New Sakshi sandboxes are live. Try agent governance on synthetic data, in your browser. Get a sandboxarrow_forward
witness · Architecture

Built so an auditor never has to take our word for it.

Sakshi sits beside the agents you already run. It tokenizes identifiers before they are ever stored, records every decision on a hash chain anyone can recompute, bounds what an agent may do on its own, and turns that evidence into regulator-mapped packs, all inside your environment.

# any auditor, in your environment
GET /chain/verify
{ valid: true,
    checked: 1,842,
    head: "3af9a89e…8bf1" }

The path of a decision

Every decision takes the same accountable path

Whether the agent reaches Sakshi through the SDK, the gateway, or MCP, the same pipeline runs: no raw identifier is stored, and nothing skips the record.

01

Agent acts

SDK · gateway · MCP

02

Tokenize

PII replaced at ingest

03

Witness

decision + context captured

04

Hash onto chain

each record seals the last

Autonomy is bounded before the action fires; the kill switch outranks everything.

The spine

A tamper-evident chain, not a database you're asked to trust

Every decision is hashed into a chain where each record seals the one before it. Recomputing the chain re-derives every hash, so tampering with a single field is not just detectable, the chain says exactly where it happened. There is nothing to take on faith.

PII is tokenized before it is ever written, so the record is defensible and the raw identifier is never in it.

The layer

Five modules on one spine

Not another model or dashboard: the accountable record beneath the agents you run. Each module writes to, and reads from, the same chain.

Register, Know Your Agent

A live inventory of every agent and model: owner, autonomy tier, provenance, and a cryptographically verifiable identity. Nothing runs unregistered.

Witness, the flight recorder

Every decision hashed into the chain with the model identity, inputs (tokenized), and outcome, independently recomputable by anyone.

Bound, autonomy envelopes

Each action routes to auto, human review, or block by stakes and confidence, with oversight telemetry, formal envelope linting, and a drilled kill switch.

Vidhi & Vishwas, evidence & fairness

Regulator instruments encoded as data and scored against live evidence, plus declared-first fairness screens gated on statistical significance so they do not cry wolf.

How agents connect

Three ways in, no rewrites

Wrap your calls with the SDK in a few lines, put an un-instrumentable agent behind the zero-code gateway, or add the MCP server to any MCP host. Every path lands on the same chain, and Sakshi never sees a raw identifier.

Developer docs & API reference arrow_forward

Where it runs

Single-tenant, in your environment

Sakshi installs in your VPC or on-prem and stays there. We never see your evidence and could not hand it over if compelled. Data residency is India by default, and the deployment model is the security model.

Enterprise deployment arrow_forward

See the whole path on your own screen

Spin up a synthetic-data sandbox, register an agent, and recompute the chain yourself.