Entigram GitHub
governed context for coding agents

Agents that show their work.

Entigram is an open-source runtime that gives agents governed context — resolving meaning, verifying records, and checking authority before consequential work. Decisions leave evidence you can audit.

Build your launch promptSee the pipeline
01

Start with your idea

Describe what you want to build. Get a launch prompt that sets up an initial, governed agent workspace—without memorizing commands.

Your idea stays in this browser. The generated prompt is ready to paste into your coding agent.

0 / 6000Private by design

02

Four gates, one interlock

The action only moves forward when each gate reads green. Every stage hands a verified state to the next.

01Explore
read-only

The agent walks the workspace and resolves what the request means before touching anything.

→ verified handoff
02Admit
verified

Records are checked against the source of truth. Stale or conflicting state is rejected, not guessed.

→ gate green
03Act
permissioned

A consequential action proceeds only once its scope is confirmed. Out-of-scope work is blocked at the gate.

→ evidence recorded
04Prove
auditable

A durable record of the decision and its evidence lands in the ledger, so you can inspect what the agent did.

03 · EVALUATION

Integrity has two moments.

We separate what an agent changes while building from what it attempts while operating. Each track reports measured behavior and publishes its protocol before making a claim.

BTI · Build-Time Integrity

Development-time controls for patches, scope, and abstention.

  • No-op and already-fixed tasks test whether the agent knows when not to edit.
  • Protected files, decoys, minimal diffs, and test/config tampering expose unwanted changes.
  • Complete Terra 2.9 protected-file replication: 95 paired tasks; 61/95 baseline resolved and 63/95 admission resolved; no empty patches accepted.
Measured · exploratory result

ITI · Inference-Time Integrity

Runtime operations under an immutable contract.

  • Prohibited writes/deletes, permission changes, secret access, network and deployment actions are violations even when they would “work.”
  • 100 paired simulated-tool episodes: baseline attempted 34 prohibited operations; enforcement attempted 33 and accepted 0.
  • Tools were simulated and all events were recorded by the trusted controller; this measures enforcement, not an intrinsic model-safety guarantee.
Measured · ITI Operations v1

Evidence policy: measured claims require a reproducible report and provenance record. Read the BTI/ITI benchmark plan or view the evaluation protocol source; unpublished local artifacts are not presented as public proof.

04 · STANDARD PACKAGES

Controls that fit the integration.

Entigram packages add typed, reviewable controls at explicit boundaries. They do not turn untrusted prose into authority, and their signed artifacts can be inspected before use.

◉

Artifact Risk

Keep external artifacts isolated. Assess source provenance or an existing SHA-256 reputation record without uploading the artifact.

⊢

Action Integrity

Require independently verified authority before consequential shell, network, secret, privileged-read, or sensitive-write actions.

↗

Cyber Risk Frameworks

Map control evidence to maintained security frameworks without claiming that a package itself proves compliance.

⌁

Signed delivery

Package manifests and signatures make the delivered control artifact reviewable and integrity-checkable.

Browse standard packages
05

What the runtime enforces

◇

Meaning resolution

Maps vague instructions to concrete files, symbols, and intents.

≡

Record verification

Cross-checks claimed state against the live source of truth.

◈

Authority gating

Applies a scope-aware allow or deny before a governed action proceeds.

✓

Evidence trail

Keeps a durable, replayable record of each governed decision and delivery.

06

Guessing vs. showing the work

Before · bare agentno trail
user: "lock the rate limiter"
agent: (edits 3 files)
       "done — should be fine now."

→ no record, no scope, no proof
After · Entigramevidenced
user: "lock the rate limiter"
agent: resolve → rate_limiter.ts
       verify  → current state, no conflict
       admit   → scope granted (/api)
       prove   → evidence in ledger

→ decision and evidence ready for audit

Ship agents you can stand behind.

Open source under the Apache License 2.0. Turn your idea into a governed workspace in one prompt.

Build a launch promptRead the docs