Babel

Open-source coding agent for real software work.

Babel plans, reviews, and implements engineering tasks in your local workspace. For each task it loads a small, explicit instruction stack—behavioral rules, domain expertise, skills, model adapters, and project overlays—so you can see how the agent will work before it edits a line of code.

Product	Local coding agent — TUI/REPL + chat / plan / deep
Primary UI	Interactive terminal session (babel)
Release	v0.1.0 · pre-1.0
Source of truth	This repo — gthgomez/Babel
Edge	Inspectable stacks, domain routing, plan → review → execute

This is the canonical public source. Day-to-day Babel development happens here. Pin consumers with a release tag + commit SHA. See ADR-0001 and docs/guides/RELEASE.md.

What Babel does

You need to…	Babel
Ship a feature or fix	Plan the work, then run an implementation session
Stay safe on risky changes	Prefer plan or deep when you want stronger gates before or during execution
Work across domains	Route backend, frontend, mobile, and more to the right expertise—not one generic prompt
Reuse hard-won workflows	Skills for testing, review, governance, release hygiene, and more
Trust the agent’s “brain”	Preview the exact stack from the catalog before the model runs
Integrate other tools	Read-only MCP for stack and manifest inspection

Most coding agents hide the system prompt. Babel makes the agent’s operating instructions modular, versioned, and testable—engineered software, not an improvised blob.

Quick start

cd .\babel-cli
npm install
npm run build
node .\dist\index.js doctor

Validate the public surface (no model required):

cd ..
pwsh -File .\tools\validate-public-release.ps1

See the stack the agent would load for a backend task:

pwsh -File .\tools\resolve-local-stack.ps1 `
  -TaskCategory backend `
  -Project example_saas_backend `
  -Model codex `
  -PipelineMode chat `
  -Format json

Golden preview: examples/manifest-previews/backend-verified.json

Mobile lane:

pwsh -File .\tools\resolve-local-stack.ps1 `
  -TaskCategory mobile `
  -Project example_mobile_suite `
  -Model codex `
  -Format json

Golden: examples/manifest-previews/mobile-direct.json

With local provider credentials configured:

cd .\babel-cli
node .\dist\index.js                 # interactive TUI (chat)
node .\dist\index.js plan "..."      # plan mode
node .\dist\index.js deep "..."      # governed deep path

Full onboarding: START_HERE.md · CLI reference: docs/CLI_QUICKSTART.md

How the agent works

Before the model acts, Babel resolves an ordered stack from prompt_catalog.yaml:

1. Behavioral rules — how the agent plans, acts, verifies, and stays safe
2. Domain architect — backend, frontend, mobile, … for this kind of work
3. Skills — reusable workflows the task needs
4. Model adapter — tune the same agent for the model you chose
5. Overlays — project- or task-specific context
6. Manifest — the exact load order you can preview and test

That is the product idea: a coding agent whose instructions are engineered, not improvised.

Product modes

Mode	When to use
chat	Default conversational agent (TUI and one-shot tasks)
chat-headless	Same engine for scripts / CI
plan	Plan first, then apply with approval
deep	Full governed pipeline for higher-risk work

Legacy aliases (verified→deep, manual→plan, direct→chat) still work with deprecation warnings.

Inspect vs execute

Mode	Purpose
Inspect	Validate catalog, preview stack, compare goldens, MCP — no model required
Execute	babel (TUI), babel plan, babel deep — model-backed sessions

Pre-1.0 honesty: model-backed runs are real and typechecked, but need local credentials and workspace setup. You can still validate and preview stacks from a clean clone with no API keys.

What’s in the repo

Path	Role
00_–06_ prompt layers	Router, behavior, domains, skills, adapters, overlays
prompt_catalog.yaml	Routable stack contract
babel-cli/	Coding agent runtime (TUI + chat / plan / deep)
examples/	Golden stack previews and first-success walkthroughs
tools/	Validate, scrub, content policy, release gates
docs/	Architecture, vision, CLI guides

Babel/
├── START_HERE.md
├── BABEL_BIBLE.md         # agent / integration entry
├── prompt_catalog.yaml
├── 00_System_Router/
├── 01_Behavioral_OS/
├── 02_Domain_Architects/
├── 02_Skills/
├── 03_Model_Adapters/
├── 04_Meta_Tools/
├── 05_Project_Overlays/
├── 06_Task_Overlays/
├── babel-cli/
├── examples/
├── docs/
└── tools/

Status (v0.1.0)

• Canonical open-source coding agent surface — clone, validate, run locally
• Public Release Gate on every PR: security → public-content-policy → linux-validation (+ windows-portability)
• main is PR-only; v* tags protected; secret scanning + push protection on
• Optional pre-commit hooks; CI is authoritative (CONTRIBUTING.md)

Pin when ready:

{
  "babel": {
    "tag": "v0.1.0",
    "sha": "8184bbbbfa818001382fdeaf8e9d51ba8bf6003d"
  }
}

Agent contracts and CLI surfaces may still change before 1.0.0.

Verification

pwsh -File .\tools\validate-public-release.ps1
pwsh -File .\tools\check-public-content-policy.ps1
pwsh -File .\tools\check-canonical-independence.ps1

Maintainers may use a confidential supplemental policy stored outside this repo:

pwsh -File .\tools\validate-public-release.ps1 -Strict `
  -RequireSupplementalPolicy `
  -SupplementalPolicyPath $env:BABEL_PRIVATE_SCRUB_POLICY_PATH

MCP (read-only):

cd .\babel-cli
npm run build
node .\dist\index.js mcp

Docs

• START_HERE.md — first success
• docs/CLI_QUICKSTART.md — TUI, chat / plan / deep, doctor, MCP
• docs/VISION.md — product direction
• docs/architecture/ARCHITECTURE.md — system shape
• BABEL_BIBLE.md — invocation contract for models and wrappers
• CONTRIBUTING.md

Contributing

Highest-value work:

• stronger agent behavior (skills, domains, adapters)
• clearer plan/run UX and diagnostics
• end-to-end coding task examples
• tighter stack selection and release safety

Keep out: private dependency fingerprints, credentials, machine paths, operator-only notes.

License

MIT. Use it, fork it, and build on it.

Full text: LICENSE