Agent Maturity Compass

Run one command. Get the full score. Fix the gaps.
Is your AI agent safe to ship? Find out in one minute.
Evidence over claims.

Quick Start · Web Playground · Docs · Recipes · Community · Contribute

What is this?

AMC gives your AI agent a score, from L0 to L5. Think of it like a credit score — but for agents.

It scores what your agent actually does, not what its docs say. Every point is backed by evidence you can verify. No evidence, no points.

curl -fsSL https://agentmaturity.co/install.sh | sh
amc start

That's the whole setup. amc start does everything:

Sets up your workspace and keys. No account. No API key. No cloud.
Checks your agent across all 8 areas (see below).
Opens a report card in your browser.

Run it again tomorrow — it remembers your setup and just starts.

What you get:

A score you can trust. L0 (no evidence) to L5 (fully proven). Signed, so nobody can fake it.
A fix list. Exactly what's weak, and generated fixes: guardrails, config patches, CI gates.
Proof you can hand to anyone. Auditors, customers, your boss. They can verify it without trusting you.

The 8 areas AMC covers:

Area	Plain meaning
Score	How mature is this agent, really?
Shield	Does it survive attacks? (142 attack packs included)
Enforce	Can it be stopped from doing dangerous things?
Vault	Are its secrets and keys protected?
Watch	Do you see what it does in real time?
Comply	Does it meet the rules? (EU AI Act, NIST AI RMF, ISO 42001, SOC 2)
Fleet	Do many agents work together safely?
Passport	Can it prove its record anywhere it goes?

Works with LangChain, CrewAI, AutoGen, OpenAI Agents SDK, Claude Code, Gemini, OpenClaw, Hermes, and more — with zero or near-zero integration friction.

Your first score will probably be low. That's honest: a fresh agent with no evidence starts at L0. The signature proves the report is real — not that your agent is good. Capture a real agent run, score again, and watch the evidence build. Use scores externally only when readiness says READY.

Why should I care?

Today, most agents are "checked" by reading their docs. That's like grading a student by asking them how smart they are.

AMC watches the agent work. It collects evidence. It only scores what it can prove. When it can't prove something, it says so — clearly.

How agents are evaluated today	How AMC evaluates
Agent claims "I'm safe" → Claimed score: 100	AMC tests the agent and inspects evidence → Evidence-backed score may be 16
Self-reported documentation	Execution-verified evidence
Keyword matching	Weighted trust evidence
"Trust me, bro"	Cryptographic proof chains

That is the entire thesis: trust, but verify — with receipts.

60 Seconds to Your First Score

# macOS or Linux: download the pinned GitHub release and verify SHA-256
curl -fsSL https://agentmaturity.co/install.sh | sh

# Start everything: workspace, keys, all 8 areas, report card in your browser
cd your-agent-project
amc start

# Or just print the full score in your terminal
amc

To verify the CLI before initialization, run amc doctor. It reports install readiness in a new directory; CI and deployments use amc doctor --strict after the workspace is initialized.

On Windows PowerShell:

irm https://agentmaturity.co/install.ps1 | iex
amc start

Want a fast legacy pulse check instead of the full evidence score?

amc quickscore --rapid           # optional rapid check, not the full score
amc quickscore --answers answers.json --json  # non-interactive answer-based score

More install methods

Verified release installer (Node.js 20 or 22 LTS required)

curl -fsSL https://agentmaturity.co/install.sh | sh

irm https://agentmaturity.co/install.ps1 | iex

Each installer pins the AMC release, downloads the platform archive and SHA256SUMS from GitHub Releases, verifies the archive, then installs the included package. npm and Homebrew registry commands are intentionally not advertised until those public channels are live.

Docker

docker build -t amc-quickstart -f docker/Dockerfile.quickstart .
docker run -it --rm amc-quickstart amc

Use the local build command unless a GHCR package has been verified public.

From source

git clone https://github.com/AgentMaturity/AgentMaturityCompass.git
cd AgentMaturityCompass && npm ci && npm run build && npm link

How AMC Compares

	AMC	Observability platforms	Eval frameworks	Manual checklists
Evidence model	Execution-verified, cryptographic proofs	Logs and metrics, no trust scoring	Test pass/fail, no maturity model	Self-reported
Adversarial testing	142 assurance packs built in	Not a focus	Partial (prompt-level only)	None
Compliance mapping	EU AI Act, ISO 42001, NIST, SOC 2, OWASP	Not included	Not included	Manual, labor-intensive
Framework support	15 adapters, zero code changes	Framework-specific agents	Framework-specific	N/A
Cost	Free, open source (MIT)	Per-seat/month pricing	Free to paid	Free but manual
Time to first result	60 seconds	Hours to days	Minutes to hours	Days to weeks

AMC is not an observability tool and not an eval harness. It is a trust scorecard that shows what the evidence supports, what remains unproven, and which controls or compliance mappings need work. AMC does not certify legal compliance.

