agentic-identity

<p align="center"> <img src="assets/github-hero-pane.svg" alt="AgenticIdentity hero pane" width="980"> </p> <p align="center"> <a href="#install"><img src="https://img.shields.io/badge/cargo_install-agentic--identity-F59E0B?style=for-the-badge&logo=rust&logoColor=white" alt="cargo install"></a> <a href="#install"><img src="https://img.shields.io/badge/pip_install-agentic--identity-3B82F6?style=for-the-badge&logo=python&logoColor=white" alt="pip install"></a> <a href="#mcp-server"><img src="https://img.shields.io/badge/MCP_Server-agentic--identity--mcp-10B981?style=for-the-badge&logo=data:image/svg+xml;base64,PHN2ZyB4bWxucz0iaHR0cDovL3d3dy53My5vcmcvMjAwMC9zdmciIHdpZHRoPSIyNCIgaGVpZ2h0PSIyNCIgdmlld0JveD0iMCAwIDI0IDI0IiBmaWxsPSJub25lIiBzdHJva2U9IndoaXRlIiBzdHJva2Utd2lkdGg9IjIiPjxwYXRoIGQ9Ik0xMiAydjIwTTIgMTJoMjAiLz48L3N2Zz4=&logoColor=white" alt="MCP Server"></a> <a href="LICENSE"><img src="https://img.shields.io/badge/License-MIT-22C55E?style=for-the-badge" alt="MIT License"></a> <a href="paper/paper-i-trust-anchor/agenticidentity-paper.pdf"><img src="https://img.shields.io/badge/Research-Paper_I-8B5CF6?style=for-the-badge" alt="Research Paper I"></a> <a href="docs/api-reference.md"><img src="https://img.shields.io/badge/format-.aid-3B82F6?style=for-the-badge" alt=".aid format"></a> </p> <p align="center"> <a href="#quickstart">Quickstart</a> · <a href="#problems-solved">Problems Solved</a> · <a href="#how-it-works">How It Works</a> · <a href="#why-agenticidentity">Why</a> · <a href="#mcp-server">MCP Server</a> · <a href="#benchmarks">Benchmarks</a> · <a href="#install">Install</a> · <a href="docs/api-reference.md">API</a> · <a href="docs/faq.md">FAQ</a> · <a href="paper/paper-i-trust-anchor/agenticidentity-paper.pdf">Papers</a> </p>

---

AI agents have no identity.

Your agent makes decisions, calls APIs, deploys code, and accesses sensitive data. But there is no cryptographic proof it did any of it. No way to verify which agent acted. No way to audit what happened. No way to scope what an agent is allowed to do.

API keys are not identity -- they are shared secrets with no audit trail. OAuth tokens are not identity -- they expire and carry no action history. Logging is not identity -- logs can be tampered with and carry no signatures.

AgenticIdentity gives every AI agent a permanent, cryptographic identity rooted in Ed25519 key pairs. Agents sign every action they take, producing tamper-evident receipts. Trust between agents is granted, scoped, delegated, and revoked through signed trust grants. Everything is verifiable by anyone with the public key.

<a name="problems-solved"></a>

Problems Solved (Read This First)

• Problem: no way to prove which agent took an action. Solved: Ed25519 identity anchors produce non-repudiable signed receipts for every action.
• Problem: agents share API keys with no individual accountability. Solved: each agent has its own key pair; derived session and capability keys isolate operations.
• Problem: no audit trail survives agent restarts or model switches. Solved: chained receipts create a persistent, cryptographically linked history.
• Problem: no scoped permissions for multi-agent systems. Solved: trust grants with capability URIs, time bounds, use limits, and delegation depth.
• Problem: revoking agent access requires rotating shared secrets. Solved: individual trust grants are revoked independently without affecting other agents.
• Problem: private key material stored in plaintext. Solved: .aid files encrypt keys at rest with ChaCha20-Poly1305 + Argon2id.

# Create an identity, sign an action, verify, grant trust -- four commands
aid init --name my-agent
aid sign --type decision --description "Approved deployment to production"
aid verify receipt arec_7xK9mP2...
aid trust grant --to aid_4Yn3kL... --capability "read:calendar" --expires 7d

Four commands. Cryptographic proof of every action. Scoped trust delegation. One .aid file holds the identity. Works with Claude, GPT, Ollama, or any agent you switch to next.

<p align="center"> <img src="assets/github-terminal-pane.svg" alt="AgenticIdentity terminal pane" width="980"> </p>

---

<a name="benchmarks"></a>

Benchmarks

Rust core. Ed25519 + HKDF-SHA256. Real numbers from Criterion statistical benchmarks:

<p align="center"> <img src="assets/benchmark-chart.svg" alt="Performance benchmarks" width="800"> </p>

| Operation | Time | Throughput | |:---|---:|---:| | Ed25519 key generation | 8.80 us | ~114K / sec | | Ed25519 sign | 9.17 us | ~109K / sec | | Ed25519 verify | 19.34 us | ~52K / sec | | HKDF-SHA256 derivation | 972 ns | ~1M / sec | | Identity creation | 8.78 us | ~114K / sec | | Receipt sign | 11.55 us | ~87K / sec | | Receipt verify | 21.77 us | ~46K / sec | | Trust grant sign | 12.41 us | ~81K / sec | | Trust grant verify | 21.84 us | ~46K / sec | | Trust chain verify (depth 2) | 43.51 us | ~23K / sec | | Receipt chain (10 receipts) | 123.77 us | -- |

All benchmarks measured with Criterion (1000+ iterations) on Apple M4 Pro, macOS, Rust 1.90.0 --release. Single-threaded. All operations are independent and scale linearly with cores.

