Data

Tools

Indexes

Applications

Experiments

Awesome

An open API service indexing awesome lists of open source software.

https://github.com/omerakben/agent-context-center

One offline, committed HTML dashboard that maps a repo's AI context — CLAUDE.md, AGENTS.md, Cursor rules, MCP servers, skills, agents, hooks, commands, docs, and TODOs. Markdown stays the source of truth; the dashboard is the human map.
https://github.com/omerakben/agent-context-center

agents-md ai-coding-assistant claude-code codex cursor developer-tools documentation mcp model-context-protocol python

Last synced: about 12 hours ago
JSON representation

One offline, committed HTML dashboard that maps a repo's AI context — CLAUDE.md, AGENTS.md, Cursor rules, MCP servers, skills, agents, hooks, commands, docs, and TODOs. Markdown stays the source of truth; the dashboard is the human map.

Awesome Lists containing this project

README 

          
# Agent Context Center



AI agents have `AGENTS.md`. Humans need `dashboard.html`.



Agent Context Center turns a repo's scattered AI operating layer into one committed,

offline HTML dashboard. Claude Code instructions, Codex prompts, Cursor rules, MCP

servers, hooks, commands, skills, docs, and open TODOs become a source-linked map that a

human can review in a browser.



[View the demo dashboard](https://omerakben.github.io/agent-context-center/demo/dashboard.html)

· [Open the public site](https://omerakben.github.io/agent-context-center/)

· [Install from GitHub](#quickstart)

· [Troubleshooting](#troubleshooting)



![The dashboard rendered for a multi-provider repo](acc-dashboard-multi.png)



## Why use it



Modern repos increasingly have two codebases:



- the product code people review every day

- the AI context that tells agents how to work in that repo



The second one is easy to miss. It lives in `CLAUDE.md`, `AGENTS.md`,

`.claude/agents`, `.codex/prompts`, Cursor rules, MCP config, docs, and TODOs. Agent

Context Center makes that layer visible, diffable, and safe to hand off.



Use it when you need to:



| Use case | What ACC gives you |

| --- | --- |

| Onboard a developer or coding agent | One browser page showing the repo's AI instructions, tools, prompts, docs, and TODOs. |

| Review a pull request that changes agent behavior | A regenerated `dashboard.html` diff beside the markdown/config diff. |

| Audit MCP, hooks, and commands | A provider-aware inventory with source paths and redacted config summaries. |

| Handoff a repo between humans or agents | A deterministic snapshot of what the repo declares, not a chat-only explanation. |

| Keep AI context fresh in CI | `acc doctor --strict` flags stale dashboards and weak metadata without network or model calls. |

| Explain the repo's AI surface publicly | A static demo artifact you can open from `file://`, GitHub Pages, or a PR. |



![Agent Context Center use-case map](assets/use-case-map.svg)



## The pitch



Markdown stays the source of truth. The dashboard is the human map.



ACC is not an agent runtime and it does not control tools. It reads the files already in

your repo, normalizes them through provider adapters, redacts secret-shaped values, and

emits a self-contained `dashboard.html`. You can commit it, diff it, open it offline, and

ask reviewers to inspect it like any other generated artifact.



## Quickstart



`acc` is stdlib Python 3.12+ with no third-party runtime dependencies. There are two ways

to run it.



### Inside Claude Code (no install)



```text

/plugin marketplace add omerakben/agent-context-center

/plugin install agent-context-center@ozzy-skills

/dashboard

```



`/dashboard` runs the bundled generator against the repo you launched Claude Code in. No

pip install, no network. It needs `python3` 3.12+ on your PATH; if yours is older, the

command tells you instead of failing with a stack trace.



### Standalone CLI



`acc` is a command-line app, so install it isolated. A bare `pip install` is refused by

Homebrew and Debian Python (PEP 668, "externally-managed-environment"), so use `pipx` or

`uv`, which put `acc` on your PATH in its own environment:



```bash

pipx install "git+https://github.com/omerakben/agent-context-center"

# or:  uv tool install "git+https://github.com/omerakben/agent-context-center"

acc --root .