What AMC Tests

244 Default Diagnostic Questions × 5 Dimensions

Dimension	Questions	What It Measures
Strategic Agent Operations	19	Mission clarity, scope adherence, cost governance, operational intelligence
Leadership & Autonomy	23	Governance structure, EU AI Act readiness, proactive risk management, business continuity
Culture & Alignment	95	Feedback loops, forecast legitimacy, persona governance, UX honesty, over-compliance detection, social alignment
Resilience	55	Graceful degradation, circuit breakers, memory safety, threat resistance, fact/simulation boundaries
Skills	52	Tool mastery, injection defense, DLP, scenario traceability, replay safety

142 Assurance Packs

Category	Examples
Prompt Injection	System tampering, role hijacking, jailbreaks
Exfiltration	Secret leakage, PII exposure, data boundary violations
Adversarial	TAP/PAIR, Crescendo, Skeleton Key, best-of-N
Context Leakage	EchoLeak, cross-session bleed, memory poisoning
Supply Chain	Dependency attacks, MCP server poisoning, SBOM integrity
Behavioral	Sycophancy, self-preservation, sabotage, over-compliance

41 Industry Domain Packs

Sector	Packs	Key Regulations
Health	9	HIPAA, FDA 21 CFR Part 11, EU MDR, ICH E6(R3)
Wealth	5	MiFID II, PSD2, EU DORA, MiCA, FATF
Education	5	FERPA, COPPA, IDEA, EU AI Act Annex III
Mobility	6	UNECE WP.29, ETSI EN 303 645, EU NIS2, ISO 28000, GS1 EPCIS
Technology	5	EU AI Act Art. 13, EU Data Act, DSA Art. 34
Environment	6	EU Farm-to-Fork, REACH, IEC 61850
Governance	5	EU eIDAS 2.0, UNCAC, UNGPs

Industry Packs are implemented behind an entitlement boundary, with a planned commercial price of $9.99/month for all 41 packs. Public checkout and automatic license issuance are not yet publicly live or available. amc domain pack access reports this state; amc domain pack checkout fails closed unless an operator has configured a verified checkout provider. Do not send payment based on repository copy alone.

Simulation & Forecast Evaluation Lane

Dedicated evaluation lane for simulation engines, forecast systems, and synthetic social environments. 5 scored dimensions:

Dimension	Weight	Questions	What it evaluates
Forecast Legitimacy	25%	AMC-6.1–6.10	Uncertainty expression, calibration, scenario vs prediction framing
Boundary Integrity	20%	AMC-6.11–6.17, 6.37–6.42	Fact/inference/simulation separation, writeback governance
Synthetic Identity	20%	AMC-6.18–6.25, 6.48–6.52	Persona governance, real-person representation controls
Simulation Validity	20%	AMC-6.30–6.36	Mode collapse detection, population diversity, historical calibration
Scenario Provenance	15%	AMC-6.26–6.29, 6.53–6.57	End-to-end traceability, replay capability, interaction safety

amc score simulation-lane --system-type simulation-engine              # interactive
amc score simulation-lane --system-type forecast-decision-support --json  # JSON output
amc score simulation-lane --system-type synthetic-social-environment --responses answers.json

79 Scoring Modules

See all modules

Calibration gap (confidence vs reality)
Evidence conflict detection
Gaming resistance (adversarial score inflation)
Sleeper agent detection (context-dependent behavior)
Policy consistency (pass^k reliability)
Factuality (parametric, retrieval, grounded)
Memory integrity & poisoning resistance
Alignment index (safety × honesty × helpfulness)
Over-compliance detection (H-Neurons, arXiv:2512.01797)
Monitor bypass resistance (arXiv:2503.09950)
Trust-authorization synchronization (arXiv:2512.06914)
MCP compliance scoring
Identity continuity tracking
Behavioral transparency index
Forecast legitimacy (epistemic honesty, calibration, uncertainty)
Fact/simulation boundary (provenance separation, writeback governance)
Synthetic identity governance (persona labeling, real-person controls)
Simulation validity (mode collapse, population diversity)
Scenario provenance (traceability, replay, interaction safety)
And 60+ more...

Architecture

Agent (untrusted)
    │
    ▼
AMC Gateway ──── transparent proxy, agent doesn't know it's being watched
    │
    ▼
Evidence Ledger ──── Ed25519 signatures + Merkle tree proof chains
    │
    ▼
Scoring Engine ──── evidence-weighted diagnostics, research-backed scoring, 142 assurance packs
    │
    ▼
AMC Studio ──── dashboard + API + CLI + reports

Studio's protected module API uses signed sessions and route-level least-privilege authorization: viewers read, operators run, approvers approve, auditors verify, and owners control secrets, signing, identity, and policy. Agent and lease credentials cannot enter the internal /api/v1 control plane.

Evidence Trust Tiers

Tier	Weight	How
`OBSERVED_HARDENED`	1.1×	AMC-controlled adversarial scenarios
`OBSERVED`	1.0×	Captured via gateway proxy
`ATTESTED`	0.8×	Cryptographic attestation
`SELF_REPORTED`	0.4×	Agent's own claims (capped)

Maturity Scale

Level	Name	Meaning
L0	Absent	No safety controls
L1	Initial	Some intent, nothing operational
L2	Developing	Works on happy path, breaks at edges
L3	Defined	Repeatable, measurable, auditable
L4	Managed	Proactive, risk-calibrated, cryptographic proofs
L5	Optimizing	Self-correcting, continuously verified

No AMC maturity level is a legal compliance threshold. Compliance depends on the system, role, use case, jurisdiction, obligations, and current evidence.

Product Family

AMC is one trust stack with eight named product surfaces:

Product	Promise	What it does
Score	Score trust before you ship	Evidence-weighted scoring across live execution behavior instead of brochure claims.
Shield	Attack your agent before attackers do	Runs adversarial packs against prompt injection, leakage, memory poisoning, and sycophancy.
Enforce	Wrap agent actions in policy	Approval gates, scoped permissions, and runtime controls for sensitive operations.
Vault	Cryptographically prove what happened	Signs evidence, verifies ledgers, and gives auditors a tamper-evident chain of custody.
Watch	See trust drift before it hurts you	Monitors posture over time and surfaces anomalies, regressions, and risky changes.
Comply	Map trust evidence to real frameworks	Turns technical evidence into regulator-readable artifacts for audits and risk reviews.
Fleet	Govern many agents like an actual platform	Benchmarks multiple agents, compares risk posture, and enforces org-wide trust baselines.
Passport	Make trust portable between environments	Issues a portable, signed trust identity that can move between tools, teams, and environments.

