Skip to main content
The EDIFY wordmark, flat white type with two pools of blue ink travelling through it

EDIFY

Three documents and a map. Think hard once, then build without thinking.

Install · The four commands · The graph · Docs · Website

EDIFY is a harness that sits around an AI coding agent and makes it work reliably on large, real codebases — a million lines, fifteen years of history, four languages. It is not a model, not an agent framework, and not a wrapper around a chat interface: a stdlib-only CLI, edify, plus four short commands your agent runs.

A frontier model with a plain setup writes excellent code and still fails on serious projects, from four structural gaps no model generation closes. EDIFY closes exactly those four and gets out of the way of everything else:

  • It cannot see a codebase it cannot fit in context → a mechanically extracted graph, queried instead of re-derived every session.
  • It fragments across sessions → a spec, a plan and a task list that every session and every swarm member builds against.
  • It builds what was asked instead of what was meant → /spec marks unknowns [NEEDS CLARIFICATION] and ends at a human.
  • It writes against the library it remembers, not the one installed → a pinned, per-role MCP server registry.

Install

macOS, Linux and Windows, from one artifact. Python 3.10+ is the whole requirement — no dependencies, no compiler, no runtime. The package is edify-agents-cli on PyPI.

pip install edify-agents-cli            # or: pipx install / uv tool install edify-agents-cli
brew install edify-dev/tap/edify        # macOS, if you prefer brew

cd your-repository
edify init                              # detect, extract the graph, install, write CLAUDE.md

edify init new ./service                # or: a folder from zero, for every agent

edify init, in flat grey type, filling with blue ink

init detects the stack, extracts the graph, scrapes conventions from your config files, installs the skill entries your stack matches, seeds the server registry, writes a fifteen-line CLAUDE.md, and records every file it put there. It asks one question — which library entries to install, since they become trusted instruction — and with no terminal (CI, a pipe, a model) it asks nothing.

edify init at a terminal: the wordmark with blue ink in it, then one line per step, ending on Next: /spec
edify init --yes                        # the whole matched set, no question
edify init --skills planner,verifier    # exactly these
edify setup ./service --yes             # a new folder, root pinned there
edify setup . --fresh --yes             # reinstall the harness over this repository

edify init new installs for every agent runtime at once, after eight skippable questions whose answers go to .edify/profile.md:

runtime reads commands land in
Claude Code CLAUDE.md .claude/commands/, .claude/skills/, .claude/agents/
OpenAI Codex AGENTS.md .codex/prompts/
Cursor AGENTS.md, .cursor/rules/edify.mdc .cursor/commands/
Gemini CLI GEMINI.md .gemini/commands/ (TOML)
GitHub Copilot .github/copilot-instructions.md .github/prompts/
Windsurf .windsurf/rules/edify.md .windsurf/workflows/
.claude in ink .codex in ink .cursor in ink
.gemini in latex .github in blue latex .windsurf in ink

Per-platform detail, PATH fixes, updating, and uninstalling: INSTALL.md. It runs offline: only edify upgrade and edify self update open a socket.

The four commands

command what it does ends at
/spec a brief becomes a specification; unknowns marked [NEEDS CLARIFICATION] a human
/plan binding decisions, milestones, a partition for parallel work — only when needed a human
/tasks spec + plan + graph → leaseable tasks, then builds them: one agent or a swarm a green suite
/verify evidence on every surface — service, browser, device, live API — from a fresh session that wrote no code a report anyone can read

The build is the same loop at any scale:

edify swarm claim            # lease the next ready task nobody's work overlaps
edify swarm prompt T-7       # skill + task + decisions + graph slice + scoped servers
# … build it, run its done-check, commit on swarm/T-7 …
edify swarm done T-7         # the next claim may now unblock

One agent runs it in a loop; a swarm runs it in N worktrees at once; people on different machines add --git, and the remote's edify-swarm branch becomes the lock. Leases are phase-gated and overlap-safe: a task whose files, or their dependents, touch a live lease waits.

spec, plan, tasks, read one after another A grid of failing red squares turning to passing blue, one by one
/spec · /plan · /tasks — three reads, one after another the /tasks run — red to green

The command files live once in .edify/commands/, with thin pointers under each runtime's command directory. Skills are mirrored the same way (.claude/skills/<name>/SKILL.md points at .edify/skills/<name>.md); edify skills sync rewrites the mirror. Verification comes first: phase 2 of every build turns each spec assertion into a failing test before any core logic exists, so a cheap builder is hitting a target, not judging its own work.

