Open source · Apache-2.0 · claude + codex adapters

Your AI agent is brilliant.
It just forgets everything.

Coding OS is the cognitive operating system that gives AI coding agents memory, structure, and discipline — a knowledge graph, a task board, and an enforced engineering loop, on top of the agent you already use.

uv tool install coding-os && cos init
~78%
fewer tokens on structural queries
3
cognitive layers
20+
stack templates
The problem

Three failure modes. Three fixes.

thinking_os

Amnesia

Every session starts from zero. Yesterday's decisions are gone.

Cross-session memory: observations, learned patterns, and a session digest the next agent inherits.

graph_os

Token burn

The agent greps and re-reads half the repo to answer one structural question.

A precomputed knowledge graph answers “who calls this / what breaks” in one query.

board_os

Chaos

It codes before it thinks, skips the plan, and claims “done” without running anything.

A complexity gate + Scrumban board + verification matrix, enforced by hooks — not vibes.

The measurement

“Who references ContentType?” — asked of Django

grep, then read around the top matches0

tokens — and no way to know what the grep missed

One cos_graph_references call0

tokens — all 64 references, with the count to prove it

~78% fewer tokens

Median across the 10 highest-degree symbols in each of four public repos (requests, FastAPI, Django, coding-os) for references and rename_plan, measured against what a competent agent actually does — grep, then read around the top matches. Not against reading every file, which is the comparison that produces a prettier number nobody believes.

Where it loses: a 3-hop blast radius costs more than reading on mid-size repos (−7% on FastAPI), and on a small repo bare grep beats it outright. The harness publishes those rows too, and it refuses to score an answer whose traversal was cut short.

The architecture

One kernel, three brains

The Complexity Gate routes every request through a disciplined loop — and hooks enforce it, so the agent can't skip a step.

  1. Classify
  2. Orient
  3. Plan
  4. Execute
  5. Verify

graph_os

Navigate

Extractors index code, docs, and configs into a queryable graph: callers, blast radius, contracts, rename plans.

thinking_os

Remember

Memory, learning, and metrics. Sessions record observations; the next session starts with a digest, not amnesia.

board_os

Plan

A Scrumban board with WIP limits, definition-of-ready, and lifecycle hooks. Tasks are contracts, not vibes.

No overclaiming

Honest support matrix

“Agent-agnostic” is a design goal, not a marketing claim. Today two adapters ship: claude and codex.

RuntimeHook coverageWhat that means
Claude CodeFullEvery hook fires — gates, skills, doc-anchors, and verification. The engineering loop is enforced end to end.
Codex (CLI & app)FullFull hook parity — Bash, file edits, MCP tool calls, prompt context, and session lifecycle all fire. Same gates, same enforcement.
HumansGit hooksInstall the git hooks and the commit contract holds; the board and graph are agent-optional — ideal for mixed human/agent teams.
Bring your stack
Next.jsFastAPIDjangoGoLaravelRailsNestJSFlutterSpring BootRustVue / NuxtSvelteKitAngularAstro

Give your agent a memory it can't lose.

Stack templates ship playbooks, rules, and skills per framework. Install in one command and keep the agent you already use.