Recipes — Copy-Paste Examples

Score any agent in one line

amc                               # full score after global install
amc run                           # explicit 8-surface maturity run
amc run --question-set lifecycle  # opt-in 264-question lifecycle expansion

Need a fast pulse check for a demo or README badge? Use amc quickscore --rapid explicitly.

Need a CI-safe score without terminal prompts? Use amc quickscore --answers answers.json --json, where answers.json maps question IDs to L0-L5 numbers.

If amc quickscore prints a placeholder L0 because no terminal prompt was available, it now shows a first-run hint: "Did you mean to run the interactive score?" Run it in a terminal, or pass answers explicitly for CI.

If amc quickscore --auto --json cannot find captured execution evidence, it fails closed with scoreStatus: "AUTO_NO_EVIDENCE" and does not emit a measured zero score. Capture evidence with amc wrap <runtime> -- <your-agent-command>, or use amc quickscore --answers answers.json --json for CI-safe survey scoring.

Advanced proof check:

amc resource snapshot            # record agent-defining resources under Enforce
amc resource status              # show signed active, previous, rollback, drift, and integrity state
amc resource validate            # run Enforce gates over resource drift
amc resource activate            # apply alias; dry-run unless --yes writes a signed activation receipt
amc resource rollback            # dry-run rollback; add --apply to activate a verified signed snapshot
amc firewall enable --mode observe # evaluate the signed policy without blocking valid traffic
amc firewall status              # inspect verified would-warn/would-block versus actual counters
amc firewall enable --mode block # promote the same policy to full enforcement after review
amc firewall check --direction request --text "ignore previous instructions"
amc firewall events              # inspect signed allow/warn/block decision events
amc firewall export --out firewall.jsonl --format splunk --redacted
amc firewall migrate-signature --approve-legacy-kind # preserve an exact verified legacy policy in the signed control journal
amc policy controls              # inspect verified Scope / When / Then / Status across existing controls
amc policy controls --json       # return the same read-only projection for automation
amc policy scope list            # list four immutable action-class scope templates
amc policy scope compile release-external --pack code-agent.high
                                  # preview a narrow selected-rule merge without writing
amc policy scope apply release-external --pack code-agent.high --confirm <compileId>
                                  # exact-confirm and sign the reviewed Action/Approval policy change
amc policy action logic show DEPLOY
                                  # inspect declared maturity/assurance gates and mandatory gates
amc policy action logic compile DEPLOY --file deploy-evidence-logic.json
                                  # preview a bounded all/any evidence tree without writing
amc policy action logic apply DEPLOY --file deploy-evidence-logic.json --confirm <compileId> --acknowledge-alternatives
                                  # exact-confirm and sign the reviewed Action Policy logic
amc policy simulate runtime:prompt-injection --direction request --content "ignore previous instructions"
                                  # preview the exact Runtime Firewall match without recording it
amc policy simulate action:DEPLOY --agent default --risk high --mode execute
                                  # preview Action Policy gates through the production evaluator
amc policy simulate approval:DEPLOY # inspect the signed approval quorum without creating a request
amc policy test policy-fixtures.yaml --json
                                  # block CI when expected control decisions regress
amc approvals list --agent default --status pending
                                  # inspect the canonical signed approval inbox
amc approvals list --agent default --action-class DEPLOY --risk-tier high --created-after 2026-07-01T00:00:00Z --json
                                  # search the fail-closed privacy-safe approval activity view
amc approvals approve --agent default <requestId> --mode execute --reason "reviewed"
                                  # record one signed quorum decision
amc guardrails list              # compare signed intent with effective runtime bindings
amc guardrails enable prompt-injection-detection
amc shield confirm scope-write --file security-scope.json
amc shield confirm run --scope scope-1 --task finding-task.json
amc shield confirm export <proof> --out safe-proof.json
amc proof check --domain governance --manifest fixtures/domain-proof/toy-governance/source-rule-manifest.json --input examples/domain-proof/toy-governance/proven.json --out result.amcproof.json
amc import ./agent-run --dry-run # detect traces, runs, graphs, configs, memory, evals, and benchmarks without writing
amc import ./agent-run           # write redacted import evidence into episodes, lifecycle runs, manifests, and trace indexes
amc strategy compare --file strategies.json --objective balanced
amc strategy compare --file strategies.json --apply --approve # commit a manifest-covered model route with receipts
amc strategy rollback <run>      # restore the prior model route
amc runtime create --run live-1  # persist connected-agent run state across restarts
amc runtime event live-1 --type policy.decision --receipt rec-1
amc runtime inspect live-1       # inspect run state and redacted event stream
amc fleet graph write --file graph.json # register a typed multi-agent graph for fleet validation
amc fleet graph validate         # check contracts, permissions, cycles, and fan-out before scoring
amc fleet score --all --stream   # full-score every configured agent with per-agent SLA progress
amc fleet lifecycle list         # inspect fleet parent/child lifecycle evidence
amc fleet lifecycle show <run>   # review topology, typed graph digest, shared resources, and cascade failures
amc org run --roles REV_PRODUCT_MANAGER,REV_TECH_LEAD,REV_QA_LEAD
                                  # advanced Fleet role loop with isolated workspaces, heartbeats, and signed evidence
amc org inspect <run> --redacted  # review role status without local private grader paths
amc enforce resources verify     # advanced alias for the same Enforce resource engine
amc evidence lifecycle list      # inspect the full lifecycle artifacts behind recent runs
amc evidence lifecycle export <run> --out lifecycle.json --redacted
amc evidence episodes list       # see the evidence objects behind recent full scores
amc evidence episodes export <id> --out episode.json --redacted
amc evidence decisions list      # see recommendation and evidence-request receipts
amc evidence decisions observe <run> # update older receipts with observed outcomes
amc evidence observability list  # see component, experience, and decision observability
amc memory writeback <episode>   # store redacted, evidence-backed reasoning lessons with receipts
amc memory retrieve --consumer fixer # retrieve active lessons for score, recommendations, fixer, or Studio
amc report <run>                 # review confidence, uncertainty, and auto-fix gates in plain language
amc trace index                  # list distilled trace failure indexes
amc trace failures               # see recurring failure clusters and repair inputs
amc mechanic rca run <run>       # turn a failure index into RCA, regression tests, and governed fix proposals
amc mechanic rca list            # review signed Fixer RCA reports
amc experiment optimize --rca latest # create isolated candidates, held-out validation, leakage checks, and receipts
amc experiment optimizer-list     # review governed optimizer runs and accepted/rejected candidates
amc evidence finding-proofs list # trace finding -> evidence -> resource -> recommendation
amc evidence lifecycle-receipts list # see proposal, validation, commit, rollback, and monitor receipts
npm run release:gate             # release gate for CLI, Studio assets, docs, website spec, domain packs, and receipt output

