Agent Runtime Proof MCP server
Commit Sep 13, 2026Verifies that a running AI agent or MCP server process matches an approved artifact to prevent stale or replaced runtime attacks.
Also known as fantasyce/agent-runtime-proof, agent-runtime-proof
- Stars
- 1
- Tools
- —
- Downloads, 30 days
- —
- Latest version
- 1.1.0
- Last commit
- Sep 13, 2026
- License
- Apache-2.0
What is Agent Runtime Proof MCP?
Agent Runtime Proof verifies that a running AI agent or MCP server process matches an approved artifact to prevent stale or replaced runtime attacks.
Summary generated from the project's README
About Agent Runtime Proof MCP server
Agent Runtime Proof provides a local, read-only mechanism to verify that a running AI agent or MCP server process matches a specific approved artifact. It addresses the security gap where a file on disk might be updated while an older, potentially stale or replaced process continues to execute. By binding live process identity to an explicit expectation, it ensures that the code actually performing work is the exact version you intended to run.
This tool operates locally on macOS, Linux, and Windows. It can be used via a command line interface or as a local stdio MCP server. When used as an MCP server, it exposes three specific tools: list_local_runtime_candidates, inspect_local_runtimes, and verify_local_runtime. It is designed to be privacy-conscious, omitting sensitive data like environment values, credentials, or command lines from its responses. It works with various hosts including Claude Code, Cursor, and VS Code through provided host profiles.
What you can do with it
- 1Detect stale process runtimesVerify if a running process is actually the version currently stored on your disk to prevent replaced runtime attacks.
- 2Secure command executionLaunch a command through a witness that records process identity and provides a content-addressed launch receipt.
Written from the project's README. Every use case is checked against a quote from it.
Automated checks
Ran Nov 15, 2026Full report →Not verified yet. A server needs a score of 80, no serious finding, and enough data for us to check. How checks work · What to fix for more points
- Source repository linked and reachablefantasyce/agent-runtime-proof is public and reachablegithub.com/fantasyce/agent-runtime-proof ↗Pass
- Package names this repository as its sourceNot checked
- Build provenance (Sigstore attestation)Not checked
- Open source license (OSI)Apache-2.0spdx.org/licenses/Apache-2.0.html ↗Pass
- Commit or release in the last yearLast commit or release 2026-09-13github.com/fantasyce/agent-runtime-proof ↗Pass
- OpenSSF ScorecardNot checked
- Package name not close to a popular packageNot checked
- No critical advisory in the package or its dependenciesNot checked
- No hidden instructions to the model0 tool descriptions and the README read; no override phrases, credential paths or hidden charactersPass
- Destructive tools are labelledNot checked
- Remote endpoint answers the MCP handshakeNot checked
README
View on GitHubAgent Runtime Proof
Prove the agent runtime you launched is the artifact you approved.
Agent Runtime Proof (ARP) is a local, read-only runtime identity verifier for AI agents and MCP servers. It detects stale, replaced, mismatched, or unverifiable runtimes without uploading code, secrets, process arguments, or process data.
The failure ARP is built for is simple: you replace an Agent or MCP executable, but yesterday's process keeps doing the work. The file on disk is new; the loaded runtime is not. ARP binds live process identity to an explicit artifact expectation and reports the uncertainty instead of treating the file on disk as proof of what is running.
See the reproducible stale-runtime demonstration, or start with the five-minute quickstart.
Quickstart
The current release is
v1.1.0
for macOS arm64, Linux amd64, and Windows amd64. Download the archive and
SHA256SUMS from that release, verify the bytes, then install the single binary
in a user-owned directory. Full commands for each platform are in
docs/quickstart.md; lifecycle and rollback details are in
docs/install.md.
After installation:
agent-runtime-proof --version
agent-runtime-proof doctor --format json
agent-runtime-proof inspect --all --limit 20
Verify a process against an explicit expectation:
agent-runtime-proof verify \
--expectation /absolute/path/to/expectation.json \
--pid 1234 \
--format json
ARP returns MATCHED only when the evidence supports that conclusion.
UNKNOWN, STALE, LEAKED, CONFLICT, and NOT_RUNNING are domain verdicts,
not transport errors.
Local stdio MCP
ARP can run as a local MCP server:
{
"mcpServers": {
"agent-runtime-proof": {
"command": "/absolute/path/to/agent-runtime-proof",
"args": ["mcp"]
}
}
}
It exposes exactly three read-only tools:
list_local_runtime_candidatesinspect_local_runtimesverify_local_runtime
Host profiles are available for Codex, Claude Code, Cursor, OpenCode, DeepSeek Harness, VS Code/GitHub Copilot, and generic hosts. See the host configuration guide. ARP does not edit the host's configuration.