The graph

An infinity traced dot by dot — the animation the EDIFY CLI paints while it works

Every file, module, exported symbol, route and table, with the edges between them, as sorted checksummed text — a rebuild over unchanged source is byte-identical. It is never written by a model: an uncovered language has no nodes and meta.languages_uncovered says so. Python is parsed with a real syntax tree; the rest by a line scanner, or by Universal Ctags with --extractor ctags.

edify graph build                     # full extraction
edify graph where isExpired           # → src/auth/clock.ts:12  symbol  isExpired
edify graph dependents createSession  # everything that breaks if it changes
edify graph overlap "a.ts b.ts" "c.ts"# can these two tasks run at once

Built for repositories no context window holds. No agent reads a million-line graph, so the graph summarises itself:

command answers
edify graph map MAP.md: subsystems, hubs, and the bridges parallel work collides on
edify graph context "<task>" the slice one task needs — files with line ranges, one hop of edges — under a token budget
edify graph explain <x> · path <a> <b> every edge of one node, or how two connect
edify graph update --changed · hook re-extract only what git says changed; git hooks keep it current

Exit codes: 0 answered, 3 ambiguous (pin with file::name), 4 not found. Every other command — check, skills, mcp, governance, doctor, license, feedback, self — is in the CLI reference.

What lands in your repository

CLAUDE.md, AGENTS.md, …  # ~15 lines each: what this repo is, the commands, the graph
.edify/
  graph/                 # nodes.tsv, edges.tsv, MAP.md — mechanical, regenerable
  skills/                # the subset your stack matched, plus index.tsv
  commands/              # the four command files
  formats/               # the three format contracts and their worked examples
  mcp.md                 # the server registry: pinned, scoped per role and phase
  conventions.md         # scraped from lint/format/tsconfig/CI — no model call
  governance.tsv         # every file above: origin, provenance, licence, hash
specs/<feature>/         # spec.md · plan.md · tasks.md · verification.md

edify governance verify reports any installed file that changed since it landed. Nothing blocks: edify check prints, and --exit-code is how you turn it into a gate you own.

Editions, privacy, and what EDIFY does not do

$0, filling with blue ink

The public edition is the whole product: MIT-licensed, free, no account, any repository size. Its one cap: the MCP registry scopes a single server into spawns. The team edition, edify-agents-teams, is by licence and lifts it; a licence is a signed token verified locally, issued on request. edify license status says which edition this machine runs.

There is no telemetry — not anonymous, not aggregate, not opt-out. edify feedback writes four answers to a file on your machine and prints how to send it, if you choose.

EDIFY does not wrap the agent runtime, host a model, run a server or daemon, or enforce anything at runtime. It makes a competent model efficient; it does not make a weak one competent. On a small solo project a frontier model in auto mode is already excellent — the value starts where the codebase stops fitting in the context window.

Docs and development

Quickstart · CLI · Graph · Agents · Verification · Privacy · Troubleshooting · Changelog · Contributing

git clone https://github.com/EDIFY-agents/EDIFY && cd EDIFY
pip install -e ".[dev]"
pytest

The suite checks that this repository satisfies what it asks of yours: every skill file valid and under budget, every command file short, every worked example parsing against its format.

Licence

See LICENSE. Skill files carry their own provenance and licence in frontmatter; entries marked provenance: adapted are governed by their upstream licence.

Metadata

Release files for edify-agents-cli 0.3.1

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for edify-agents-cli 0.3.1
File Size Uploaded Provenance Actions
edify_agents_cli-0.3.1.tar.gz 281.1 kB With provenance Details for edify_agents_cli-0.3.1.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for edify-agents-cli 0.3.1
File Interpreter ABI Platform Actions
edify_agents_cli-0.3.1-py3-none-any.whl Python 3 none any Details for edify_agents_cli-0.3.1-py3-none-any.whl

Total release size: 557.0 kB

Release files / edify_agents_cli-0.3.1.tar.gz

Download URL edify_agents_cli-0.3.1.tar.gz
Size 281.1 kB
Tags Source
SHA-256 checksum
How to use checksums
22682d5c8dcf69dfe75947a41a9bd8afad69370af7090484c5c91a240e4d46b7
BLAKE2b-256 checksum
How to use checksums
957603d7ea680f918ef2073c008feeadd88e5e91e4e4aca46afc69f9161c22d8
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 19, 2026.

Transparency log

Release files / edify_agents_cli-0.3.1-py3-none-any.whl

Download URL edify_agents_cli-0.3.1-py3-none-any.whl
Size 275.9 kB
Tags Python 3
SHA-256 checksum
How to use checksums
61c47b08e0374d21d78cb7bbc1c532ff9acabc2832f9b5e0fdb9f52f9dfac320
BLAKE2b-256 checksum
How to use checksums
51dd0edd21d85d79fbb661cd66333325ffb5a5931fa3dea141a6f50aa408252b
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 19, 2026.

Transparency log

Security for edify-agents-cli 0.3.1

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

Full provenance

Both files in this release have been uploaded with provenance metadata.

Select from 2 files

Attestation for edify_agents_cli-0.3.1.tar.gz

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 19, 2026.

Transparency log

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page