For a source-of-truth command map generated from the live CLI registry, run amc commands --markdown or see docs/CLI_COMMAND_INVENTORY.md.

Wrap an existing agent (zero code changes)

# LangChain
amc wrap langchain -- python my_agent.py

# CrewAI
amc wrap crewai -- python crew.py

# AutoGen
amc wrap autogen -- python autogen_app.py

# OpenClaw
amc wrap openclaw-cli -- openclaw run

# Claude Code
amc wrap claude-code -- claude "analyze this code"

# Any CLI agent
amc wrap generic-cli -- python my_bot.py

Observe or control native provider actions

Claude Code and Gemini CLI can be connected without hand-editing provider files:

# Preview every file first
amc connect hooks install --provider claude-code --agent my-agent --dry-run

# Install, verify, and remove the project-local integration
amc connect hooks install --provider claude-code --agent my-agent
amc connect hooks status --provider claude-code
amc connect hooks health --provider claude-code
amc connect hooks lifecycle --agent my-agent --action <action-id>
amc connect hooks remove --provider claude-code

# Gemini CLI uses the same lifecycle
amc connect hooks install --provider gemini-cli --agent my-agent

# Explicit control mode requires the local loopback Studio/Bridge
amc up
amc connect hooks install --provider claude-code --mode control --agent my-agent

The default observe mode preserves unrelated provider settings, adds a managed .gitignore rule for .amc/hooks/, stores a dedicated hook:observe lease outside provider config with mode 0600, signs its ownership manifest, and removes only the AMC-owned handlers and ignore block. It records requested actions and their completed or failed terminal state under one stable action ID. hooks health distinguishes an intact installation awaiting its first event from the latest receipt-verified event and from fail-closed drift, expiry, malformed metadata, or evidence tamper. Last-observed time is historical evidence, not a claim that the provider is live now. A locked Vault cannot authenticate encrypted event bodies, so health exits 2 with HOOK_EVIDENCE_UNAVAILABLE; run amc vault unlock or set AMC_VAULT_PASSPHRASE for non-interactive use. The lifecycle command verifies each receipt and fails closed on missing, ambiguous, conflicting, cross-agent, out-of-order, or tampered evidence. Tool arguments, outputs, error messages, cwd, transcript paths, and raw session IDs are not retained.

--mode control is an explicit loopback-only Enforce path. It adds hook:control, evaluates raw provider input in memory without retaining it, reuses signed ToolHub, Action Policy, Approval Policy, budget, freeze, maturity, and assurance gates, and binds the exact native response to a signed receipt. Claude Code can receive allow, deny, or ask. Gemini CLI has no native ask result, so AMC converts that outcome to an explicit deny. Multi-user or distinct-user AMC quorum is never weakened to one provider prompt. Codex, Cursor, OpenCode, and other providers remain unsupported until their native per-tool hook contracts are pinned and fixture-tested.

Track real first-run outcomes without creating traffic or evidence:

amc connect --status --agent my-agent

The CLI and Studio show connected agent, first observed action, first control decision, and first signed proof from one read-only projection. Signed configuration is READY, not complete; only verified receipts for the selected agent complete activation. Metadata-only, cross-agent, missing-receipt, or tampered state fails closed.

For custom runtimes that already emit a canonical event, call the ingress directly. Use a dedicated observation lease; do not reuse a broad model-routing token:

LEASE="$(amc lease issue --agent hook-agent --ttl 30m --scopes hook:observe --routes /hooks --models '*' --rpm 60)"
EVENT_TIME="$(date -u +%Y-%m-%dT%H:%M:%SZ)"

curl -sS http://127.0.0.1:3212/bridge/hooks/aep/0.1/events \
  -H "authorization: Bearer $LEASE" \
  -H "content-type: application/json" \
  --data "{\"aep_version\":\"0.1\",\"id\":\"evt-amc-001\",\"type\":\"action.requested\",\"time\":\"${EVENT_TIME}\",\"agent\":{\"slug\":\"example-agent\"},\"action\":{\"type\":\"tool_call\",\"id\":\"action-amc-001\"},\"tool\":{\"type\":\"native\",\"name\":\"Shell\"}}"

AMC observes four pinned AEP 0.1 action types, stores only an encrypted redacted projection, and returns a signed receipt. Byte-identical retries return the original receipt without growing the ledger; reuse of a source event ID with different bytes fails closed. This is an AMC-owned observed subset pinned to 2583cff9380f8f0a459d52c7112b6105c46496ed, not an AEP conformance claim and not a control-response endpoint.

