<div align="center">

<img alt="GraQle — persistent organisational intelligence for AI agents." src="https://raw.githubusercontent.com/quantamixsol/graqle/master/assets/hero-dark-hq.png" width="800">

GraQle

Give your AI agents a memory of how your organisation works - and check every decision before they act.

]()

pip install graqle

Website · Quickstart · Decision Assurance · Shared memory · Changelog

<!-- mcp-name: io.github.quantamixsol/graqle -->

</div>

What GraQle does

1. Remembers. Turns your code, documents, policies and decisions into one connected knowledge graph that lasts beyond a chat session.
2. Checks. Before an agent acts, the Decision Assurance Gate checks the action and answers EXECUTE, REPLAN, HOLD, REJECT or ESCALATE - with the reason and a signed record.
3. Shares. One graph serves a whole team of agents - Claude Code, Cursor, Codex, Copilot - each with its own permissions and its own files.

Works with any model: Anthropic, OpenAI, AWS Bedrock, Ollama, Gemini and more - or fully offline.

Quickstart

pip install graqle

graq scan docs ./policies          # documents, policies, ADRs -> graph
graq scan repo .                   # code -> the same graph
graq learn knowledge "vendor DPA must be signed before any data access"
graq run "what approval is needed for a large refund?"
# -> answer + confidence + the exact sources it used

Connect it to your AI tool (Claude Code, Cursor, Windsurf, VS Code):

graq init        # detects your IDE and wires GraQle in

PDF, DOCX, PPTX and XLSX: pip install "graqle[docs]".

Check every decision

Most AI tools give you one confidence number. GraQle asks separate questions before an agent acts - and a good answer to one never hides a failed answer to another:

• Allowed? Policy valid, actor authorised, evidence current, arguments exactly the approved ones. Any "no" stops the action.
• Consistent? Contradicting facts are tracked in the graph, and a change updates every decision that depends on it.
• Sure enough? Confidence is measured per dimension, not blended into one score.
• On track? Unreliable step sequences are sent back to replan within a retry budget.
• Who approves? Authority is checked for the action's impact level.

from graqle.assurance.gate import DecisionAssuranceGate

verdict = gate.evaluate(candidate, ctx)
verdict.outcome         # EXECUTE | REPLAN | HOLD | REJECT | ESCALATE
verdict.decisive_rule   # which rule decided
verdict.reason_codes    # why

Every decision leaves a sealed record: a signed, hash-chained provenance event, exportable to W3C PROV-O and checkable with SHACL, so an auditor can verify it with standard tools. The gate is off by default: installing or upgrading GraQle changes nothing until you switch it on.

Turn it on, per project:

1. Set GRAQLE_DAG_ENABLED=true in the environment where GraQle runs (or in your .env).
2. Provide the required settings. GraQle ships no thresholds, salts or keys: you choose them. Start from the GRAQLE_DAG_* section of .env.example and fill in every placeholder (salt, thresholds, signing key id and version, calibrator version, and the others listed there). Put the values in the environment, or in a private file at GRAQLE_DAG_SECRETS_PATH (default .graqle/graqle_secrets.yaml); the environment wins.
3. Start as usual. If a required value is missing, GraQle raises a configuration error that names it the first time the gate settings load. There is no silent fallback.

Check it: python -c "from graqle.assurance.settings import is_dag_enabled; print(is_dag_enabled())" prints True.

To switch it off again, unset the flag or set it to false. Details: operating the gate.

One memory for a team of agents

Several agents on one codebase usually means several copies of the same graph in memory. GraQle can hold it once:

# where the graph lives
graq mcp serve --transport http --port 8765 --config graqle.yaml --allowed-roots /path/to/projects

# in each agent session (.mcp.json)
{ "mcpServers": { "graqle": { "command": "graq", "args": ["mcp", "relay", "--url", "http://127.0.0.1:8765"] } } }

• Less memory. Each session runs a small relay instead of a full copy (measured on a 20,000-node graph: 3 separate servers 687 MB, shared server + 3 relays 493 MB).
• Learning is shared at once across all sessions.
• Each session keeps its own gates and its own files - file work happens on each user's own machine.
• Teammates connect over an SSH tunnel, or HTTPS with a token.
• Several projects, separate graphs on one server (Team and Enterprise).

Turn on the decision gate (opt-in, baseline tier)

The decision gate is off unless GRAQLE_DAG_ENABLED is set; with it off nothing changes. Turn it on:

1. Create an ed25519 signing key (PEM, mode 0600) and a key manifest that lists its public key

(graqle.assurance.key_manifest_store.save_manifest), plus a coverage policy file.

1. Set GRAQLE_DAG_ENABLED=1 and the required GRAQLE_DAG_* settings: SIGNING_KEY_ID,

SIGNING_KEY_VERSION, SIGNING_KEY_PATH, COVERAGE_POLICY_PATH, CONFIG_SALT, CAP_VALUE, HG05_POISONING_THRESHOLD, HG07_MATERIALITY_THRESHOLD, CALIBRATOR_VERSION, TRAJECTORY_ESTIMATOR, MAX_RETRY_BUDGET, CIRCUIT_BREAKER_WINDOW, SCOPE_KEY_FIELDS_REF, MERGE_PRIORITY_REF. There are no built-in values for these; you choose them.

1. Register the actors that may take MEDIUM-risk calls (for example GRAQLE_RBAC_ACTORS_JSON). An

actor with no registered role is closed by hard gate HG-06.

1. Start your process. The packaged default runtime provider is registered at start-up. If it

cannot be built the process fails at start-up with an error naming register_runtime; supply your own provider with graqle.assurance.adapter.register_runtime(provider) if you need one.

Baseline-tier decisions are governed by rule DAG-RT-BASELINE-ADMIT. No projection score, reliability or residual risk is computed or recorded for them. EXECUTE only when risk is LOW, impact is zero and all hard gates pass; anything else is HOLD (DAG-RT-NO-EVIDENCE) or, when a hard gate fails, REJECT. MEDIUM and above routing requires explicit configuration. Every decision is written as a signed, hash-chained provenance event; if it cannot be written the call is not executed.

This is a policy-scoped authorisation at evidence tier BASELINE. It is not informed by the graph and is not a safety verification. Evidence-fed scoring arrives in 0.88 (CR-DAG-088-1). python scripts/dag_wheel_e2e.py runs these steps against the built wheel in a clean virtualenv.

Trust

• Local-first, no telemetry. The graph is a file you control; nothing phones home.
• Secure by default. The server needs authentication unless you explicitly switch on local development mode; graph uploads can never silently overwrite each other.
• Signed releases. Sigstore-signed wheels, CycloneDX SBOM, reproducible builds.
• Regulated work. Audit trail and evidence exports for your own EU AI Act file - GraQle provides evidence, it does not claim to make you compliant. Details

Report vulnerabilities to [email protected] · SECURITY.md

Editions

Recent releases

• v0.87.0 — Decision Assurance Gate: five-answer decisions, contradiction tracking, confidence per dimension, replanning, signed provenance records.
• v0.86.0 — One memory for a team of agents: shared server, thin relay, per-session gates, file work on each user's machine; secure-by-default server.

Full changelog

Patent & license

Core methods are patent-pending: EP26167849.4, EP26162901.8 and EP26166054.2. Source is fully auditable under the GraQle License - see LICENSE.

<div align="center">

Built by [Quantamix Solutions](https://quantamixsolutions.com) · github.com/quantamixsol/graqle

</div>