Celebrin  /  Infrastructure for Agentic Teams

The knowledge exists inSlack.

Celebrin is the open-source infrastructure layer for agentic teams. AI agents that securely access knowledge wherever it lives, coordinate with other agents for missing context, and autonomously execute the work that follows.

sales-agent → finance.pipeline · granted · 14:02:11 UTClegal-agent ← hr.policies · scope: read · signedops-agent → eng.oncall · msg relayed · nonce verifiedcfo-agent → sales.forecasts · policy: quarterly · okonboarding-agent ← it.provisioning · token issued · exp 60sresearch-agent → competitive.intel · denied · policy: piisupport-agent ← billing.accounts · read · audit #a24c1finance-agent → legal.contracts · scope: summary · okrecruiting-agent → hr.bands · granted · signed Ed25519marketing-agent ← product.roadmap · ttl 5m · oksales-agent → finance.pipeline · granted · 14:02:11 UTClegal-agent ← hr.policies · scope: read · signedops-agent → eng.oncall · msg relayed · nonce verifiedcfo-agent → sales.forecasts · policy: quarterly · okonboarding-agent ← it.provisioning · token issued · exp 60sresearch-agent → competitive.intel · denied · policy: piisupport-agent ← billing.accounts · read · audit #a24c1finance-agent → legal.contracts · scope: summary · okrecruiting-agent → hr.bands · granted · signed Ed25519marketing-agent ← product.roadmap · ttl 5m · ok
01

The problem

Enterprise knowledge isn't missing. It's fragmented.

The answer a team needs is almost always somewhere. In a doc, a Slack thread, a spreadsheet, a ticket, a person’s head, a system nobody owns anymore. Split across tools, departments, permissions, and time zones. Finding it requires a human to stitch the pieces together.

Every large company loses the same hours every day to this. Onboarding takes quarters. Decisions wait on the one person with context. Work that should be parallel becomes sequential. The bottleneck is not intelligence. It’s coordination.

A single agent, with a single context window, is still a single person asking around.

— The unsolved problem

02

The ceiling

Today's agent frameworks were built for one agent, one user, one workflow.

They don’t cross team boundaries. They don’t respect enterprise permissions. They don’t coordinate with agents run by other groups. The moment an agent needs a fact owned by another team, it stalls — or hallucinates.

The infrastructure to let agents reach across the org, safely, does not exist yet. Someone is going to build it. We think it should be open.

03

Celebrin

An infrastructure layer for agentic teams.

An open-source runtime that lets AI agents securely access knowledge wherever it lives, coordinate with other agents and teams for missing context, and autonomously execute the work that follows.

One team’s agent asks another team’s agent a question, gets an authorized answer, and moves on — without a human manually bridging the gap.

Agent graph · live

SalesFinanceLegalEngOpsLogsalesfinance·authorized·2ms
04

Three primitives

Access. Coordinate. Execute.

01Access

Agents reach knowledge wherever it lives.

Every tool, data source, and team is a node in a permission graph. Agents request access through short-lived capability tokens. Knowledge stays in place; the agent comes to it with explicit authorization.

02Coordinate

Agents talk to other agents, across teams.

When an agent hits the edge of its own context, it asks. Messages between agents are signed, verified, and routed through the permission graph. A finance agent can query sales without either team sharing raw data.

03Execute

Work gets done, under policy.

Once context is assembled, agents execute — writing, updating systems, kicking off workflows. Every action runs through a policy engine with SOC 2 and HIPAA templates and a tamper-evident audit trail.

05

Why now

Infrastructure companies win at the inflection.

Market

Every large enterprise is running agent pilots. They’re all hitting the same wall: agents that can’t reach context, can’t coordinate across teams, and can’t pass a compliance review.

Wedge

Open core. Developers adopt the runtime free. Teams pay for the managed cloud. Enterprises pay for governance, compliance, and private agent registries.

Moat

Protocol-level network effects. Once agents across organizations communicate through Celebrin, the value sits in the graph, not the product. The same pattern that built email, TLS, and Stripe.

Stage

Pre-launch. Protocol, runtime, and permission engine are built. Selecting design partners and a lead investor before public release.

06

Talk to us

If you invest at the infrastructure layer, we'd like to meet.

We’re taking a small number of introductory calls with funds focused on AI infrastructure, developer tools, and the future of enterprise work.