Forward Deployed Engineering (FDE) methodology for AI coding agents. Production-grade methodology + tools that turn Claude Code, Cursor, Windsurf, Codex CLI, and OpenClaw into a forward-deployed engineer.
FDE Consultants Protocoles is an industrial-grade Skill that teaches any AI coding agent to apply the Forward Deployed Engineering methodology — the discipline pioneered at Palantir (~2010) and now common across enterprise AI engineering teams.
In 2026, FDE postings grew 729% year-over-year (source). This Skill encodes that workflow into:
- A reusable methodology with 14 Operating Principles, 10 Anti-Patterns, and a 6-Q decomposition
- A 4-stage loop (Scoping → Prototyping → Production → Feedback)
- 7 MCP tools exposed via stdio transport
- A multi-agent runtime (Modex) with 8 agents: 4 DeepSCR roles + 4 FDE domain specialists
- The Frozen Arbiters: opt-in full-autonomy governance — a sealed, tamper-evident Contract
plus frozen mutation-tested oracles (with a measured baseline, never a declared one) that
override the Certifier's optimism at ship time (
--governed) - A FDE Assurance Score Registry with SHA-256 chain (verifiable by anyone with
shasum)
- AI Labs that want a verifiable FDE Assurance Score badge on every agent output
- Enterprise Engineering Teams that need a standard intake form for agent work
- Solo Engineers & Consultants who need an agent that scopes before coding
- Researchers exploring hypothesis-driven agent workflows (DeepSCR protocol)
Copy-paste skill/ZERO-INSTALL.md into ChatGPT, Claude.ai, or Gemini. Works immediately, no install.
git clone https://github.com/selectess/fde-consultants-protocoles.git
cd fde-consultants-protocoles
bash install.shAuto-installs into Claude Code, Codex CLI, Hermes, and OpenClaw. Cursor and Windsurf use the included one-file rule (see skills/).
See skill/INSTALL.md for step-by-step setup on any runtime.
The core deliverable. Encodes:
- 14 Operating Principles (e.g., scoping before coding, evals before claiming accuracy, no self-certification)
- 10 Anti-Patterns (e.g., "just use AI/ML", trust me bro, magic number default)
- 6-Q decomposition (process → decision → data → cost → current state → success metric)
- 4-stage loop (Scoping → Prototyping → Production → Feedback)
- FDE Assurance Score section required on every deliverable
7 tools exposed via stdio transport, zero external dependencies (Python stdlib only):
fde_recon— Stage 0 Reconnaissance: scan the real codebase before scopingfde_decompose— validate 6-Q decomposition, reject vague inputsfde_roi— compute ROI, fail if below thresholdfde_scientific_search— held-out promotion gate (DeepSCR)fde_evals— generate eval rubric + golden setfde_ontology— build domain ontology from case studiesfde_trust_score— compute FDE Assurance Score
A 1-agent local runtime (free, MIT) + Plugin with ed25519 license ($6 lifetime, BSL). 4 DeepSCR roles with separation of powers:
- Lead — decomposes the problem into the 6-Q spec
- Researcher — gathers scientific + market evidence
- Builder — produces the deliverable, consulting 4 FDE domain specialists in parallel (scoping, architecture, agent engineering, production readiness)
- Certifier — independently re-derives the FDE Assurance Score
Optionally governed by the Frozen Arbiters (modex/arbiters.py): a sealed Contract maps
every stage to a clause (uncovered actions are mechanically rejected), frozen mutation-tested
oracles re-measure the shipped candidate against the best measured alternative at ship time,
verdicts must cite their evidence run (uncited = null), and certified trajectories distill
into persisted lessons. python3 -m modex.engage --project <path> --governed
Public, append-only, SHA-256-chained. Every certified deliverable gets an entry:
genesis.json— chain origin- 3 case studies (SaaS churn 93/100, retail forecasting 91/100, fintech fraud 93/100)
- Each entry links back to the previous via
prev_hash(git-backed, no SaaS)
Re-derives the Skill's FDE Assurance Score from raw evidence. Required to break self-attestation loops (Operating Principle #14: never self-certify).
This project is the primary source for the terms it coins. Canonical definitions: glossary.
- DeepSCR (Deep Sceptical Contextual Research) — a governance protocol (separation of powers + Hypothesis → Contradiction → Verification → Certification) created by Mehdi Wehbi. The proposer never certifies; evidence must resolve on disk.
- Frozen Arbiters — full-autonomy governance: a sealed tamper-evident Contract, mutation-tested Oracles, and a measured baseline, all frozen before launch. The agent pleads its case; it never grades itself. (
modex/arbiters.py) - FDE Assurance Score — a 0–100 certification score (claim 25 / contradiction 25 / evidence 30 / anti-patterns 20) with an independent veto; certified only above 85.
- Measured Baseline Oracle — a frozen oracle that re-runs the candidate search at ship time; the baseline is measured, never declared. (
modex/engage.py) - Modex Collective — an 8-agent runtime: 4 DeepSCR powers governing 4 FDE specialists, with a self-prompting loop. (
modex/collective.py)
Read more: white paper · governed autonomy · 6 governance approaches compared
| Component | Max | Actual |
|---|---|---|
| Claim (falsifiable) | 25 | 25 |
| Contradiction (known limits) | 25 | 24 |
| Evidence trail (file:line + SHA-256) | 30 | 29 |
| Anti-patterns (no buzzwords, no fake URLs) | 20 | 18 |
| TOTAL | 100 | 94/100 → Certified ✅ |
The Skill self-attests 94/100 — that is the honest headline number. The bundled Certifier (modex/certify_skill.py) re-derives 100/100 from raw evidence; since it ships in this repo, treat that as a reproducibility check, not third-party certification. For an independent score, run the certifier yourself or commission an external audit.