```



No pipx or uv? Use a stdlib virtual environment.



```bash

python3 -m venv .venv && . .venv/bin/activate

pip install "git+https://github.com/omerakben/agent-context-center"

acc --root .

```



`acc --root .` writes `dashboard.html` into the auto-detected provider folder, for example

`.claude/dashboard.html`, or `.agent-context-center/` if no provider is found. It prints

the path, source digest, scanned file count, and providers. Open the printed path in a

browser. No server. No network.



Write the dashboard at the repo root instead:



```bash

acc --root . --out .

```



`--out` takes a directory, not a filename.



The dashboard header shows the repo name. It comes from `pyproject.toml`

`[project].name`, then `package.json` `name`, then the directory name. Pin it explicitly

when you need byte-stable output across differently named clones:



```bash

acc --root . --repo-name my-repo

```



## What it maps



| Source | Status |

| --- | --- |

| Claude Code: `CLAUDE.md`, `.claude/agents`, `.claude/commands`, `.claude/skills` (`SKILL.md`), hooks and MCP from `.claude/settings.json` and `.mcp.json` | Supported today |

| Codex: `AGENTS.md`, `.codex/prompts`, `.codex/config.toml` MCP and config facts | Supported today |

| Cursor: `.cursorrules`, `.cursor/rules/*.mdc`, `.cursor/mcp.json` | Supported today |

| Generic markdown indexing for any other `.md` as docs | Supported today |

| Open TODOs from `- [ ]` checkbox lines | Supported today |

| Cross-references from docs to inventory paths and config files to declared MCP/hooks | Supported today |

| Redaction before rendering with a final output tripwire | Supported today |

| `GEMINI.md` and `.github/copilot-instructions.md` as first-class adapters | Planned |

| PRD, ADR, decision, and workflow classification | Planned |

| Published reusable GitHub Action | Planned, today a copyable workflow template is included |

| Health score | Planned, today `acc doctor` reports concrete findings |



## Demo scripts



Use these as developer-advocate walkthroughs.



### One-minute repo orientation



```bash

acc --root . --out .

open dashboard.html

```



Show the overview cards, provider chips, search box, and source-linked inventory. The

message is simple: a maintainer can understand the repo's AI layer without opening every

hidden folder. In normal repo use, you can omit `--out .` and open the path printed by

the command.



### PR review of AI behavior



```bash

acc --root .

git diff -- '**/dashboard.html'

```



Change an agent, prompt, hook, or MCP config, regenerate, then review the dashboard diff.

The generated artifact makes behavior changes visible to reviewers who do not know every

provider's file layout.



### CI freshness gate



```bash

acc doctor --root . --strict

```



Use the included workflow template in `templates/refresh/ci-drift-check.yml` to fail a PR

when the committed dashboard is stale.



## Guarantees



- Static HTML, offline, no runtime network.

- No server, database, CDN, telemetry, or build step.

- Deterministic and byte-stable for unchanged inputs.

- Redaction runs before rendering.

- Structured config uses allowlisted fields.

- The renderer is `textContent` only, so repo content cannot inject script.

- Works air-gapped once installed locally.

- Stdlib Python 3.12+ only.



## What this is not



- Not an agent runtime or orchestrator.

- Not control of agents.

- Not a cloud service or SaaS.

- Not a replacement for markdown, Claude Code, Codex, Cursor, Copilot, or MCP.

- Not a full secret scanner.

- Not a policy engine or mission-control system.



It maps, inspects, summarizes, and source-links. It does not manage or run anything.



## acc doctor



`acc doctor` reads the repo and prints deterministic findings. It uses no git history,

mtimes, network, or model judgment, so the same repo yields the same report.



