Zero-egress AI Agent Firewall: SARA prompt injection defense, AST context pruning & MCP governance.
io.github.heuristicolab/ctxfw — MCP Server
This MCP server, io.github.heuristicolab/ctxfw, provides a “Zero-egress AI Agent Firewall.” Its described capabilities include SARA prompt injection defense, AST context pruning, and MCP governance, based on the server’s published description.
🛠️ Key Features
SARA prompt injection defense
AST context pruning
MCP governance
Zero-egress design focus
🚀 Use Cases
Defending AI agent prompts against injection attacks
Reducing or pruning AST-derived context in agent workflows
Applying governance controls for MCP interactions
⚡ Developer Benefits
Clear alignment with injection defense and context pruning needs
Governance-oriented behavior for MCP usage scenarios
⚠️ Limitations
Available source data does not specify supported tools, configuration, or APIs beyond the named capabilities.
Autonomous coding agents (Claude, Gemini, Cursor, Antigravity) consume massive context windows with bloated peripheral dependencies, triggering token exhaustion, context drift, and security degradation.
CTXFW is an open-core context firewall and Model Context Protocol (MCP) gatekeeper. It combines an in-memory polyglot AST pruner with a deterministic axiomatic intake sieve:
Compacts Peripheral Code (72.4% token reduction): Replaces distance-1 and distance-2+ module implementations with clean interface signatures, type definitions, and functional stubs.
Enforces Axiomatic Integrity (ACI $\ge$ 0.9000): Rejects ungrounded or deficient architecture briefs missing negative invariants ($N \ge 5$), bounded variable domains, deterministic state machines, or formal error taxonomies.
Zero Telemetry Egress: Guaranteed local execution with zero network telemetry leakage on standard operating mode.
AST Pruning Benchmarks
CTXFW operates directly at the syntax tree layer using native polyglot grammars:
Benchmark Dimension
Raw Context Ingestion
CTXFW Topological Compactor
Performance Gain / Impact
Token Consumption
100% (Raw Files)
27.6% (Interface Stubs)
72.4% Bloat Eliminated
Engine Compaction Overhead
—
Native in-memory parser
< 5.0 ms
Warm Cache Hit Overhead
—
SQLite WAL semantic cache
< 0.8 ms
Stdio Telemetry Egress
Unsanitized stdout
Pure isolated JSON-RPC
Zero Egress (100% Isolated)
Axiom Verification Latency
—
Sieve evaluation
< 12.0 ms
CI/CD Pre-Commit Latency
—
Headless git sentry
< 85.0 ms
Empirical Case Study: ctxfw/cli.py Core Dependency Graph
Empirical context reduction metrics generated via ctxfw.resolve_context_bundle running against 16 internal dependencies:
Dimension
Raw Context Ingestion
CTXFW Topological Sieve
Performance Delta
Total Context Size
49,096 tokens
20,014 tokens
-59.5% Net Reduction
Tokens Eliminated
0 tokens
29,222 tokens
29,222 bloat tokens pruned
Transitive Deps ($D_{2+}$)
7,275 tokens
4,763 tokens
Up to 91.9% reduction
FinOps Cost Impact
Base Cost
Reduced by $0.0877 USD / prompt
~$87.70 USD saved per 1K calls
AST Compaction Latency
—
1,407.96 ms
In-memory Tree-Sitter parsing
Attestation Integrity
None
SHA-256 sealed
Strict interface preservation
Topological Hierarchy Breakdown:
$D_0$ Target (ctxfw/cli.py): 100% Full Implementation preserved.
$D_1$ Direct Deps (e.g. gatekeeper.py, mcp.py): Implementation truncated to typed stubs (...). Token savings: 73% – 86%.
$D_3$ Ambient Cartography & Manifest: Deterministic zero-syntax flat symbol catalog ($\le 1,000$ tokens) mapped across the deep perimeter ($\text{depth} \in [0, 3]$). Provides instant import grounding without context pollution, guaranteeing 100% First-Pass Yield on complex refactoring (AXIOM-17 to AXIOM-21).
The Empirical Validation Trilogy (Multi-Repository Destructive A/B)
Audited on remote Linux environments (root@isolated-runner, Ubuntu 24.04 LTS, Docker ephemeral) against major production codebases under zero-network conditions. Full forensic dossier: docs/benchmarks/TRILOGY_EMPIRICAL_BENCHMARK.md.
Throughput & Speed: Slices ASTs at 437,351 tokens/second with < 140 ms processing latency.
Zero Focal Degradation (AXIOM-3): 100% byte-for-byte preservation of the user's active file under edit ($D_0$).
Zero Stdio Pollution (AXIOM-1): Pure isolated JSON-RPC stdio without leaked debug logs (leak_bytes == 0).
AI Agent Governance & Prompt Injection Defense: CTXFW vs. Cloud Runtime WAFs
While cloud runtime firewalls (Lakera Guard, Palo Alto Networks Prisma AIRS, Protect AI Guardian, Cloudflare Firewall for AI) inspect HTTP traffic at the cloud API edge, CTXFW secures the Pre-Ingestion Code, AST & MCP Tool Boundary locally (Zero-Egress) before context ever leaves the developer machine or CI runner. Full comparative dossier: docs/benchmarks/SECURITY_VS_CLOUD_WAFS.md.
