ANF v0.3.0 — Agent Native Format reposition

Agent Native Format v0.3.0

The project is now Agent Native Format (ANF): a token-minimal view format that translates live system state into far fewer tokens for AI agents. The execution-contract runtime (formerly Agent Contract Protocol) stays as a secondary component.

Changed since v0.2.0

• Repositioned the repo, README, and About to lead with ANF (the view format).
• Renamed the Go module to github.com/Clawdlinux/agent-native-format.
• Retitled the paper to "Agent Native Format: Token-Efficient Execution Contracts for Autonomous Agent Tool Use" (attached).
• Superseded the old ACP naming decision.
• Patched crypto/tls CVE (GO-2026-5856) via go1.26.5.

Format

• Spec: FORMAT.md (CC BY 4.0)
• Go encoder: pkg/anf (Apache 2.0)
• Kubernetes translator: translators/kubernetes

Runtime (formerly ACP)

• Execution-contract server, MCP bridge, and policy proxy. Same benchmarks: 64-97% tool-context token reduction vs raw MCP, one round trip instead of up to 21.

Install

go install github.com/Clawdlinux/agent-native-format/cmd/acp-server@v0.3.0

Paper PDF attached.

ANF v0.2.0 — Frontier Benchmark & Paper

ACP: Intent-Resolved Execution Manifests for Token-Efficient Agent Tool Use

Paper PDF attached below. Benchmarked on Claude Opus 4.7, Sonnet 4.6, GPT-5.5, GPT-5.4.

Highlights

• 120 API calls across 4 frontier models × 5 scenarios × 2 paths (MCP/ACP)
• Provider-reported input token reduction: 26–93% depending on model/scenario
• tiktoken/cl100k_base control column for deterministic normalization
• Tool-call correctness preserved or improved (GPT-5.4 S4: 0% → 100%)
• Competitive landscape: positioned against TSCG (arXiv:2605.04107) and MCP-Zero (arXiv:2506.01056)
• BFCL (ICML 2025) selected as open benchmark standard for next track

What's in this release

• paper/acp.md — Markdown paper source (canonical)
• paper/acp.tex + paper/references.bib — LaTeX source for preprint
• benchmark/frontier/run_frontier.py — Frontier benchmark harness
• results/frontier/ — Raw JSON + summary for all 120 runs
• ACP-Paper-2026.pdf — Compiled paper (attached)

Install

go install github.com/Clawdlinux/ninevigil-acp/cmd/acp-server@v0.2.0-paper

v0.1.0-spec — Public ANF Spec Launch

ACP v0.1.0-spec — Public Spec Launch

Agent Context Protocol — one API call, complete execution context, minimal tokens.

ACP sits on top of MCP. The agent sends one intent; ACP returns one scoped manifest with auth pre-injected, ordering pre-computed, and security boundaries declared. Existing MCP servers keep working — ACP ingests their tools/list and produces token-minimal manifests for execution.

Why this matters

	MCP today (raw)	ACP on top of MCP
Round trips before first action	3–21 (measured)	1
Tokens for representative workflows	373–9,223	111–359 (64–97% reduction)
Auth in agent context	yes	never
Execution ordering	agent infers	server declares

Full benchmarks: results/2026-05-02-week3-summary.md · Full positioning: docs/positioning.md.

Install (pick one)

# Go binary, no clone (requires Go 1.25+)
go install github.com/Clawdlinux/ninevigil-acp/cmd/acp-server@v0.1.0-spec
ACP_AUTH_TOKEN=dev-token acp-server --addr :8080

# Docker (no toolchain needed)
docker run --rm -p 8080:8080 -e ACP_AUTH_TOKEN=dev-token \
  ghcr.io/clawdlinux/ninevigil-acp:v0.1.0-spec
# or, from Docker Hub:
docker run --rm -p 8080:8080 -e ACP_AUTH_TOKEN=dev-token \
  goodra007/acp-server:v0.1.0-spec

# Source build
git clone https://github.com/Clawdlinux/ninevigil-acp && cd ninevigil-acp
make build
ACP_AUTH_TOKEN=dev-token ./bin/acp-server --addr :8080

Python adapters (pip from git, no PyPI required)

pip install "git+https://github.com/Clawdlinux/ninevigil-acp.git@v0.1.0-spec#subdirectory=adapters/python/acp_common"
pip install "git+https://github.com/Clawdlinux/ninevigil-acp.git@v0.1.0-spec#subdirectory=adapters/python/acp_langgraph"
pip install "git+https://github.com/Clawdlinux/ninevigil-acp.git@v0.1.0-spec#subdirectory=adapters/python/acp_openai"
pip install "git+https://github.com/Clawdlinux/ninevigil-acp.git@v0.1.0-spec#subdirectory=adapters/python/acp_crewai"

Quickstart

curl -sS -X POST http://localhost:8080/v1/context \
  -H "Authorization: Bearer dev-token" \
  -H "Content-Type: application/json" \
  -d '{"intent":"query customer data, render report, email the team","agent_id":"demo"}' \
  | python3 -m json.tool

What's included

• Spec (SPEC.md, CC BY 4.0) — protocol definition, manifest format, threat model.
• Go server (cmd/acp-server, Apache-2.0) — intent resolver (keyword + embedding), manifest builder, auth-injection proxy.
• Go SDK (pkg/acp, Apache-2.0) — acp.NewClient(...) with token + manifest helpers.
• Python adapters (Apache-2.0) — acp_common (shared client + types), acp_langgraph, acp_openai, acp_crewai.
• Benchmark harness (benchmark/) — 5 scenarios × 50 runs/scenario, tiktoken/cl100k_base.

Status

PoC. The protocol surface is stable for v0.1; the resolver and adapters are alpha-quality and will move with feedback. File issues at Clawdlinux/ninevigil-acp/issues.

Image digest

sha256:804051ad391d2dc398b24f9f50930257672d6ada9771cf36ff5863433a1432b3 — same digest for all three tags (v0.1.0-spec, v0.1, latest).