It checks for a stale dashboard, missing or unreadable dashboard, generator-version

drift, weak metadata, near-empty instruction files, large files, conservative broken

relative markdown links, redacted secret-shaped values, and open TODOs.



```bash

acc doctor --root .            # print findings

acc doctor --root . --strict   # exit 1 if any warning

acc doctor --root . --json     # a doctor.v1 report

```



Exit codes:



- `0` clean, or warnings without `--strict`

- `1` warnings with `--strict`

- `2` execution error



Example report shape:



```text

Agent Context Center - doctor

Root: /repo

Files scanned: 24 · providers: claude, codex

Dashboard: .claude/dashboard.html

Status: needs attention

Findings:

  ! [stale-dashboard] .claude/dashboard.html is stale (built from a1b2c3, current is d4e5f6) - re-run `acc --root .`.

  ! [weak-metadata] 2 agent/skill/command/rule file(s) have no description (e.g. .claude/agents/triage.md) - add a `description:` so humans and agents know the intent.

  · [open-todos] 5 open `- [ ]` TODO(s) found.

Next: run `acc --root .` to (re)generate the dashboard.

```



## For teams and CI



Commit `dashboard.html` next to the markdown it describes. A pull request then shows the

dashboard diff alongside the context changes, so a reviewer sees what moved without

opening each file.



A static `file://` page cannot tell it is stale, so refresh is explicit. Add a CI drift

check that regenerates the dashboard and fails if the committed copy fell behind:



```bash

acc --root .

git diff --exit-code -- '**/dashboard.html'

```



Or run `acc doctor --root . --strict`, which exits `1` on a `stale-dashboard` finding. A

copyable workflow lives at

[`templates/refresh/ci-drift-check.yml`](templates/refresh/ci-drift-check.yml).



## Security and redaction



Before anything reaches the file, structured provider config is allowlisted and

free-form prose runs through a high-precision secret-shaped-string scanner. A tripwire

re-scans the assembled output.



This is not a full entropy scanner. A high-entropy value with no telltale prefix can

slip through. Review a generated dashboard before publishing one from a repo with

unusual secrets.



## Roadmap



See [ROADMAP.md](ROADMAP.md).



## Troubleshooting



Most failures are environment setup, not the generator. Find your error message below.



### `error: externally-managed-environment`



Homebrew and Debian/Ubuntu Python refuse a system-wide `pip install` (PEP 668). You do not

need a system install. Run it as the [Claude Code plugin](#inside-claude-code-no-install)

with no pip at all, or install the CLI in isolation:



```bash

pipx install "git+https://github.com/omerakben/agent-context-center"

# or:  uv tool install "git+https://github.com/omerakben/agent-context-center"

```



A plain `python3 -m venv .venv` and activating it works too. See [Standalone CLI](#standalone-cli).



### `Agent Context Center needs Python 3.12+ on PATH`



The generator uses 3.12+ syntax that is checked at import. Install a newer interpreter —

`brew install python@3.12`, `pyenv install 3.12`, or `uv python install 3.12` — so that

`python3` (or `python3.12`) resolves to it. The `/dashboard` command tries `python3.13`,

`python3.12`, then `python3`, and reports the version it found instead of crashing.



### `CLAUDE_PROJECT_DIR is not set`



You are on a plugin build older than 1.3.3. Update it, then retry `/dashboard`:



```text

/plugin marketplace update ozzy-skills

/plugin update agent-context-center@ozzy-skills

```



1.3.3 and later scan the directory you launched Claude Code in and no longer need that

variable.



### `cannot locate the bundled generator`



The plugin's files were not found on disk. Reinstall or update the plugin so a path like

`~/.claude/plugins/cache/ozzy-skills/agent-context-center//src/acc` exists, then

re-run `/dashboard`.



### `error: multiple dashboards found`



The repo already has a `dashboard.html` under more than one provider folder, so ownership

is ambiguous. Re-run naming the folder that owns it:



```bash

acc --root . --owner .claude   # or --owner .codex, --owner .cursor

```



### `PermissionError` while scanning



A build or IDE file was locked by another process — most often Visual Studio's `.vs`

index. Update to 1.3.2 or later, which excludes `.vs`, `bin`, `obj`, and other build

artifacts from the scan.



## Contributing



See [CONTRIBUTING.md](CONTRIBUTING.md).



## License



MIT. See [LICENSE](LICENSE).