Business Users
Describe the agent in chat or through the guided wizard — at most three clarification rounds; unanswered questions become labeled assumptions — and marshal writes the specification.
Build · Govern · Deploy — open source under Apache-2.0, self-hosted on AWS
marshal.build
Turn plain-English ideas into governed, deployed AI agents on AWS.
Describe what you need in chat. marshal shapes it into a structured specification — requirements, design, tasks — generates a serverless agent (Lambda, API Gateway, Amazon Bedrock) from that spec, validates the code against deterministic gates, scores its risk against an admin-tunable rubric, and deploys it into a governed AWS environment with a time-to-live, a budget cap and a full audit trail.
Self-hosted in your own AWS account. The repository is the product.
- Build an expense receipt triage agent for our finance team
- requirements.md — user stories & acceptance criteria
- design.md — architecture (Mermaid) & data flow
- tasks.md — implementation plan
- validation gate — service allowlist · packaging · API key on every route
- risk score: medium → reviewer approval recorded
- Deployed · isolated AWS account (Innovation Sandbox) or your own account · TTL · budget cap · audit trail
How it works
Build. Govern. Deploy.
Chat to specification to deployed agent. Governance attaches at fixed seams in the code — one for model calls, one for access decisions, one for audit — so a feature cannot bypass it.
-
From plain English to deployed AI agent
Describe the agent in chat or through the guided wizard; attach a PDF, Word document or existing spec as grounding. marshal shapes it into a structured specification — requirements, design, tasks — kept as append-only versions with diff and rollback, then generates a serverless agent (Lambda + API Gateway + Amazon Bedrock) from the spec.
-
Validated, scored, reviewed
A deterministic validation gate blocks any build with a finding: AWS service allowlist, packaging shape, API-key contract on every route, forbidden patterns such as hard-coded credentials. A seven-factor rubric scores the spec; low risk auto-approves, medium and high route to named reviewers whose decision is recorded. Every model call and every mutating action lands in one audit trail.
-
Into a governed AWS environment
Approved agents deploy as CloudFormation stacks into your installation account (direct) or into an isolated account leased from AWS Innovation Sandbox (isb) — with health probes, a post-deploy smoke check, a time-to-live, a budget cap and clean teardown. The CloudFormation timeline streams to the UI as it happens.
Who it’s for
Three roles, one governed pipeline
Roles — business, power, admin — come from Cognito groups; an administrator provisions every user, and the same gates apply to all of them.
Power Users
Pick models from the template allowlist and tune parameters within platform bounds; every clamp is recorded and shown as the effective value. Declare capabilities — memory, tools, connectors, packaged dependencies, agent composition — in the spec itself.
Platform Admins
Model allowlist and parameter bounds; monthly cost caps per user, project and platform that alert or block; deployment policies for concurrency, TTL, budget and endpoint authentication; a versioned risk policy; an audit log exportable as CSV or NDJSON.
Security reviewers
The threat model and control map (docs/security.md) names each control, where it lives in the code and what was tested. Vulnerability reports go to security@marshal.build and are acknowledged within three business days.
Features
Everything between the idea and the audit log
Chat, spec engine, validation gate, templates and deployments — one governed pipeline.
Streaming AI chat
Claude models on Amazon Bedrock through cross-region inference profiles; responses stream over SSE. Attach a PDF, Word document or existing spec as grounding — extraction is local and deterministic, no OCR, no model call.
Full spec generation
Requirements, design with a Mermaid architecture diagram, and an implementation plan — each structurally validated with one model repair pass, with per-document progress streamed as it runs.
Spec editor with diff & rollback
Monaco editor with version pickers and side-by-side diff. Specs are append-only versions; rollback writes the chosen version as a new latest version and re-runs the same checks.
Templates & guardrails
Templates carry model allowlists, token ceilings, temperature bounds and the capabilities a project may declare. The rails are enforced server-side on every call, and a guardrail edit bumps the template version.
Deterministic validation gate
Pure-local checks on every generated build — AWS service allowlist, packaging shape, API-key contract, console and connector wiring, forbidden patterns, spec conformance. Any finding fails the build; there is no "deploy anyway".
