When an AI coding assistant works across a multi-file repository, written rules in a prompt eventually get ignored: the AI changes a function in File A and breaks a caller in File B, follows an outdated design note after the code has changed, or claims “all tests pass” without running them.
Resonant (`resonant-exocortex`) is an MCP server for Antigravity, Cursor, and Claude Code—linked to a Cloud CI/CD Pull-Request Gate—that checks edits in milliseconds before saving and binds project memory directly to your code.
Before your AI writes a file to disk, Resonant inspects the modified function, tests boundary inputs in memory (such as empty lists [] or 0), compares old vs. new behavior, and checks for leaked secrets or in-place caller list mutations.
Every architectural note stored in Resonant’s SQLite database is linked to the SHA-256 fingerprint of the actual code function. The moment you edit that function later, outdated notes about it expire automatically—and fixed bug patterns become local 0-token rules.
Extracts only the target function, its callers, and Git co-change partners during coding, and links directly to resonant_ci_gate (POST /api/mcp/ci-gate) so Pull Requests are checked against the same learned guardrails before merge.
Most MCP servers on the market are either thin API wrappers (GitHub, Slack, file readers) or keyword/vector memory banks. Resonant applies five peer-reviewed computer-science techniques directly to your code in under 2 milliseconds at 0 extra LLM token cost:
| Resonant Pillar | What Standard AI Tools Do | What Resonant Does (In Plain English) | Peer-Reviewed CS & Security References |
|---|---|---|---|
1. Pre-Save Edge & Security Testingresonant_verify_cegar |
Ask a second AI model to “review” the first AI’s diff (LLM-as-a-Judge), which takes 5–15s, costs extra tokens, and misses edge cases. | Runs your modified function in a safe memory sandbox against tricky inputs (0, -1, None, empty lists []) and checks 40+ security and 60fps rules in <2ms before saving. |
• Clarke et al., CAV 2000 (CEGAR) • Peng et al., CWEval (2025) • Vero et al., BaxBench (2025) |
2. Surgical Caller & Git Co-Change Mapresonant_architect_orient |
Chop files into arbitrary text paragraphs and search by word similarity, missing callers in other files. | Maps each function’s inputs, state changes, and return guarantees alongside its callers and files that historically changed together in your Git commits (~87% fewer tokens). |
• Hoare, CACM 1969 ({P} C {Q}) • Zimmermann et al., IEEE TSE 2005 • Weiser, IEEE TSE 1984 (Slicing) |
3. Self-Expiring Code-Linked Memoryresonant_memory_jtms |
Save static text notes in a database that quietly go out of date weeks later when code is refactored (“memory rot”). | Attaches every note to a SHA-256 fingerprint of the actual code structure. The moment anyone edits that function, old notes automatically flag as STALE_AST_DRIFT. |
• Doyle, Artificial Intelligence 1979 (JTMS) • Wu et al., LongMemEval (ICLR 2025) |
4. 0-Token Structural Bug Guardrailsresonant_evolve_guardrail |
Append another text warning to AGENTS.md or CLAUDE.md, bloating prompt costs until the AI ignores half the file. |
Compares the broken code tree against the fixed code tree to extract the shared structural mistake and saves a <1ms local rule that costs 0 prompt tokens forever. |
• Plotkin, Machine Intelligence 1970 (Anti-Unification) • Rolim et al., IEEE/ACM ICSE 2017 |
5. Fact-Checked Completion Receiptsresonant_honesty_audit |
Trust the AI when it says “I updated the file and all tests passed” without checking disk state. | Breaks the AI’s summary into individual claims and verifies every file path, code symbol, and test exit code (PROVED, HYPOTHESIS, or REFUTED). |
• Necula, ACM POPL 1997 (PCC) • Min et al., FActScore (EMNLP 2023) |
Every coding task shows resonant_architect_orient before editing and resonant_verify_cegar (verified: true) right after editing.
Call resonant_product_telemetry at any time to inspect cumulative raw tokens avoided, API dollars saved, CEGAR bugs caught, and active evolved guardrails.
Run tb_mcp_server.py --self-test (verifies 6/6 tools in <1s) and check the PROVED / REFUTED receipts in every completion audit.
You don’t need to install Python packages or upload your repository. Connect Claude Code, Cursor, Google Antigravity, or Windsurf directly to our live Python verification engine (https://resonantexo.com/api/mcp — 50 free calls/day per IP in an isolated sandbox, or 5,000 calls/day with a Pilot Key):
claude mcp add --transport http resonant-exocortex https://resonantexo.com/api/mcp
{
"mcpServers": {
"resonant-exocortex": {
"url": "https://resonantexo.com/api/mcp"
}
}
}
“Refactor this function and pass both old_code and code to resonant_verify_cegar before saving to check for boundary bugs, silent None returns, or dropped guards.”
“Pass these two files in files to resonant_architect_orient and show me the extracted function contracts, AST hashes, and coupled callers.”
“Use resonant_memory_jtms to commit an architectural rule linked to this function, then modify the function and run reconcile to see it flag STALE_AST_DRIFT.”
| Capability | Free Public Sandboxresonantexo.com/api/mcp |
VIP Pilot Key (Evaluators)X-Resonant-Key: rk_pilot_... |
Full Local & CI/CD Editionresonant-exocortex |
|---|---|---|---|
| Verification Engine | 100% Real Python CS Engine (AST contract diffing, in-memory Python bytecode differential execution, polyglot & CWE checks). | 100% Real Python CS Engine across Python, TypeScript/JS, C/C++, Rust, and C#. | 100% Real Python CS Engine running locally offline with zero network latency. |
| Multi-File Context Slicing | Inline single-file (code) or multi-file map (files) up to 64 KB per call. |
Inline multi-file map up to 256 KB per call. | Full local Git repository + automatic 30-commit Git co-change history and caller graph. |
| AST-Linked SQLite Memory | Isolated ephemeral per-IP sandbox DB (tests commit, query, and STALE_AST_DRIFT live). |
Persistent isolated tenant SQLite DB across sessions for your team or evaluation pilot. | Local SQLite WAL database (resonant_jtms.db) committed alongside your repo. |
| Daily Quota & Guardrail Evolution | 50 tool calls/day per IP (zero sign-up required). | 5,000 tool calls/day + custom guardrail evolution. | Unlimited local calls + pre-save IDE hook + POST /api/mcp/ci-gate PR blocker. |
The public endpoint at https://resonantexo.com/api/mcp runs our live Python verification engine on Google Cloud Run (us-central1, monitored 24/7 by Google Cloud Monitoring) with 50 free calls/day per IP in an isolated sandbox. For technical diligence reviewers, startup/grant evaluators, and engineering teams testing multi-session SQLite memory, Google Vertex AI / Gemini workflows, or self-hosted NVIDIA NIM / vLLM GPU clusters, we issue dedicated X-Resonant-Key: rk_pilot_... tokens with 5,000 calls/day and sub-2ms (0.06ms warm) tenant storage.
Request Pilot Key / Partnership Inquiry: founder@thunderbearstudios.org • Technical Support: dev@thunderbearstudios.org • Architecture & Literature: Technical Overview & Published Research