Skip to content

FAQ

How is this different from an MCP gateway, Okta or Cloudflare?

Section titled “How is this different from an MCP gateway, Okta or Cloudflare?”

Those decide who may connect. Deedglass decides what each tool call may do, runs it, then checks the result against the system of record and seals a signed deed of what actually happened. Use them together: an identity provider for people, a network edge for traffic, Deedglass for agent actions. See How Deedglass works.

That Deedglass compared what the agent reported with an independent observation and found no difference. The level says which: the tool’s own answer (L0), a read-back (L1), or the change Postgres committed (L2). It doesn’t mean the action was a good idea. The exact definition.

It adds a decision, two short database transactions and a seal to every call. In three 60-second open-loop runs at 200 calls a second on an M1 laptop (on AC power, with Postgres on the same machine), an allowed call with no witness took 5.8 ms at the median and 19.4 ms at p99, end to end over loopback (bench/results/2026-10-05-darwin-arm64-bench-gate.json, the median of the three runs). That includes HTTP, the database writes and the sealed deed. The median is just above the 5 ms target and the p99 is under the 25 ms one; a CPU profile shows the time going to the two Postgres round trips (admission, then the outcome and the seal), not to policy evaluation, which is under 1 ms at p99. A run on a server comes before launch. An L2 verdict arrives after the change log is read, so it adds no time to the call when you don’t wait for it.

Nowhere you didn’t configure. Self-hosted, Deedglass sends no telemetry. Payloads stay in your database until retention prunes them; deeds hold only digests. Data handling and privacy.

Because Postgres keeps its own record of every committed change, the change log, and Deedglass can read it. That makes L2 possible: comparing the agent’s report with what the database actually committed. Other systems are checked by reading back (L1).

Not through Deedglass: denied calls never run, stopped agents are refused, and the attack suite tests the ways a manipulated model tries. But an agent with another path to a system (its own credentials, a direct database URL) isn’t governed on that path. Scan your setup to find those paths.

Calls fail closed: the MCP wrapper and the SDKs refuse to run a governed tool they can’t get a decision for, and say so. Nothing runs ungoverned. A call that was running when Deedglass went away is recorded as unknown, then UNVERIFIABLE.

The core is free and Apache-2.0. Pro adds the Action on private repos, the full report and Pro packs, from $29 a month.

Deedglass is built by Arjun Prabhakaran, a solo founder in India.