Back to Gap Map Product / UX / Agent identity & authorization

Halofy

Halofy
open source / Overall score: 1.4

Halofy is a self-hostable kernel that governs what organizational context AI agents may read and write. Agents reach it over MCP or HTTP with credentials the server resolves into a namespace, actor and role, every read and write passes the same access checks, and each outcome, denials included, lands in an append-only audit log. It also stores durable context and issues signed erasure certificates. Halofy builds it.

Openness

5 high confidence
5.0
license
AGPL-3.0 (or later; kernel and console)+Apache-2.0 (driver kit in kernel/src/driver, MCP server in kernel/src/mcp and integrations in kernel/integrations)
source
public
core features withheld
no — the 543-path tree at 3763e64 carries AGPL-3.0 at the root and Apache-2.0 in the three subtrees NOTICE names, with no ee/, enterprise/ or license-key path; the AI gateway, managed channels, hosted connector catalogue, tuned extraction prompts and cross-organization federation are sold as a separate managed service
commercial
Halofy managed service and commercial license (the commercial license removes the AGPL obligation for the same code)

Halofy's kernel and console are AGPL-3.0 and its MCP server, driver kit and integrations Apache-2.0, and the open build carries the whole identity, access, audit and erasure layer with no license key. Halofy sells a managed service around it, with model routing, hosted connectors, federation across organizations and a tuned extraction prompt set that is not published.

Adoption

1 low confidence
1.0

Halofy is installed from its repository, and the kernel's package name is not published on npm, so GitHub stars are the only signal. A star is not a use, so this shows attention rather than deployments.

Capability

2 medium confidence
2.0

Halofy gives each agent a credential that the server resolves into its namespace and role, and checks every read and write of organizational context against them before it runs. It has no delegation between agents and no human approval step, and its audit log is append-only but not hash-chained or signed.

Verified 2026-11-20, except some scores, still awaiting re-confirmation