, the AI features stay off. Run a model on the same machine, on another server, or through a hosted provider that speaks the OpenAI API.

Two uses, one source of truth

Shared memory for agents. Knowledge management for people.

One set of pages, one set of permissions, one log. Agents do the reading and the drafting. People decide what stands.

Shared memory for your agents

Agents from different vendors share one memory. A Claude Code agent, a Cursor agent and a script you wrote read and write the same space under their own names. One agent records what it learned, another picks it up on its next visit, and a third asks a question and gets an answer that cites both. Facts, decisions and open questions live in one place instead of inside each vendor's own memory.

Knowledge management for your team

Your people write in Markdown and organise pages into spaces. They share a space with owner, member and viewer roles. They search by meaning or by phrase, ask questions and get cited answers, and see how pages relate and why. They approve what agents propose.

More on knowledge management