The v1.1.0 Release also includes
agent-runtime-proof_1.1.0.mcpb for compatible MCP clients. Its Registry
metadata binds the bundle URL to its SHA-256 digest; the same three native
binaries remain available as ordinary platform archives.
The same release is published in the
official MCP Registry
as io.github.fantasyce/agent-runtime-proof.
What ARP proves
These controls answer different questions and work best together:
| Evidence | Question it answers |
|---|---|
| SBOM | What dependencies were declared in this build? |
| Signature or build attestation | Who produced these artifact bytes, and through which build? |
| Checksum | Did the downloaded bytes change? |
| Agent Runtime Proof | Is the live local process bound to the artifact expectation I approved? |
ARP does not replace signing, provenance, SBOMs, sandboxing, or host policy. It closes the gap between an approved file and the process that is actually doing the work.
Privacy and safety boundary
ARP is intentionally narrow:
- local processes and local files only;
- CLI, local
stdioMCP, and an optional launch Witness; - read-only observation of Agent and host state;
- no daemon, network listener, repair action, or configuration write;
- no dependency on Across Agents Assistant, Across Context, Across Orchestrator, or Across Autopilot.
MCP responses omit raw argv, environment values, command lines, file contents, credentials, and transcripts. Read the exact guarantees and limitations in docs/data-handling.md, docs/privacy-model.md, and docs/threat-model.md.
CLI and Witness
Inspect one process or a bounded current-user inventory:
agent-runtime-proof inspect --pid 1234 --format json
agent-runtime-proof inspect --all --limit 100
Run a local command through the byte-transparent launch Witness:
agent-runtime-proof witness --expectation expectation.json -- command arg
The Witness records process identity and a content-addressed launch receipt,
then proxies stdin/stdout without rewriting protocol bytes. Hosts that own
process creation can embed the same contract through sdk/witness.
Exit codes are 0 for inspection, doctor, or MATCHED; 2 for a determinate
negative verdict; 3 for UNKNOWN; 64 for invalid input; and 70 for an
internal failure. JSON mode writes one JSON value to stdout; sanitized
diagnostics use stderr.
Supported platforms and limits
Release archives are built for macOS 14+ arm64, Linux amd64, and Windows 11 amd64. Passive inspection of interpreter and declared-tree runtimes remains conservative when the active entrypoint cannot be observed. An on-disk digest alone is never reported as a loaded-runtime match.
Remote attestation, a daemon, network listeners, repair actions, and Agent configuration writes remain outside v1.
Architecture and acceptance
The same contracts and application layer drive the CLI, MCP server, Witness, and data-only Host Profiles. Technical design is documented in architecture-development-acceptance.md.
Acceptance records:
- Phase 0 contracts
- Phase 1 macOS core
- Phase 1 Windows core
- Phase 2 MCP
- Phase 3 Witness
- Phase 4 host matrix
- Phase 5 open-source release
Maintainers can run the complete source gate with:
bash scripts/check.sh
Release assets include CycloneDX SBOMs, SHA-256 checksums, and GitHub artifact attestations. Publication state is determined by the public GitHub Release, not by a local acceptance record.
Agent Reliability Toolkit
ARP is one independent part of a small, local-first reliability toolkit:
- Agent Recall Trail keeps private Agent memory separate from reviewed shared knowledge.
- Agent Runtime Proof verifies that a live Agent or MCP runtime matches the artifact you approved.
- Agent Residue Evidence records task-scoped files, processes, and listening ports left by tests and builds.
Each project remains separately installable and keeps its own trust boundary.
Community
Contributions and independent host/runtime cases are welcome. For a sanitized Codex or DSH integration, share the host, version, launch relationship, and observed proof boundary in the design-partner issue. DSH users can also join the official community discussion. Read CONTRIBUTING.md, CODE_OF_CONDUCT.md, and SUPPORT.md. Report vulnerabilities privately as described in SECURITY.md; do not put secrets or sensitive process data in a public issue.
Apache-2.0 licensed. See LICENSE.
Agent Runtime Proof MCP questions
What does Agent Runtime Proof actually verify?
It verifies if the live local process is bound to the specific artifact expectation you approved, closing the gap between an approved file and the actual running process.
Does this tool upload my code or secrets?
No, it is a local, read-only verifier that detects mismatched runtimes without uploading code, secrets, process arguments, or process data.
What platforms are supported?
The tool is available for macOS 14+ (arm64), Linux (amd64), and Windows 11 (amd64).