Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 8 additions & 2 deletions CONTRIBUTING.md
Original file line number Diff line number Diff line change
Expand Up @@ -9,7 +9,7 @@ Thanks for considering a contribution — this project grew out of a personal wo
3. Copy `.env.example` to `.env` and add tokens for whichever services you're working on (see README for the full list)
4. `npm run setup` to point your local Claude Desktop config at this repo
5. Make your changes
6. Test manually against a real Claude Desktop instance (see below)
6. Test your changes (see below)

## Adding a new tool module

Expand All @@ -24,7 +24,13 @@ Keep functions small and single-purpose — one API call in, one plain-object re

## Testing your changes

There's no automated test suite yet (see the roadmap in the README), so please test manually:
An automated test suite (`vitest`) covers every tool module under `tests/`:

```bash
npm test
```

Run it before opening a PR. On top of that, please also test manually against a real Claude Desktop instance:

1. Point your local Claude Desktop config at your working copy
2. Restart Claude Desktop
Expand Down
32 changes: 18 additions & 14 deletions ROADMAP.md
Original file line number Diff line number Diff line change
Expand Up @@ -5,6 +5,8 @@ GitHub, Vercel, Supabase, Slack, Gmail, Notion, Terraform, Docker,
Postgres/MySQL, a local encrypted secrets manager, Sentry, plus project
intelligence, workflow automation, and a full security layer (redaction,
permission levels, approval gates, path protection, structured audit logs).
An automated `vitest` test suite now covers every tool module (`npm test`),
on top of manual verification against real disposable resources.
See [BUILD_LOG.md](./BUILD_LOG.md) for how we got here.

## Next up
Expand All @@ -13,12 +15,11 @@ See [BUILD_LOG.md](./BUILD_LOG.md) for how we got here.

An opt-in, local `server_attested` receipt adapter now covers only `ship_change`, `verify_ci_fix`, and `deploy_project`. Before treating it as production infrastructure, key custody, rotation/revocation, multi-process persistence, and operational deployment policy still need explicit designs. The receipt layer remains evidence only and does not participate in Causly authorization.

### Deferred: Azure / GCP / AWS — direct cloud infrastructure
### Planned: Azure / GCP / AWS — direct cloud infrastructure

**Deliberately out of scope for now.** Testing any of these properly
requires a live cloud subscription and burns real credit/cost the moment
a key is generated — not worth it until there's an actual need driving it.
Revisit if/when that need shows up, not on a fixed timeline.
Direct cloud-provider tools are on the roadmap and will be built — they're sequenced after the
hosted server (below) rather than immediately, since testing them properly requires a live cloud
subscription and real cost the moment a key is generated.

- Account/subscription/resource-group listing
- Compute: list, get, deploy, restart
Expand All @@ -29,13 +30,22 @@ Revisit if/when that need shows up, not on a fixed timeline.
Note: Terraform tools (already shipped) cover a good chunk of this
indirectly — provisioning and destroying cloud resources via `plan`/
`apply`/`destroy` works today, independent of any direct-SDK cloud
integration. Direct cloud tools would add runtime introspection/debugging
integration. Direct cloud tools will add runtime introspection/debugging
that Terraform alone doesn't give you.

### Hosted MCP server
### Planned: Hosted MCP server (Causly Hosted)

Open-source local server is feature-complete for the current tool set.
Focus shifts to building the hosted version.
A managed, hosted version — run for you instead of on your own machine — is in active
development, with an early-access waitlist open today. Focus shifts here next.

### Planned: Local AI runtime environment

A dedicated, lightweight execution environment for the AI to run in locally — not a VM, not
Docker. Purpose-built, scales up and down with load automatically, and isn't tied to a specific
language runtime (the server itself is Node today, but this environment is independent of that
choice). Design and build not yet started; this is a confirmed future direction, distinct from
the `npx -y causly-server` setup installer under "After that" below.

## After that

Expand All @@ -49,12 +59,6 @@ Focus shifts to building the hosted version.
`causly://deployment/{id}/status`, `causly://sentry/{project}/issues`) —
same pattern as the existing project resources, extended to the services
we already talk to.
- **Automated test suite** — mocked API tests for every service module,
plus unit tests for the security layer (redaction, classification, path
denial) and the workflow tools. Currently verified by hand against real
disposable resources (a throwaway GitHub repo, Docker containers, DB
instances, a Sentry project — see BUILD_LOG.md) — good enough to ship,
not good enough to stay unmonitored as the surface grows.
- **`npx -y causly-server`** installer — currently `git clone` + `npm
install` + `.env` setup + `npm run setup`. Roadmap goal is a single
command that detects the environment, configures Claude Desktop, and
Expand Down
Loading