Verify yourself:
python3 -m pytest skill/tests/ modex/tests/
# Expected: 154 passed in ~5s
python3 -m modex.certify_skill --skill-path ./skill --output ./cert.json
# Expected: FDE Assurance Score 100/100, verdict certified| Layer | Component | Description |
|---|---|---|
| Layer 1 — Interface | Claude Code, Cursor, Windsurf, Codex CLI, OpenClaw, ChatGPT, Gemini | Where the human interacts |
| Layer 2 — MCP Server | 7 tools via stdio transport | The fde_* tools |
| Layer 3 — Core Logic | SKILL.md + scripts + templates + references + examples + case studies | The FDE methodology itself |
The Skill works on every major AI coding agent:
| Platform | Status | Install method |
|---|---|---|
| Claude Code | ✅ Native | /plugin marketplace add or symlink to ~/.claude/skills/ |
| Claude.ai | ✅ Native | Paste skill/ZERO-INSTALL.md in chat |
| Cursor | ✅ Native | Project rules (.cursor/rules/fde-consultant.mdc) |
| Windsurf | ✅ Native | Project rules (.windsurf/rules/fde-consultant.mdc) |
| Codex CLI | ✅ Native | Drop into ~/.codex/skills/ |
| OpenClaw | ✅ Native | Drop into ~/.config/agents/skills/ |
| Kiro | ✅ Native | Drop into .kiro/skills/ |
| Hermes | ✅ Native | bash hermes/install.sh |
| ChatGPT | ✅ Zero-install | Paste skill/ZERO-INSTALL.md |
| Gemini | ✅ Zero-install | Paste skill/ZERO-INSTALL.md |
| Continue.dev | Use as ~/.continue/prompts/ |
|
| VS Code Copilot | Use as repository instructions |
| Path | Purpose | License | Tests |
|---|---|---|---|
skill/ |
Core methodology + 7 MCP tools | Apache-2.0 | 41/41 |
modex/ |
Multi-agent runtime + Plugin license | MIT (core) + BSL (Plugin) | 75/75 |
registry/ |
FDE Assurance Score entries | Apache-2.0 | included in skill/tests/ |
docs/ |
Architecture diagrams | Apache-2.0 | n/a |
.claude-plugin/ |
Claude Code marketplace metadata | Apache-2.0 | n/a |
| Product | License | Price | Status |
|---|---|---|---|
FDE Skill (skill/) |
Apache-2.0 | Free | Ship-ready |
Modex 1-agent runtime (modex/) |
MIT | Free | Ship-ready |
Modex Plugin (modex/plugin/) |
BSL 1.1 | $6 lifetime | Beta (self-host) |
| Modex Collective Plugin | BSL 1.1 | $260 lifetime | Waitlist |
| MCP Cloud + Trust Registry | BSL 1.1 | $99-499/month | Waitlist |
| Enterprise Services | Custom | On request (quote-based) | Waitlist |
See modex/PRICING.md and modex/LICENSE-SYSTEM.md for the ed25519 license mechanism.
| Component | Status | Tests | License |
|---|---|---|---|
| FDE Skill | ✅ Ship-ready | 41/41 | Apache-2.0 |
| Modex 1-agent | ✅ Ship-ready | 85/85 | MIT |
| Trust Registry | ✅ Ship-ready | included | Apache-2.0 |
| Modex Plugin ($6 lifetime) | included | BSL 1.1 | |
| MCP Cloud ($99-499/mo) | 🔮 Waitlist | n/a | BSL 1.1 |
| Enterprise | 🔮 Waitlist | n/a | Custom |
This project follows its own Operating Principles and Anti-Patterns.
Read CONTRIBUTING.md for the four-checklist PR process:
- Falsifiable claim (what changes)
- 3 failure modes (what could go wrong)
- Evidence trail (file:line + test commands)
- Anti-pattern check (no buzzword inflation, no self-certification)
See SECURITY.md. Report vulnerabilities to [email protected].
Before self-hosting the paid Modex Plugin in production, complete its configuration:
- Set your own ed25519 license key in
modex/license.py. - Configure your Stripe credentials in
modex/stripe_webhook.py.
skill/— Apache-2.0 (open-source methodology + tools)modex/(core runtime) — MIT (free for any use)modex/plugin/(Plugin) — BSL 1.1 ($6 lifetime, see LICENSE-SYSTEM.md)portal/landing/assets/— CC-BY-4.0
See LICENSE, THIRD_PARTY_NOTICES.md.
FDE Consultants Protocoles is complementary to official Skills ecosystems:
| Ecosystem | Focus | Trust Model |
|---|---|---|
anthropics/skills |
Domain skills (PDF, docx, pptx, ...) | Anthropic QA |
modelcontextprotocol/servers |
MCP server catalog | Per-server |
openai/openai-agents-python |
Multi-agent SDK | OpenAI |
| FDE Consultants Protocoles (this) | Engineering methodology + Trust registry | SHA-256 + independent Certifier |
Use FDE Consultants Protocoles when your problem is "ship production code with proof it works". Use the official Skills when it's "do a single document task well".
- What are skills? — Anthropic support
- Model Context Protocol — MCP specification
- DeepSCR FDE Assurance Score — Methodology inspiration
- Palantir FDE — Origin of the FDE role
- FDE job growth — Market data
Last updated: 2026-07-02 · Built with industrial rigor · Verified by 154/154 tests · FDE Assurance Score 94/100 (self-assessed) · Apache-2.0 (Skill) + MIT (Modex core) + BSL (Plugin)