OPEN SOURCE BY Enric

We Build the Tools We Wished Existed

Enric is an engineering studio behind three open source projects — an autonomous coding orchestrator, a browser-to-API recon pipeline, and a small-model code-generation framework. No product to buy. No login required. Just the tools, on GitHub.

Explore the ProjectsView on GitHub

Built and maintained in the open — every guarantee below is enforced in code you can read.

Open Source Projects
3
Vendor Lock-In
0
Auditable Source
100%

Our Open Source Projects

Actively Maintained

Leerie

The autonomous task driver for Claude Code

  • Evidence-gated confidence is the only hard gate before a worker's code counts as done
  • Isolated git worktrees keep parallel subtasks from ever colliding
  • Resumable by design — a reboot, rate limit, or Ctrl-C loses nothing
PythonMIT License
Leerie →View on GitHub
Actively Maintained

Barnacle

Turn browser recon into a deterministic API integration

  • Recon once, replay forever: direct HTTP hot path with automatic browser fallback
  • Zod schemas, rate limits, and headers generated straight from captured traffic
  • A nightly smoke test catches contract drift before it reaches production
TypeScriptMIT License
Barnacle →View on GitHub
Coming Soon

Beacon

Small models, senior-level code — grounded in how programmers read

  • AST-aware indexing with dependency and call-graph analysis
  • Four cognitive modes — recognize, comprehend, design, forage
  • Risk reporting surfaces viscosity, hidden dependencies, and bus-factor concentration
Beacon →
Actively Maintained

How Leerie Works

The model writes code. The program runs everything else.

Leerie classifies a task, decomposes it into right-sized subtasks, and hands each one to an isolated worker in its own git worktree. Every worker's output is JSON-schema-validated and self-gated on evidence-anchored confidence before the orchestrator — ordinary Python, not another model — decides the run is done.

  1. 1

    Program, not agent

    Phases, wave scheduling, retries, and merge logic are ordinary Python — real loops and conditionals that cannot drift, deciding when work is done instead of letting a model declare it.

  2. 2

    No silent failures

    Every worker's output is JSON-schema-validated before the orchestrator acts on it — malformed output or a confident hallucination can't push the system into an undefined state.

  3. 3

    Confidence is the only hard gate

    Each implementer clears domain-specific evidence gates — file-and-line citations, reproductions, falsification attempts — before a confidence score of 9 or higher is trusted.

  • ✓Evidence-gated confidence is the only hard gate before a worker's code counts as done
  • ✓Isolated git worktrees keep parallel subtasks from ever colliding
  • ✓Resumable by design — a reboot, rate limit, or Ctrl-C loses nothing
View on GitHub
Recon once. Replay forever.

How Barnacle Works

A real Stagehand-driven browser walks your described user flow once, wiretapping every API call it triggers. Those captures are replayed with plain HTTP to prove which calls survive without a browser — the survivors become committed constants your production traffic hits directly.

1
Browser recon, once
A real browser walks your described user flow while a session-level listener wiretaps every API call it triggers, before recon ever ends.
2
HTTP replay proves the contract
Captured calls are replayed with plain fetch — no browser, no AI — to prove which endpoints work standalone before anything ships to production.
3
Fallback, not failure
If the direct HTTP path ever breaks, the browser session re-engages automatically while a nightly smoke test flags the drift for a small, targeted PR.
1Browser recon, once

A real browser walks your described user flow while a session-level listener wiretaps every API call it triggers, before recon ever ends.

2HTTP replay proves the contract

Captured calls are replayed with plain fetch — no browser, no AI — to prove which endpoints work standalone before anything ships to production.

3Fallback, not failure

If the direct HTTP path ever breaks, the browser session re-engages automatically while a nightly smoke test flags the drift for a small, targeted PR.

View on GitHub
Coming Soon

How Beacon Works

A cognitive loop, not a bigger model.