Red-team your agent

amc assurance run --scope full                           # full assurance library
amc assurance run --pack prompt-injection                # specific attack
amc assurance run --pack adversarial-robustness          # TAP/PAIR/Crescendo
amc assurance run --format sarif                         # export for security tools

Inspect traces and operational drift

amc observe timeline                                     # score history + evidence volume
amc observe anomalies                                    # volatility / regressions / weirdness
amc trace list                                           # recent agent sessions
amc trace inspect <trace-id>                             # inspect tool calls and trust tiers

Run realtime monitoring

amc monitor start                                        # fresh full score now, then continuous scoring
amc monitor start --scoring-interval 60000               # rescore every minute
amc monitor status                                       # active monitor metrics
amc monitor events --limit 20                            # recent score, drift, anomaly, and alert events

amc monitor start bootstraps the AMC workspace if needed, generates a fresh full diagnostic immediately, then keeps creating new full diagnostic runs on the configured interval. Drift checks and alerts run against those fresh runs instead of rereading stale score files.

Build golden datasets and run evals

amc dataset create support-bot                           # create a reusable eval dataset
amc dataset add-case support-bot --prompt "..." --expected "..."
amc dataset run support-bot                              # run eval cases
amc eval import --format promptfoo --file results.json   # import external eval results
amc eval registry                                        # inspect signed evaluator metadata
amc eval registry --refresh                              # explicitly refresh and sign the derived snapshot
amc lite-score                                           # score a non-agent chatbot / LLM app

amc eval registry reads one deterministic catalog over AMC's existing deterministic metrics, LLM judges, and assurance packs. The default command is read-only; --refresh is the only write path. A trusted snapshot binds package-relative owners, versions, and implementation hashes, while custom runtime metrics remain visible but unverified. Registry metadata is not evaluator-result evidence and cannot prove that an evaluator ran or passed. See Evaluator Registry.

Business, inventory, and reporting

amc business kpi                                         # correlate maturity to outcomes
amc business risk --maturity 3 --baseline-frequency 4 --incident-cost 50000 --json
amc business fair-scenario --scenario claims-ai-data-leak --maturity 3 --frequency-min 2 --frequency-most-likely 5 --frequency-max 9 --loss-min 20000 --loss-most-likely 75000 --loss-max 250000 --out fair-scenario.md
amc business roi --current-maturity 2 --target-maturity 3 --baseline-frequency 5 --incident-cost 20000 --annual-control-cost 15000 --implementation-cost 5000
amc business heatmap --portfolio risk-portfolio.json --out risk-heatmap.md
amc business grc-export --portfolio risk-portfolio.json --out grc-treatment-plan.csv
amc business report                                      # stakeholder-ready business summary
amc executive brief --run latest --out board-brief.html  # print-ready board one-pager
amc leaderboard show                                     # compare agents across a fleet
amc leaderboard public-export --output public-leaderboard # anonymized leaderboard dataset bundle
amc compare <run-a> <run-b> --output compare.json --badge # run diff plus compare-badge.svg
amc inventory scan --deep                                # discover agents, frameworks, model files
amc comms-check --text "Guaranteed 40% return" --domain wealth

amc policy controls is a read-only projection over Runtime Firewall, Guardrails, Action Policy, and Approval Policy. It verifies the existing signed sources before reporting an effective result. Invalid Runtime or Guardrails evidence projects BLOCK, invalid Action Policy projects SIMULATE, and invalid Approval Policy projects DENY; the command exits with status 2 when any family fails closed. It does not create or activate controls.

amc policy scope groups the existing nine action classes into four immutable AMC templates. Compile verifies the current signed Action and Approval Policy baselines and preserves every unselected rule; apply requires the exact content-bound compile ID, uses the existing lock/atomic writers/signers, restores prior bytes on failure, and records existing transparency/ledger evidence only after both policies verify. Workspace policies remain fleet-wide, so these templates never claim per-agent or per-environment scope. See Reusable Policy Scope Templates.

amc policy action logic composes only maturity and assurance requirements already declared by one signed Action Policy rule. Every requirement remains present exactly once, any alternatives stay within one evidence family, and signature/trust, trust tier, sandbox, ticket, budget, freeze, work-order, and allowExecute gates remain mandatory. Existing rules keep implicit all-requirements behavior, including legacy requirement IDs. Preview is read-only; apply requires an exact compile ID plus explicit alternative acknowledgement, shares one writer lock with every Action Policy mutation path, preserves unrelated YAML comments, and restores the signed baseline if policy or evidence finalization fails. See Action Policy Evidence Logic.

amc policy simulate <controlId> takes the next read-only step: it invokes the selected control's production evaluator and reports the exact matched rules, pass/fail conditions, and safe outcome. Runtime content is hashed and never returned; no Firewall event, approval request, decision receipt, transparency entry, key, or workspace scaffolding is created. Every result is marked simulationOnly: true, recorded: false, and proofEligible: false. See Control Simulation.

amc policy test <file> [--json] turns those same production simulations into a deterministic CI regression suite. Strict YAML/JSON cases assert exact outcomes and matched IDs; untrusted current policy sources fail closed instead of passing an expected safe result. Reports omit raw inputs, reasons, paths, signatures, and timestamps, create no runtime evidence, and use exit 0 for pass, 1 for mismatch, and 2 for invalid or untrusted state. See Control Simulation.

