# JACS
**Sign it. Prove it.** -- Agent trust infrastructure for a world where AI agents cross organizational boundaries.
Cryptographic signatures for AI agent outputs so anyone can verify who said what, whether it was changed, and hold agents accountable. No server. No account. Three lines of code.
## Quick Start
Zero-config -- one call creates a persistent agent with keys on disk.
### Password Setup
Persistent agents use encrypted private keys. Rust/CLI flows require an explicit password source before `quickstart(...)`. Python/Node quickstart can auto-generate one, but production deployments should still set it explicitly.
```bash
# Option A (recommended): direct env var
export JACS_PRIVATE_KEY_PASSWORD='use-a-strong-password'
# Option B (CLI convenience): path to a file containing only the password
export JACS_PASSWORD_FILE=/secure/path/jacs-password.txt
```
Set exactly one explicit source for CLI. If both are set, CLI fails fast to avoid ambiguity.
For Python/Node library usage, set `JACS_PRIVATE_KEY_PASSWORD` in your process environment (recommended).
### Python
```python
import jacs.simple as jacs
info = jacs.quickstart(name="payments-agent", domain="payments.example.com")
print(info.config_path, info.public_key_path, info.private_key_path)
signed = jacs.sign_message({"action": "approve", "amount": 100})
result = jacs.verify(signed.raw)
print(f"Valid: {result.valid}, Signer: {result.signer_id}")
```
### Node.js
```javascript
const jacs = require('@hai.ai/jacs/simple');
async function main() {
const info = await jacs.quickstart({
name: 'payments-agent',
domain: 'payments.example.com',
});
console.log(info.configPath, info.publicKeyPath, info.privateKeyPath);
const signed = await jacs.signMessage({ action: 'approve', amount: 100 });
const result = await jacs.verify(signed.raw);
console.log(`Valid: ${result.valid}, Signer: ${result.signerId}`);
}
main().catch(console.error);
```
### Rust / CLI
```bash
# Install (requires Rust toolchain)
cargo install jacs-cli
# Or download a prebuilt binary from GitHub Releases
# https://github.com/HumanAssisted/JACS/releases
jacs quickstart --name payments-agent --domain payments.example.com
jacs document create -f mydata.json
# Start the built-in MCP server (stdio transport)
jacs mcp
```
`jacs mcp` is local stdio-only transport. No separate install step needed.
### Homebrew (macOS)
```bash
brew tap HumanAssisted/homebrew-jacs
# Install JACS CLI
brew install jacs
# Install HAI SDK CLI separately
brew install haisdk
```
## Verify a Signed Document
No agent needed. One command or one function call.
```bash
jacs verify signed-document.json # CLI -- exit code 0 = valid
jacs verify --remote https://example.com/doc.json --json # fetch + verify
```
```python
result = jacs.verify_standalone(signed_json, key_directory="./keys") # Python, no agent
```
```typescript
const r = verifyStandalone(signedJson, { keyDirectory: './keys' }); // Node.js, no agent
```
[Full verification guide](https://humanassisted.github.io/JACS/getting-started/verification.html) -- CLI, Python, Node.js, DNS, verification links.
## Which Integration Should I Use?
Find the right path in under 2 minutes. [Full decision tree](https://humanassisted.github.io/JACS/getting-started/decision-tree.html)
| Python + LangChain/LangGraph | `from jacs.adapters.langchain import jacs_signing_middleware` | [LangChain Guide](https://humanassisted.github.io/JACS/python/adapters.html) |
| Python + CrewAI | `from jacs.adapters.crewai import jacs_guardrail` | [CrewAI Guide](https://humanassisted.github.io/JACS/python/adapters.html) |
| Python + FastAPI | `from jacs.adapters.fastapi import JacsMiddleware` | [FastAPI Guide](https://humanassisted.github.io/JACS/python/adapters.html) |
| Node.js + Express | `require('@hai.ai/jacs/express')` | [Express Guide](https://humanassisted.github.io/JACS/nodejs/express.html) |
| Node.js + Vercel AI SDK | `require('@hai.ai/jacs/vercel-ai')` | [Vercel AI Guide](https://humanassisted.github.io/JACS/nodejs/vercel-ai.html) |
| Node.js + LangChain.js | `require('@hai.ai/jacs/langchain')` | [LangChain.js Guide](https://humanassisted.github.io/JACS/nodejs/langchain.html) |
| MCP Adapter (Python, partial) | `from jacs.mcp import create_jacs_mcp_server` | [Python MCP Guide](https://humanassisted.github.io/JACS/python/mcp.html) |
| MCP Adapter (Node.js, partial) | `require('@hai.ai/jacs/mcp')` | [Node.js MCP Guide](https://humanassisted.github.io/JACS/nodejs/mcp.html) |
| A2A Protocol | `client.get_a2a()` / `client.getA2A()` | [A2A Guide](https://humanassisted.github.io/JACS/integrations/a2a.html) |
| Rust / CLI (canonical MCP server) | `cargo install jacs-cli` | [Rust Guide](https://humanassisted.github.io/JACS/rust/installation.html) |
| Any language (standalone) | `import jacs.simple as jacs` | [Simple API](https://humanassisted.github.io/JACS/python/simple-api.html) |
## MCP Support Matrix
| Canonical full `jacs_*` server | Yes | No | No | No |
| Transport proxy / protocol glue | N/A | Yes | Yes | Demo only |
| Middleware / framework adapter | N/A | Yes | Yes | No |
| Language-native tool registration | Canonical | Partial compatibility layer | Partial compatibility layer | No |
| Demo / example only | No | No | No | Yes |
The canonical full MCP contract is exported from Rust at [`jacs-mcp/contract/jacs-mcp-contract.json`](jacs-mcp/contract/jacs-mcp-contract.json). Node and Python adapters are tested against that snapshot so contract drift is explicit.
