FORGHD — The operating system for prompt engineering

The Lifecycle

From draft to production. In one continuous pipeline.

Every prompt has a story — author, branch, review, deploy, observe. FORGHD captures all of it so nothing ships unsigned and nothing gets lost in Slack.

01 / Author

Draft with intent

Rich editor with variables, tests, and metadata. Schema-first authoring keeps prompts predictable.

02 / Branch

Fork, edit, propose

Branches and PR-style proposals. Diff every token. Squash, rebase, and merge — without leaving the canvas.

03 / Review

Review & approve

Threaded feedback, suggestions, and required sign-offs on every proposal — with a full audit trail from idea to merge.

04 / Firewall

Gate the release

Block unsigned, untested, or unapproved prompts before they reach production. Policy as code.

05 / Deploy

Promote in one call

Targeted rollouts to environments and tenants. Roll back atomically when telemetry says so.

06 / Observe

Live performance

Latency, cost, completion quality, eval scores — all per version, per env, in one timeline.

Built for serious teams

Everything a prompt needs
between idea and production.

Versioning & Branching

Treat prompts like code.

Branch off any version, propose changes, and merge with a full diff. Every commit is signed, traceable, and reversible. Roll forward and back with a single API call.

a3f9c1refine system role@maya

b7e210add safety eval@kio

ƒ mergeinto main@maya

c92044tighten temperature@dev

d8aa3bv1.4 → v1.5@bot

e1b7c0production tag@kio

Prompt Firewall

Nothing unsigned ships.

Policy gates between staging and prod. Untested or unapproved versions fail loud at deploy.

PASScheckout/refund-summaryv2.4

BLOCKsupport/auto-replyv1.8

HOLDonboarding/welcomev3.0

PASSsearch/rerankv0.9

Analytics

See what's working.

Latency, cost, eval scores per version. A/B without leaving the platform.

312ms

p50 latency

$.21/k

cost / req

94%

eval score

MCP bridge

Your registry in every MCP client.

Serve FORGHD over the Model Context Protocol so assistants can list prompts, pull pinned versions, and run against policy — without copy-paste or drift.

Connected clients

CL

Claude Desktop

live

CU

Cursor

live

CH

ChatGPT

live

CN

Cline

live

Prompt Firewall

The release gate
your prompts never had.

Define the policy once — required tests, required reviewers, schema validation, max cost, allowed models, redaction rules. Every deploy hits the wall before it hits prod.

Inbound deploys

v2.4refund-summarysigned

v1.8auto-replyeval

v3.0welcomeno review

v0.9reranksigned

FW.policy.v3

Block first, ship after.

Policies live in YAML alongside your code. Diff-on-PR. Replayable on any historical deploy.

✓ require: 2 approvers from #ml-team

✓ tests.coverage ≥ 0.85 over golden set

✓ deny: PII regex match in system role

✓ max.cost.per.run ≤ $0.012

To production

apiprod-eu-1live

apiprod-us-1live

apicanary-5%ramp

edgeworkerslive

Teams & Workflow

Built for how teams actually ship.

Approval workflows, shared context, and direct MCP bridges to the LLMs your team already uses — so prompts stop being one engineer’s secret weapon and start being team infrastructure.

Approval workflow

Configurable reviewers per project, branch protection, and required checks. Merging without sign-off is impossible.

Shared context

Variables, schemas, golden datasets and snippets live in one workspace. Reference them from any prompt; updates ripple everywhere.

MCP for any LLM soon

Connect Claude, ChatGPT, Cursor, Cline — any MCP-aware client — straight to your prompt registry. Edit, fetch, and run live without leaving the chat.

Roles & permissions

Granular RBAC — author, reviewer, deployer, viewer — scoped per project, env, or tenant. SAML SSO included.

Gated ship path

Review once, promote once.

Every proposal carries owners, checks, and pinned context. The card below is the same structure we attach to merges — so approvers see policy, datasets, and MCP bridges before anyone touches prod.

See the underlying approval flow →

refund-summary / propose

2 of 2 approved

MA

Tighten refund tone & add JSON schema

@maya · feature/v2.4 → main · 4 files changed

+12 −4 2 tests added

✓

@kio approved

reviewer · ml-team

2m ago

✓

@anya approved

reviewer · safety

just now

✓

Eval suite passed

checks · golden-set

94.2%

⟳

Firewall policy check

policy · FW.v3

running…

·

Promote to prod

deployer required

queued

@contextrefund-policy.mdpinned

@datasetgolden/refunds.v6 · 1,240 casespinned

@schemaRefundDecision.jsonpinned

MCP bridges connected

CL

Claude Desktop

live

CU

Cursor

live

CH

ChatGPT

live

CN

Cline

live

Product tour

Motion first. Then the panels you live in.