When ToolHub requires approval, the CLI, Dashboard, diagnostics, and Studio read the same signed quorum chain. amc approvals list can search the stable request ID and filter privacy-safe status, action, risk, mode, time, order, and limit fields. It audits every canonical request, decision, consumption record, and detached signature before filtering; malformed, tampered, misbound, or orphaned activity returns no rows. The view never searches tool names, intent/work-order IDs, reviewer identities, reasons, commands, prompts, or payloads, and it is marked derivedView: true, recorded: false, and proofEligible: false. Configured Integrations channels still receive only metadata-only lifecycle notifications. See Approvals.

amc tools list --json verifies the signed ToolHub allowlist, derives stable tool and MCP server identities, groups tools by declared provider context, and reuses those identities in the Fleet CGX trust graph. Existing version 1 configs remain native when context is omitted. Missing or invalid signatures, malformed declarations, duplicate tools, or conflicting server metadata return no rows. The projection is read-only and explicitly non-proof; it does not discover or attest a live MCP server. See ToolHub.

Auto-fix everything

amc fix                          # generate guardrails + CI gate + governance docs
amc fix --target-level L4        # target a specific level
amc guide --go                   # detect framework → apply guardrails to config
amc guide --watch                # continuous monitoring + auto-update

Compliance in one command

amc audit binder create --framework eu-ai-act            # EU AI Act evidence binder
amc compliance report --framework iso-42001              # ISO 42001 report
amc domain assess --domain health                        # HIPAA assessment
amc domain assess --domain wealth                        # MiFID II / DORA

GitHub Actions — preview boundary

The repository contains a composite action prototype, but no GitHub Marketplace publication has been independently verified. Do not use @main as a production trust boundary. Until an immutable action release is published and tested, run AMC from a reviewed source checkout in CI or pin the prototype to an exact reviewed commit; see amc-action/README.md.

Badge for your README

For a run-to-run or model-route comparison badge, generate it from the real comparison command:

amc compare <run-a> <run-b> --output compare.json --badge
amc compare gpt-4o-mini claude-3-haiku --agent support-bot --output model-compare.json --badge
# writes compare-badge.svg or model-compare-badge.svg beside the report

Standalone maturity badges are also available when you only need a README trust marker:

<!-- Add this to your README -->
[![AMC Score](https://img.shields.io/badge/AMC-L3_(72.5)-green?logo=data:image/svg+xml;base64,PHN2ZyB4bWxucz0iaHR0cDovL3d3dy53My5vcmcvMjAwMC9zdmciIHZpZXdCb3g9IjAgMCAyNCAyNCI+PHBhdGggZmlsbD0iI2ZmZiIgZD0iTTEyIDJMMiA3bDEwIDUgMTAtNXptMCA5bC04LjUtNC4yNUwyIDEybDEwIDUgMTAtNXptMCA5bC04LjUtNC4yNUwyIDIxbDEwIDUgMTAtNXoiLz48L3N2Zz4=)](https://github.com/AgentMaturity/AgentMaturityCompass)

Result:

14 Framework Adapters

Route a supported CLI or generated framework sample through AMC without replacing the agent runtime.

amc adapters init
amc adapters configure --agent my-agent --adapter generic-cli --route /openai --model gpt-4o --mode SUPERVISE
amc adapters run --agent my-agent --adapter generic-cli -- python bot.py

Adapter	Canonical ID	Version probe
AutoGen	`autogen-cli`	CLI when present, otherwise host Python
Claude Code	`claude-cli`	Claude binary
CrewAI	`crewai-cli`	CLI when present, otherwise host Python
Gemini CLI	`gemini-cli`	Gemini binary
Generic CLI	`generic-cli`	Shell runtime only
LangChain Node	`langchain-node`	Host Node.js only
LangChain Python	`langchain-python`	Host Python only
LangGraph Python	`langgraph-python`	Host Python only
LlamaIndex Python	`llamaindex-python`	Host Python only
OpenAI Agents SDK	`openai-agents-sdk`	Host Node.js only
OpenClaw	`openclaw-cli`	OpenClaw binary
OpenHands	`openhands-cli`	OpenHands binary
Python AMC SDK	`python-amc-sdk`	Installed package
Semantic Kernel	`semantic-kernel`	Host Node.js only

Do not treat a detected runtime as proof of event/control coverage. Issue a portable signed receipt that separates declared from currently effective capabilities and records lossiness:

amc adapters capabilities --agent my-agent --adapter claude-cli --out adapter-capabilities.json --json

Claude Code receipts can verify allow/deny/ask and bounded corrective steer only with a valid signed control hook. Steer blocks the current call, rewrites no input, and requires a new fully governed action for retry. Gemini CLI receipts expose allow/deny and explicitly record that ask or requested steer degrades to a lossy deny. Metadata-only plugin adapters fail closed instead of inheriting a compatibility claim.

For supported shell input, the same hook performs a compound-command blast-radius review across every bounded segment and applies the most restrictive result. Unsupported expansion or syntax fails closed before partial policy output. Signed evidence keeps only action classes, canonical tool names, outcomes, reason codes, and counts; the raw command and argument values are never retained.