## Who Is JACS For?
**Platform teams** building multi-agent systems where agents from different services -- or different organizations -- need to trust each other's outputs.
**Compliance and security engineers** in regulated industries (finance, healthcare, government) who need cryptographic proof of agent actions, not just log files.
**AI framework developers** adding provenance to LangChain, CrewAI, FastAPI, Express, Vercel AI, or MCP pipelines without changing their existing architecture.
**Researchers and labs** running public-facing agents that need verifiable identity without exposing operator information.
## When You DON'T Need JACS
Honesty builds trust, so here is when JACS is probably overkill:
- **Single developer, single service.** If all your agents run inside one process you control and you trust your own logs, standard logging is fine.
- **Internal-only prototypes.** If data never leaves your organization and you are not in a regulated environment, the overhead of cryptographic signing adds no value yet.
- **Simple checksums.** If you only need to detect accidental corruption (not prove authorship), SHA-256 hashes are simpler.
JACS adds value when data crosses trust boundaries -- between organizations, between services with different operators, or into regulated audit trails.
## The Trust Blind Spot
A 2026 survey of 29 multi-agent reinforcement learning publications found that **zero** addressed authentication, integrity, or trust between agents (Wittner, "Communication Methods in Multi-Agent RL," [arxiv.org/abs/2601.12886](https://arxiv.org/abs/2601.12886)). Every paper optimized for communication efficiency while assuming a fully trusted environment.
That assumption holds in a research lab. It does not hold in production, where agents cross organizational boundaries, outputs are consumed by downstream systems, and regulatory auditors expect cryptographic proof.
JACS provides the missing trust layer: identity (who produced this?), integrity (was it changed?), and accountability (can you prove it?).
## Post-Quantum Ready
JACS supports ML-DSA-87 (FIPS-204) post-quantum signatures alongside classical algorithms (Ed25519 and RSA-PSS). The `pq2025` algorithm preset is the default across quickstart/create paths.
[Algorithm Selection Guide](https://humanassisted.github.io/JACS/advanced/algorithm-guide.html)
## A2A Interoperability
Every JACS agent is an A2A agent -- zero additional configuration. JACS implements the [Agent-to-Agent (A2A)](https://github.com/a2aproject/A2A) protocol with cryptographic trust built in. For details on how signing, A2A, and attestation relate, see the [Trust Layers Guide](https://humanassisted.github.io/JACS/getting-started/trust-layers.html).
Built-in trust policies control how your agent handles foreign signatures: `open` (accept all), `verified` (require key resolution, **default**), or `strict` (require local trust store entry).
```python
from jacs.client import JacsClient
client = JacsClient.quickstart(name="a2a-agent", domain="a2a.example.com")
card = client.export_agent_card("http://localhost:8080")
signed = client.sign_artifact({"action": "classify", "input": "hello"}, "task")
```
```javascript
const { JacsClient } = require('@hai.ai/jacs/client');
const client = await JacsClient.quickstart({
name: 'a2a-agent',
domain: 'a2a.example.com',
});
const card = client.exportAgentCard();
const signed = await client.signArtifact({ action: 'classify', input: 'hello' }, 'task');
```
For trust bootstrap flows, use a signed agent document plus an explicit public key exchange:
```python
agent_json = client.share_agent()
public_key_pem = client.share_public_key()
client.trust_agent_with_key(agent_json, public_key_pem)
```
## Cross-Language Compatibility
A document signed by a Rust agent can be verified by a Python or Node.js agent, and vice versa. The signature format is language-agnostic -- any JACS binding produces and consumes the same signed JSON.
Cross-language interoperability is tested on every commit with both Ed25519 and post-quantum (ML-DSA-87) algorithms. Rust generates signed fixtures, then Python and Node.js verify and countersign them. See the test suites: [`jacs/tests/cross_language/`](jacs/tests/cross_language/mod.rs), [`jacspy/tests/test_cross_language.py`](jacspy/tests/test_cross_language.py), [`jacsnpm/test/cross-language.test.js`](jacsnpm/test/cross-language.test.js).
## Use Cases
**Prove that pipeline outputs are authentic.** A build service signs every JSON artifact it emits. Downstream teams and auditors verify with a single call; tampering or forgery is caught immediately. [Full scenario](USECASES.md#1-verifying-that-json-files-came-from-a-specific-program)
**Run a public agent without exposing the operator.** An AI agent signs every message but only publishes the public key via DNS. Recipients verify origin and integrity cryptographically; the operator's identity never touches the internet. [Full scenario](USECASES.md#2-protecting-your-agents-identity-on-the-internet)
**Add cryptographic provenance in any language.** Finance, healthcare, or any regulated environment: sign every output with `sign_message()`, verify with `verify()`. The same three-line pattern works identically in Python, Node.js, Rust, and Go. [Full scenario](USECASES.md#4-a-go-node-or-python-agent-with-strong-data-provenance)
## Links
- [Documentation](https://humanassisted.github.io/JACS/)
- [Decision Tree](https://humanassisted.github.io/JACS/getting-started/decision-tree.html)
- [Algorithm Guide](https://humanassisted.github.io/JACS/advanced/algorithm-guide.html)
- [API Reference](https://humanassisted.github.io/JACS/nodejs/api.html)
---