Beacon indexes a codebase into an AST with dependency and call-graph analysis, then cycles a task through four cognitive modes modeled on how expert programmers actually build understanding of unfamiliar code — producing a reviewable diff without requiring a frontier-scale model.

Recognize, comprehend, design, forage

Beacon cycles through four cognitive modes — recognize, comprehend, design, forage — modeled on how expert programmers actually build understanding of unfamiliar code.

Metacognitive self-calibration

A metacognitive layer tracks bias, calibrates confidence, and escalates a knowledge_request when Beacon genuinely lacks the context to proceed.

Judged, GREEN-tier routing

Only call types that scored 91-100% factual accuracy in a 296-call LLM-as-judge audit (July 2026) are routed to a smaller model — everything else stays on the primary model.

Quickstart

Get Running in Under a Minute

Every project ships as a real, working install — no waitlist, no signup, just a command.

Leerie

One command installs the CLI and sets up the container runtime for you.

curl -fsSL https://raw.githubusercontent.com/enricai/leerie/main/scripts/install.sh | bash

Barnacle

Clone the repo, install dependencies, and copy the environment template.

pnpm install
cp .env.example .env

Beacon

Beacon's repository isn't public yet — the deep dive above covers how it works.

Coming Soon
Open Source

Built Different, By Design

Three projects, one philosophy: replace supervision with evidence. Here's how each one stacks up against a named alternative.

Leerie

What Matters
The Named Alternative
Enric's Approach
Who decides a task is done
GitHub Copilot (Agent mode)The model judges its own completion — it inspects errors and test results, then decides for itself when to stopSourceAs of 2026-07-28
A deterministic evidence gate: confidence of 9/10 or higher on root cause and solution, backed by tested falsifiersLEERIE.md — "Confidence is the only hard gate"
Billing model
Cognition DevinPlans run $0-$200/mo, but usage beyond the plan is billed "at API pricing" — cost still scales with useSourceAs of 2026-07-28
Runs on your existing Claude Code subscription — no Anthropic API key, no per-call billingLEERIE.md — opening paragraph and "Requirements"
Recovery from a stuck or interrupted run
GitHub Copilot coding agentA stuck session times out after an hour; recovery means manually unassigning and reassigning the issueSourceAs of 2026-07-28
Resumable by design — --resume picks up from the last completed wave after any reboot, kill, or rate limitLEERIE.md — "Resumable by design"
Who decides a task is done
The Named AlternativeGitHub Copilot (Agent mode)The model judges its own completion — it inspects errors and test results, then decides for itself when to stopSourceAs of 2026-07-28
Enric's ApproachA deterministic evidence gate: confidence of 9/10 or higher on root cause and solution, backed by tested falsifiersLEERIE.md — "Confidence is the only hard gate"
Billing model
The Named AlternativeCognition DevinPlans run $0-$200/mo, but usage beyond the plan is billed "at API pricing" — cost still scales with useSourceAs of 2026-07-28
Enric's ApproachRuns on your existing Claude Code subscription — no Anthropic API key, no per-call billingLEERIE.md — opening paragraph and "Requirements"
Recovery from a stuck or interrupted run
The Named AlternativeGitHub Copilot coding agentA stuck session times out after an hour; recovery means manually unassigning and reassigning the issueSourceAs of 2026-07-28
Enric's ApproachResumable by design — --resume picks up from the last completed wave after any reboot, kill, or rate limitLEERIE.md — "Resumable by design"