Keyed endpoints by default
Generated APIs require an API key and usage plan on every route unless the requirements state "Endpoint authentication: PUBLIC", and administrators can forbid that opt-out. The owner reveals the key live from the deployed stack; the platform stores nothing.
Connectors, tools and composable agents
Specs declare capabilities in plain sentences — "SHALL keep conversation memory", "SHALL use connector <slug>", "SHALL call agent <slug>" — and the platform builds the memory table, the tool loop or the agent graph. Templates decide which rungs a project may use.
Enclave deployments
In isb mode each deployment gets its own AWS account leased from AWS Innovation Sandbox, with a budget and an expiry, recycled at teardown. In direct mode agents deploy into the installation account through a dedicated deployment role that is denied the control plane’s own data and roles.
Governance & cost attribution
Every stack is tagged with its project, owner and deployment id. Model spend is charged to the caller and Enclave spend to the project owner; monthly caps per user, project and platform alert or block.
Run it yourself
One edition, two install tiers
Everything in the repository is available to every installation under Apache-2.0; nothing is license-gated or entitlement-checked. The code is the product.
Evaluate — SANDBOX_PROVIDER=direct (default)
Generated agents deploy as CloudFormation stacks into the installation account. One AWS account, nothing else to set up; the backend assumes a dedicated, scoped deployment role for each deployment. Run evaluations in a dedicated account.
Operate — SANDBOX_PROVIDER=isb
AWS Innovation Sandbox vends an isolated AWS account per deployment — an Enclave — with its own budget and lease, recycled at teardown. Needs an AWS Organization, IAM Identity Center and the Innovation Sandbox solution: about 0.5–1 day of setup plus account-creation lead time. Setup guide.
You need an AWS account with administrator credentials in a named CLI profile, AWS CLI v2 ≥ 2.9, Node.js 22, Python 3.12 with uv, a CDK bootstrap, and Bedrock access to the default Anthropic models. Region: us-east-1 only. No Docker — the three images build in CodeBuild.
- git clone https://github.com/xyme/marshal.build.git && cd marshal.build
- cp scripts/deploy-env.local.sh.example scripts/deploy-env.local.sh # then set AWS_PROFILE, MARSHAL_AWS_ACCOUNT_ID, OPS_ALERT_EMAIL
- (cd infra && npm ci)
- ./scripts/deploy-cloud.sh --bootstrap
- about 40 minutes on a fresh account (measured 41), seven steps, fail-closed health smoke at the end
Idle floor of the control plane in us-east-1 (on-demand pricing, before Bedrock usage and before any Innovation Sandbox account): ≈ $120–125/month at MARSHAL_TIER=evaluate — single-AZ RDS, one task per service — and ≈ $190–195/month at the default production sizing — Multi-AZ RDS, two tasks per service; measured at ≈ $6.3/day on a fresh-account install. Innovation Sandbox adds roughly $10–30/month of control infrastructure plus each leased account’s own spend.
Contributions are welcome under a Contributor License Agreement; the checks and conventions are in CONTRIBUTING.md.
An audit trail, budget caps and a risk gate on every deployment
- Runs in your own AWS account
- Cognito sign-in with TOTP MFA; Entra ID / Okta federation by configuration
- Backend not internet-reachable; WAF at the edge
- Every deployment under a TTL and a budget cap
- Every model call and every mutating action audited
- Supply-chain gate on every image build
- Apache-2.0 · one edition · nothing license-gated
Built in the open
The figures behind the release
Release history with root causes is in CHANGELOG.md; the engineering walkthrough of every functional area is docs/ARCHITECTURE.md. In the week before release the supply-chain gate refused a deploy over 14 fixable advisories; the dependencies were upgraded, not allowlisted.
616
backend tests at v0.1.0; the suite needs no AWS credentials and no Postgres
95
browser smoke checks at v0.1.0, run against a deployed installation at the end of every deploy
7
deterministic validation checks on every generated build; any finding fails it
7 Oct 2026
v0.1.0 public source release under Apache-2.0
Ready to turn an idea into an agent?
Clone the repository, deploy into your own AWS account in about 40 minutes, and ship your first agent through the risk gate.
Self-hosted in your own AWS account.