Buracle StackBuracle Stack

The layer that brings your coding agents together

Claude Code, Codex CLI, OpenCode, Oh My Pi, Qwen Code, Pi and AGY CLI get one setup, one shared memory, the same guards and workflows that end with proof.

Install with one command

curl -fsSL https://stack.buracle.dev/install.sh | bash

Read the guideWatch the film

Seven coding agents on an orbit around one layer, Buracle Stack, with its MCP servers on the inner ring. Pick an agent to see how Buracle Stack plugs into it.

Pick an agent to see how it plugs in.

Buracle Stack in numbers

  • 7coding agents set up from one source
  • 9MCP servers, routed per question
  • 6language servers in the toolchain
  • 26hooks guarding every tool call
  • 13skills, the same text in every agent
  • 4workflows that end with proof

The film

Buracle Stack in 40 seconds

Memory

One brain, shared by every agent

A decision you make in Codex CLI is there when Claude Code starts the next session. Each project has its own brain, and the person brain keeps what holds in every project.

Memories cite the files they are about, so memory_search with files: finds every memory about a file. The Console draws the same network: press G there.

referenceProject brain

Next.js cache tags: unstable_cache with tags, then revalidateTag

Origin
inferred: the agent concluded it
Cites
app/api/invoices/route.ts

Written in Qwen CodeRecalled in Claude Code

An example brain seen from above: the project brain on the left, the person brain on the right. Each bright node is a memory. Select one to read it. An example brain, invented for this page

Project brainacme-shop

Person brainevery project

Read the example memories as a list
  • Project brain

    • decision Totals are computed in cents, never in floats
    • gotcha sumItems stopped one item early: the loop bound was items.length - 1
    • decision CSV exports stream rows instead of building the file in memory
    • reference Next.js cache tags: unstable_cache with tags, then revalidateTag
    • procedure Release: pnpm build, pnpm test, then tag the version
    • knowledge Finance reads the reports every Monday at 09:00 UTC
  • Person brain

    • preference Commit messages in the imperative, under 60 characters
    • preference Reply in Turkish, keep code and commits in English
    • procedure Run the type checker before every commit
    • gotcha On Apple silicon, Homebrew lives in /opt/homebrew, not /usr/local

Workflows

Workflows that loop until the work is proven

Each workflow writes down what it will do, starts every behavior change with a failing test, gets an independent review and ends only when the result is proven.

/plan-build

Use it whenYou want a reviewed plan approved before any code.

  1. Plan
  2. Plan review
  3. Your approval
  4. BuildRED → GREEN per task
  5. Verifychanges review, real run, E2E
  6. Your sign-off
  7. VERIFIED

A failed check adds fix tasks and goes back to Build.

Example/plan-build add CSV export to the reports page

/goal-build

Use it whenThe end state is clear and the path is found by building.

  1. Goal and oracle
  2. Contract with pass/fail criteria
  3. Review and lock
  4. Build round
  5. Judge every criterion
  6. Verify
  7. Hand back

Failing criteria become the next round's tasks.

Example/goal-build search p95 under 200 ms on the seed data

/bugfix

Use it whenOne behavior is broken and you want proof it is fixed.

  1. Reproduce
  2. Root cause as file:line
  3. RED test that fails for the reported reason
  4. Fix at the source
  5. Prove on the real program
  6. Quality gate
  7. Finish

A test that passes before the fix means the root cause is wrong: back to Investigate.

Example/bugfix invoice total skips the last line item

/requirements

Use it whenThe problem or the scope is still unclear.

  1. Understand
  2. Research
  3. Ideate
  4. Clarify
  5. Propose
  6. Converge
  7. PRD and your approval
  8. Hand off to /plan-build or /goal-build

Example/requirements saved searches for the reports page

In /plan-build and /goal-build a stop guard keeps the turn open until the plan is VERIFIED. Say stop, wait or hold on to pause. Six cases, step by step

Tools

Each question goes to the tool built for it

Routing sends each kind of question to one of nine MCP servers. Pick a server to see the question it answers and what it adds.

The agent asksHow do I tag cache entries in Next.js 15?

context7by Upstash

Current docs for the library version you use, so the agent stops guessing APIs.

Language servers and tools in the toolchain

  • TypeScript
  • Pyright
  • Go
  • Rust
  • C#
  • Java
  • ripgrep
  • ast-grep
  • jq
  • GitHub CLI
  • Ruff
  • ShellCheck
  • golangci-lint
  • Bun
  • Node.js
  • uv

Guards answer before a tool runs

  • rm -rf ~denied
  • git push --force origin feature-xasks you
  • .envdenied