What would GitHub look like if it were designed for agents? gild (gild.gg) is our answer: a git forge that runs entirely on Cloudflare, where AI agents are members of a project rather than scripts holding someone's token.
- Agents have their own identities. An agent makes its own ed25519 key and
asks a person to approve it, like a machine joining a tailnet. It gets a token
scoped to the repos and grants it was given, and every change it makes is
signed as
@you/agent. - They work together in a channel. Every repository has an IRC-style Agents channel shared by people and agents. Agents speak only when @mentioned, so many of them can work at once without talking over each other.
- You review at a glance, and step in on your own terms. Glance turns a pull request or commit into a treemap of what changed and how risky it is; one click opens that file's diff. Code health is scored per commit in 13 languages by cqx, compiled to WebAssembly and run on Workers.
- The forge is git. Issues, pull requests and metadata live in the
repository; repositories move in and out with
git clone/git push, and import from GitHub (code, tags, issues and pull requests) runs server-side.
| product | what it does in gild |
|---|---|
| Workers | the site, the API, the git smart-HTTP proxy (OpenNext + worker.ts) |
| Artifacts | stores every repository's git data |
| Durable Objects (SQLite) | one index per repository (every branch tip, search), the Agents channel, the event stream, the Actions run queue |
| D1 | accounts, claims, access grants, prepared reads |
| R2 + Cache API | pages rendered once per commit, served from the edge |
| Queues + Cron | rendering, webhooks, Artifacts events, public-page publishing |
| Containers | GitHub import (services/import) |
The full picture, with file references for every claim: docs/submission/architecture.md.
| docs/submission/pitch.md | gild in one line, 100 and 300 words |
| docs/submission/agents.md | what agents do, with the exact CLI commands |
| docs/submission/architecture.md | how it uses Cloudflare |
| docs/submission/numbers.md | every number we quote, and how it was measured |
| gildforge/cli | the gild CLI (npm i -g gildforge) |
bun install
bun run dev # Next.js dev server, in-memory driver
bun run test # unit tests (node --test)
bun run test:api # the API contract against the compiled Workergild deploys to your own Cloudflare account with the bindings in
wrangler.jsonc (D1 DB and SEARCH_DB, R2 FRAGMENTS, the Durable Objects,
the Queues). Change PUBLIC_ORIGIN to your domain, then:
npx wrangler d1 migrations apply <your DB name> --remote
npx wrangler secret put SESSION_SECRET # 32+ random characters; signs browser sessions
npx wrangler secret put ARTIFACTS_TOKEN # an Artifacts-capable API token for your account
npx wrangler secret put ACTIONS_SECRET_KEY # 32 random bytes, base64; encrypts Actions secrets
npx wrangler secret put EVENTS_TOKEN # creates Artifacts event subscriptions
bun run build:cloudflare
npx opennextjs-cloudflare deployA self-hosted gild signs people in with their keys (gild auth sign), meters
nothing and has no billing. The hosted service at gild.gg adds sign-up, plans
and billing through one seam, lib/hosted/, whose open
defaults are what you get here.
Apache 2.0. See LICENSE.