Barnacle

What Matters
The Named Alternative
Enric's Approach
Infra cost of a browser-driven request at scale
BrowserbaseEvery production request still spins up and bills a browser session — $20/mo for 100 browser-hours, then meteredSourceAs of 2026-07-28
A browser drives the flow once during recon; the surviving direct-HTTP calls become the production hot pathBARNACLE.md — "The mental model" and "Hot-path fallback triggers"
Per-request cost of a managed scraping API
Bright Data / FirecrawlBright Data charges $1.5 per 1,000 requests; Firecrawl runs about $0.00083 per page — every single call is billedSourceAs of 2026-07-28
Identical concurrent requests coalesce into one upstream call, and successful responses cache for 15 minutesBARNACLE.md — "Cache deduplication"
Infra cost of a browser-driven request at scale
The Named AlternativeBrowserbaseEvery production request still spins up and bills a browser session — $20/mo for 100 browser-hours, then meteredSourceAs of 2026-07-28
Enric's ApproachA browser drives the flow once during recon; the surviving direct-HTTP calls become the production hot pathBARNACLE.md — "The mental model" and "Hot-path fallback triggers"
Per-request cost of a managed scraping API
The Named AlternativeBright Data / FirecrawlBright Data charges $1.5 per 1,000 requests; Firecrawl runs about $0.00083 per page — every single call is billedSourceAs of 2026-07-28
Enric's ApproachIdentical concurrent requests coalesce into one upstream call, and successful responses cache for 15 minutesBARNACLE.md — "Cache deduplication"

Beacon

What Matters
The Named Alternative
Enric's Approach
Codebase context strategy for code-generation grounding
CursorContext comes from embeddings-based pattern matching, with no documented AST parsing or call-graph analysisSourceAs of 2026-07-28
AST-aware codebase indexing via tree-sitter, with dependency and call-graph analysisBEACON.md — "Features"
Basis for routing a call to a smaller model
CursorAbout 50 models are offered, but which one handles a call is a manual, per-conversation choice with no accuracy auditSourceAs of 2026-07-28
Routing to a smaller model is limited to the three call types a 296-call audit found most accurate (91-100%)BEACON.md — "Optional: Qwen routing experiment"
What the tool does to your files when it finishes
Cursor / DevinBoth write directly to disk and can commit or open PRs on their own, with checkpoints as the rollback mechanismSourceAs of 2026-07-28
Never modifies your source files — produces reviewable diffs in memory that you apply yourselfBEACON.md — "What Beacon Produces"
Codebase context strategy for code-generation grounding
The Named AlternativeCursorContext comes from embeddings-based pattern matching, with no documented AST parsing or call-graph analysisSourceAs of 2026-07-28
Enric's ApproachAST-aware codebase indexing via tree-sitter, with dependency and call-graph analysisBEACON.md — "Features"
Basis for routing a call to a smaller model
The Named AlternativeCursorAbout 50 models are offered, but which one handles a call is a manual, per-conversation choice with no accuracy auditSourceAs of 2026-07-28
Enric's ApproachRouting to a smaller model is limited to the three call types a 296-call audit found most accurate (91-100%)BEACON.md — "Optional: Qwen routing experiment"
What the tool does to your files when it finishes
The Named AlternativeCursor / DevinBoth write directly to disk and can commit or open PRs on their own, with checkpoints as the rollback mechanismSourceAs of 2026-07-28
Enric's ApproachNever modifies your source files — produces reviewable diffs in memory that you apply yourselfBEACON.md — "What Beacon Produces"
3Leerie
2Barnacle
3Beacon

THE ENRIC APPROACH

Why These Three Projects Matter

Every Enric project starts from the same non-negotiables.

  • Evidence over assertion

    Confidence scores are backed by falsifiers, citations, and reproductions — never a model's self-report.

  • Deterministic control flow

    Real loops and conditionals decide what happens next in every project — not a model narrating its own plan.

  • Built to run unattended

    Resumable state, isolated execution, and graceful fallback mean a reboot, rate limit, or site change never loses work.

  • Open source, auditable

    Every one of these projects ships its source — you can read exactly how the guarantees above are enforced.

GET INVOLVED

Contribute, Star, or Just Follow Along

These are living projects. Issues, pull requests, and questions are welcome on every repo — no account or login required to read the source.

Leerie
Barnacle
BeaconComing Soon
View All Repositories
Enric — Open Source AI Developer Tools: Leerie, Barnacle & Beacon