The recording walks an end-to-end change in the real product. The screenshots are the static views reviewers and auditors grab today — timeline, diff, merge, and editor — each in its own tab so nothing is buried under a single “demo” switch.

Screen recording

From draft to signed version in one continuous session.

Follow a refund prompt as it moves through edit, diff against what prod pinned, three-way merge, and the approval queue. This is the path every environment enforces before traffic sees new tokens.

Full session capture — edit · diff · merge · firewall queue

Screenshots

Frames for the people who sign off and file tickets.

Each still maps to a real workflow: shipping history you can diff, line-level changes you can revert, merge decisions you can defend, and authoring with variables and versions in one surface.

Every save is a version. Select any two points to diff or merge; prod pins never drift quietly.

Line-level changes with one-click restore — audit-friendly without round-tripping through a repo.

Resolve each conflict from base, yours, or theirs — then land a single merged prompt with a clear provenance trail.

Typed variables, AI assist, and the versions rail stay in one frame so authors never leave the canvas to ship.

Developer SDK

Slot it in.
Three lines is enough.

Drop FORGHD into any service. Fetch the active version, run with telemetry, and let the firewall enforce policy at the edge — all with one client.

Drop-in client

Python, TypeScript, and Go SDKs. Zero-config caching. Streaming support.

Edge-ready telemetry

Every call ships latency, tokens, cost, and eval signal back without blocking your hot path.

Provider-agnostic

OpenAI, Anthropic, Google, open-source — switch model without touching your prompt source.

app.ts

app.py

curl

forghd@1.0.4

// Fetch the production-pinned version, run, ship telemetry.
import { forghd } from "@forghd/sdk";

const fg = forghd({ project: "checkout" });

export async function summarizeRefund(order: Order) {
  const prompt = await fg.prompt("refund-summary", { env: "prod" });

  // Firewall runs server-side — denies on policy violation.
  const { text, runId } = await prompt.run({
    model: "claude-haiku-4-5",
    vars: { order },
  });

  return { text, runId }; // runId joins the live timeline.
}

# Fetch the production-pinned version, run, ship telemetry.
from forghd import Forghd

fg = Forghd(project="checkout")

def summarize_refund(order):
    prompt = fg.prompt("refund-summary", env="prod")

    # Firewall runs server-side — denies on policy violation.
    result = prompt.run(
        model="claude-haiku-4-5",
        vars={"order": order},
    )

    return result.text, result.run_id

# One request — version resolution, run, telemetry.
curl https://api.forghd.dev/v1/run \
  -H "authorization: Bearer $FORGHD_KEY" \
  -H "content-type: application/json" \
  -d '{
    "project": "checkout",
    "prompt":  "refund-summary",
    "env":     "prod",
    "model":   "claude-haiku-4-5",
    "vars":    { "order": { "id": "ord_3F2k" } }
  }'

The operating system
for prompt engineering.

From draft to production. In one continuous pipeline.

Draft with intent

Fork, edit, propose

Review & approve

Gate the release

Promote in one call

Live performance

Everything a prompt needs
between idea and production.

Treat prompts like code.

Nothing unsigned ships.

See what's working.

Your registry in every MCP client.

The release gate
your prompts never had.

Inbound deploys

FW.policy.v3

Block first, ship after.

To production

Built for how teams actually ship.

Review once, promote once.

Motion first. Then the panels you live in.

From draft to signed version in one continuous session.

Frames for the people who sign off and file tickets.

Slot it in.
Three lines is enough.

Make "it shipped the wrong prompt"
impossible.

The operating system for prompt engineering.

From draft to production. In one continuous pipeline.

Draft with intent

Fork, edit, propose

Review & approve

Gate the release

Promote in one call

Live performance

Everything a prompt needs between idea and production.

Treat prompts like code.

Nothing unsigned ships.

See what's working.

Your registry in every MCP client.

The release gate your prompts never had.

Inbound deploys

FW.policy.v3

Block first, ship after.

To production

Built for how teams actually ship.

Review once, promote once.

Motion first. Then the panels you live in.

From draft to signed version in one continuous session.

Frames for the people who sign off and file tickets.

Slot it in. Three lines is enough.

Make "it shipped the wrong prompt"impossible.

The operating system
for prompt engineering.

Everything a prompt needs
between idea and production.

The release gate
your prompts never had.

Slot it in.
Three lines is enough.

Make "it shipped the wrong prompt"
impossible.