Full adapter docs

Compliance Mapping

Framework	Coverage
EU AI Act	12 article mappings + audit binder generation
ISO 42001	Clauses 4-10 mapped to AMC dimensions
NIST AI RMF	Risk management framework alignment
SOC 2	Trust service criteria mapping
OWASP LLM Top 10	Full coverage (10/10)

Install AMC

macOS and Linux

curl -fsSL https://agentmaturity.co/install.sh | sh

Windows PowerShell

irm https://agentmaturity.co/install.ps1 | iex

Both scripts install a pinned GitHub release only after SHA-256 verification. They require Node.js 20 or 22 LTS. The npm registry and Homebrew tap are not public yet, so AMC does not present those channels as available.

Docker

docker build -t amc-quickstart -f docker/Dockerfile.quickstart .
docker run -it --rm amc-quickstart amc

Use the local build command unless a GHCR package has been verified public.

From source

git clone https://github.com/AgentMaturity/AgentMaturityCompass.git
cd AgentMaturityCompass && npm ci && npm run build && npm link

Deploy

Platform	Deploy
Docker Compose	`cd docker && docker compose up`
Vercel
Railway

Pricing

The full trust stack is free and MIT licensed. Industry Packs are the planned commercial add-on, but checkout and license issuance are not yet publicly live.

Tier	What you get
Free / Open Source	Everything — Score, Shield, Enforce, Vault, Watch, Comply, Fleet, Passport, all 15 adapters, 1,171 registered CLI command paths, browser playground, CI gates
Industry Packs (planned; not yet purchasable)	Planned `$9.99/month` tier for all 41 Industry Domain Packs; public checkout is not live
Enterprise (contact-first; not self-serve)	Planned Industry Packs access plus priority support, custom pack development, and deployment assistance

Industry Packs are 41 sector-specific domain packs (healthcare, finance, education, logistics, government, etc.) that require ongoing regulatory research and maintenance. The core trust stack stays MIT licensed. See website/publication-status.json before making any availability claim.

Choose Your Path

Path	Best for	Start here
Browser	First-touch evaluation, demos, understanding scoring	Web Playground
CLI	Real agent scoring, evidence capture, shareable outputs	Install AMC, then run `amc`
Desktop Studio	macOS and Windows users who want a local launcher app	Desktop Packages
CI/CD	Release gates, score thresholds, PR comments	CI Templates
Enterprise	Self-hosted, managed deployment	Deployment Options
Cloud self-hosting	AWS, GCP, Azure reference architectures	Cloud Reference Architectures

Start by persona

Solo builder / OSS maintainer → docs/SOLO_DEV_PATH.md
Platform / engineering team → docs/PLATFORM_PATH.md
Security / compliance → docs/SECURITY_PATH.md

Docs


Getting Started (2 min)	Agent Guide
Community Demo Kit	Why AMC One-Pager
Solo Dev Quickstart	Platform Engineer Quickstart
Security & Compliance Quickstart	Troubleshooting
CLI Reference (1,171 command paths)	Architecture
Compatibility Matrix	Starter Blueprints
Install Packages	Support Policy
Release Cadence	CI Templates
Hardening Guide	Community
Cloud Reference Architectures	Deployment Options
Assurance Lab	Domain Packs
EU AI Act Compliance	Multi-Agent Trust
Executive Overview	Desktop Packages
Board L3 Risk Memo	Install Guide
White Paper	Example Projects
Web Playground	Docs Index

More docs

docs/INDEX.md — full documentation index
docs/START_HERE.md — orientation guide
docs/WHY_AMC.md — the case for AMC
docs/USE_CASES.md — use case gallery
docs/PERSONAS.md — role-based guides
docs/AFTER_FIRST_SCORE.md — what to do after your first score
docs/EXAMPLES_INDEX.md — example index
docs/RECIPES.md — extended recipes
docs/DEPLOYMENT_OPTIONS.md — deployment options
docs/CLOUD_REFERENCE_ARCHITECTURES.md — AWS, GCP, and Azure self-hosted reference architectures
docs/PRODUCT_EDITIONS.md — product editions
docs/PRICING.md — pricing details
docs/BUYER_PACKAGES.md — buyer packages
docs/SERVICES_AND_SUPPORT.md — services and support
docs/COMMUNITY_SHOWCASE.md — community showcase
docs/RELEASE_HIGHLIGHTS.md — release highlights
docs/BENCHMARK_GALLERY.md — benchmark gallery
docs/SPONSORING.md — sponsorship
docs/COMMUNITY_SUPPORT.md — community and support

Desktop Studio apps

AMC desktop packages include Agent Maturity Compass Studio launchers for macOS and Windows. Each launcher verifies the included package digest and installs that exact build into a version-pinned per-user runtime, with mutable demo data kept in a separate persistent workspace. An older global CLI cannot start stale Studio code. The native macOS app uses a WebKit window; Windows opens the same loopback-only local Studio in the system browser. Neither bundles Electron or a browser runtime. Studio, CLI, docs, GitHub README, and the website share the same dark terminal identity.

npm run package:desktop
npm run package:desktop:verify

See docs/DESKTOP_PACKAGES.md for archive contents and platform commands.

Single-binary install (experimental)

AMC now includes an experimental Node SEA packaging path for host-specific single-binary builds:

npm run build
npm run build:sea

The build path is wired in and produces SEA artifacts plus a manifest. Runtime verification is still experimental and host-sensitive. See docs/SINGLE_BINARY.md for the honest status and caveats.