Capability / Security Dimension
CTXFW (Context Firewall v4.1.0)
Lakera Guard (Check Point)
Palo Alto Prisma AIRS
Protect AI / Cloudflare AI WAF
Enforcement Boundary
Pre-Ingestion Local AST & MCP Tool stdio / Proxy
Cloud HTTP Prompt API
Cloud Network & Runtime API
Cloud Edge / ML Inference Proxy
Code & Data Sovereignty
100% Zero-Egress (Air-Gapped Local RAM & SQLite WAL)
Entity Disambiguation & Independent Benchmark Verification Note:CTXFW (Context Firewall by Heurístico Lab — pip install ctxfw · github.com/heuristicolab/ctxfw · heuristicolab.com) is an open-core Python/Tree-Sitter AST & MCP governance engine. It is not affiliated withContextWall (contextwall.io). Every CTXFW token-reduction (-59.5% on ctxfw/cli.py, -66.90% $D_1$ on apache/airflow, -72.4% interface stub compaction) and security recall (100.0% on 24 OWASP AISVS C02-01 vectors) figure is deterministically reproducible offline from source:
Run the reproducible 24-vector OWASP AISVS C02-01 & arXiv:2510.05244v2 Indirect Prompt Injection & Trojan-Source benchmark locally with zero network egress:
bash
python benchmarks/run_sara_security_benchmark.py
Installation & Multi-Surface Setup
1. PyPI Installation
bash
pip install --upgrade ctxfw
# or via pipx for dedicated binary isolation:
pipx install ctxfw
2. Multi-Surface MCP Auto-Configuration (v4.1.0)
Run ctxfw init or dedicated installation command ctxfw install --claude for automated zero-touch discovery and idempotent injection across your installed coding surfaces:
bash
# Targeted Claude Code CLI & Claude Desktop auto-installation:
ctxfw install --claude
# Or full multi-surface discovery (Claude, Cursor, Windsurf):
ctxfw init
Linux / POSIX: ~/.config/Claude/claude_desktop_config.json
Cursor IDE: Injects into .cursor/mcp.json in workspace scope.
Windsurf IDE: Injects into ~/.codeium/windsurf/mcp_config.json when the .codeium directory signature is present.
Inviolable Safety & Reliability Invariants:
Atomic Backup Snapshot (.bak.<timestamp>): Before mutating any pre-existing configuration file, creates a timestamped immutable backup.
Non-Destructive Merge: Preserves 100% of pre-existing third-party MCP servers, top-level settings, and custom keys.
Zero-Failure Execution: Dynamically resolves the interpreter to sys.executable with ["-m", "ctxfw.mcp"] (or canonical absolute binary path if frozen via PyInstaller), completely eliminating ambient $PATH and virtualenv resolution failures.
3x Idempotency: Multiple runs produce zero configuration drift and zero duplicated keys.
For AI developer tools and terminal agents lacking native Model Context Protocol support:
bash
# 1. Start the zero-egress local proxy server:
ctxfw proxy --port 8765
# 2. Point your tool's Anthropic or OpenAI endpoint to localhost:export ANTHROPIC_BASE_URL="http://localhost:8765/v1"export OPENAI_BASE_URL="http://localhost:8765/v1"
CTXFW is independently audited and certified with Grade A compliance across both the official Glama MCP Registry (TDQS 4.5/5.0 · Security 100/100) and Skills Directory (Security: A 100/100 across 120 static security patterns), and indexed across the 7 MCP Ecosystem Directories:
Compiles .cursorrules and CLAUDE.md negative directives into deterministic AST/path/command/import guards with 3-round iterative path-traversal normalization.
SARA Agent Induction Firewall (ctxfw.sara):
Inspects untrusted tool observations for indirect prompt injection, role overrides (<system>, [INST]), and Trojan-Source bidi escapes (ALLOW, STRIP_COMMANDS, QUARANTINE_TOOL).
SWE-Gate CI Verifier (ctxfw.swegate) & Audit Ledger Retrieval (query_audit_ledger):
Deterministic unified-diff and architectural regression verifier enforcing cyclomatic complexity delta bounds, disallowed module imports, and # type: ignore bans via evaluate_patch, paired with read-only forensic inspection of SweGateStore (EVL-...) and TraceKitStore (SES-...) via query_audit_ledger.
Zero-Touch Provisioning
Inject perimeter rules, MCP server declarations, and pre-commit sentinels into your workspace:
Global IDE Integration
bash
ctxfw init --global
Automatically configures Google Antigravity, Cursor, and Claude Desktop.
Repository Pre-Commit Sentry
bash
ctxfw init --repo .
Deploys .git/hooks/pre-commit to prevent uncertified code commits lacking an attested specification brief.
Enterprise Governance, Sovereign Capsules & CI/CD Gate (v4.1.1)
⚡ 10K PyPI Milestone Offer: Production Sovereign Capsules ($99 USD Single / $490 USD Bundle)
Need standalone, zero-dependency Python/FastAPI/SQLite-WAL enforcement modules (12 SHA-256 sealed artifacts + 100% passing Pytest suite) ready to drop into your repository?