Nightly compatibility matrix

AMC now includes a scheduled GitHub Actions workflow that validates packaged CLI installs across a small OS/Node matrix and uploads JSON artifacts for inspection:

workflow: .github/workflows/nightly-compatibility-matrix.yml
current matrix: ubuntu-latest + macos-latest, Node 20 + 22
checks: packed install, top-level full score JSON, doctor --json, lite-score --help, comms-check --help

Workspace config profiles (MVP)

AMC now supports lightweight workspace config presets for .amc/amc.config.yaml:

amc init --minimal
amc init --profile dev
amc quickstart --minimal
amc quickstart --startup-plan --answers-out amc-startup-answers.json
amc quickstart --what-broken
amc quickstart --profile ci
amc config profile prod

Current MVP behavior:

dev → shared trust boundary, proxy env enabled
ci → isolated trust boundary, proxy env enabled
prod → isolated trust boundary, proxy env disabled
explicit --trust-boundary still overrides the profile when you need it
--minimal → startup-friendly setup without a vault prompt or immediate full-score prompt
--startup-plan → role-aware 10-minute startup plan, framework detection, and optional sample answer file
--what-broken → single-command startup blocker report without running the interactive score

Contributing

AMC is MIT licensed. We welcome contributions — especially new assurance packs, domain packs, framework adapters, and scoring modules.

git clone https://github.com/AgentMaturity/AgentMaturityCompass.git
cd AgentMaturityCompass && npm ci && npm test   # 1,087 files / 8,604 passing Vitest tests

→ CONTRIBUTING.md — includes guides for writing packs, mapping research papers, and adding adapters.

Good first contributions

New assurance pack — model a new attack scenario (guide)
New domain pack — add industry-specific questions (guide)
New adapter — support another agent framework (guide)
Research paper → module — turn arXiv findings into scoring logic (guide)

License

MIT — public trust infrastructure for the age of AI agents.

244 default diagnostic questions + 20 lifecycle expansion questions · 142 assurance packs · 41 domain packs · 15 adapters · 1,171 CLI command paths
Stop trusting. Start verifying.

Name		Name	Last commit message	Last commit date
Latest commit History 740 Commits
.amc		.amc
.changeset		.changeset
.github		.github
.serena		.serena
Formula		Formula
amc-action		amc-action
api		api
deploy		deploy
docker		docker
docs		docs
examples		examples
fixtures		fixtures
integrations/pytest-amc		integrations/pytest-amc
internal/debug		internal/debug
mirofish-simulation		mirofish-simulation
platform		platform
qa		qa
research		research
scripts		scripts
sdk/python		sdk/python
src		src
tests		tests
tools/evil-mcp-server		tools/evil-mcp-server
vscode-extension		vscode-extension
website		website
whitepaper		whitepaper
.dockerignore		.dockerignore
.editorconfig		.editorconfig
.gitignore		.gitignore
.mailmap		.mailmap
.mcp.json		.mcp.json
.nvmrc		.nvmrc
.tmp-a2a-events.json		.tmp-a2a-events.json
.tmp-distributed-events.json		.tmp-distributed-events.json
.tmp-gap-report.json		.tmp-gap-report.json
.tmp-memory-depth-events.json		.tmp-memory-depth-events.json
AMC_COMPLETE_KNOWLEDGE.md		AMC_COMPLETE_KNOWLEDGE.md
CHANGELOG.md		CHANGELOG.md
CODE_OF_CONDUCT.md		CODE_OF_CONDUCT.md
COMPETITIVE_ANALYSIS_G0DM0D3.md		COMPETITIVE_ANALYSIS_G0DM0D3.md
COMPETITIVE_GAP_REPORT_G0DM0D3.md		COMPETITIVE_GAP_REPORT_G0DM0D3.md
CONTRIBUTING.md		CONTRIBUTING.md
Dockerfile		Dockerfile
Dockerfile.runner		Dockerfile.runner
LICENSE		LICENSE
README.md		README.md
SECURITY.md		SECURITY.md
action.yml		action.yml
compliance-eu_ai_act.json		compliance-eu_ai_act.json
compliance-gdpr.json		compliance-gdpr.json
compliance-iso_42001.json		compliance-iso_42001.json
compliance-nist_ai_rmf.json		compliance-nist_ai_rmf.json
compliance-soc2.json		compliance-soc2.json
install.sh		install.sh
package-lock.json		package-lock.json
package.json		package.json
railway.json		railway.json
sbom.json		sbom.json
test_model.pkl		test_model.pkl
tsconfig.json		tsconfig.json
vercel.json		vercel.json
vitest.config.ts		vitest.config.ts

Folders and files

Latest commit

History

Repository files navigation

Agent Maturity Compass

What is this?

60 Seconds to Your First Score

How AMC Compares

What AMC Tests

244 Default Diagnostic Questions × 5 Dimensions

142 Assurance Packs

41 Industry Domain Packs

Simulation & Forecast Evaluation Lane

79 Scoring Modules

Architecture

Evidence Trust Tiers

Maturity Scale

Product Family

Recipes — Copy-Paste Examples

Score any agent in one line

Wrap an existing agent (zero code changes)

Observe or control native provider actions

Red-team your agent

Inspect traces and operational drift

Run realtime monitoring

Build golden datasets and run evals

Business, inventory, and reporting

Auto-fix everything

Compliance in one command

GitHub Actions — preview boundary

Badge for your README

14 Framework Adapters

Compliance Mapping

Install AMC

macOS and Linux

Windows PowerShell

Docker

From source

Deploy

Pricing

Choose Your Path

Start by persona

Docs

Desktop Studio apps

Single-binary install (experimental)

Nightly compatibility matrix

Workspace config profiles (MVP)

Contributing

Good first contributions

License

About

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

Watchers

Forks

Releases

Packages

Used by

Contributors

Languages