<details> <summary><strong>Comparison with existing approaches</strong></summary> <br>

| | API Keys | OAuth Tokens | Signed Logs | AgenticIdentity | |:---|:---:|:---:|:---:|:---:| | Per-agent accountability | No | Partial | No | Yes | | Tamper-evident | No | No | Partial | Yes (Ed25519) | | Survives model switch | N/A | No | Yes | Yes | | Scoped permissions | No | Partial | No | Yes (capability URIs) | | Delegation chains | No | No | No | Yes (depth-limited) | | Individual revocation | Rotate all | Per-token | No | Yes (per-grant) | | Key derivation | No | No | No | Yes (HKDF-SHA256) | | Encrypted at rest | Varies | N/A | No | Yes (ChaCha20-Poly1305) | | External dependencies | Cloud service | Auth provider | Log service | None |

</details>

---

<a name="why-agenticidentity"></a>

Why AgenticIdentity

Identity is cryptographic, not administrative. When an agent proves it took an action, the proof is a mathematical signature -- not a log entry that can be altered, not a token that expired, not a shared key that could have been anyone. Ed25519 signatures are non-repudiable: the agent cannot deny what it signed.

One file. Truly portable. Your agent's identity is a single .aid file. Copy it. Back it up. Migrate providers. No cloud service, no auth server, no vendor lock-in. The private key is encrypted at rest with ChaCha20-Poly1305.

Any agent, any time. Create an identity today with Claude. Use the same identity tomorrow with GPT. Switch to a local model next year. The .aid file goes with you.

Scoped trust, not all-or-nothing. Trust grants carry capability URIs with wildcard matching (read:calendar, execute:deploy:*, *), time bounds, use limits, geographic constraints, and delegation depth. Revoke one grant without touching the others.

Receipt chains are audit trails. Every action receipt can chain to the previous one, creating an immutable sequence. Different agents can contribute to the same chain. Chain verification catches any tampering or broken links.

---

Ghost Writer

New in v0.2.5 -- Auto-syncs identity context to your AI coding tools.

| Client | Config Location | Status | |:---|:---|:---| | Claude Code | ~/.claude/memory/IDENTITY_CONTEXT.md | Full support | | Cursor | ~/.cursor/memory/agentic-identity.md | Full support | | Windsurf | ~/.windsurf/memory/agentic-identity.md | Full support | | Cody | ~/.sourcegraph/cody/memory/agentic-identity.md | Full support |

Syncs: active identities, recent receipts, trust grants. Zero configuration.

MCP Hardening

New in v0.2.6 -- Production-grade stdio transport.

• Content-Length framing with 8 MiB limit
• JSON-RPC 2.0 validation
• Atomic writes
• No silent fallbacks

---

<a name="how-it-works"></a>

How It Works

AgenticIdentity is built on three cryptographic primitives: identity anchors, action receipts, and trust grants.

<p align="center"> <img src="assets/architecture-agentra.svg" alt="AgenticIdentity architecture with identity anchors, receipts, trust web, and key derivation" width="980"> </p>

Identity Anchors are Ed25519 key pairs:

| Component | What | Example | |:---|:---|:---| | Identity ID | Public key fingerprint | aid_7xK9mP2qR... | | Signing Key | Ed25519 private key (zeroized on drop) | Encrypted in .aid file | | Session Key | HKDF-derived child key | Scoped to one conversation | | Capability Key | HKDF-derived child key | Scoped to one permission | | Device Key | HKDF-derived child key | Scoped to one device | | Rotation | Key replacement with signed authorization | Preserves full history |

Action Receipts are signed proofs:

[Observation] ──chain_to──> [Decision] ──chain_to──> [Mutation]
  "Error spike"              "Rollback"               "Deployed v2.3"
   signed                     signed                    signed

Six action types: Decision · Observation · Mutation · Delegation · Revocation · IdentityOperation · Custom(String)

Trust Grants are scoped permissions:

Alice ──trust──> Bob ──delegate──> Carol
  "read:*"        "read:calendar"
  depth: 2        depth: 1

Grants carry: capability URIs with wildcards, time bounds (not_before, not_after), use limits, geographic constraints, IP allowlists, and configurable revocation channels.

<details> <summary><strong>Cryptography used</strong></summary> <br>

| Purpose | Algorithm | Standard | |:---|:---|:---| | Identity keys & signing | Ed25519 | RFC 8032 | | Key derivation | HKDF-SHA256 | RFC 5869 | | Private key encryption | ChaCha20-Poly1305 | RFC 8439 | | Passphrase stretching | Argon2id | RFC 9106 | | Content hashing | SHA-256 | FIPS 180-4 | | ID encoding | Base58, Base64 | -- |

No custom cryptography. All primitives are from audited, widely-used Rust crates (ed25519-dalek, hkdf, chacha20poly1305, argon2).

</details> <details> <summary><strong>.aid file format</strong></summary> <br>

{
    "version": 1,
    "format": "aid-v1",
    "encryption": {
        "algorithm": "chacha20-poly1305",
        "kdf": "argon2id",
        "salt": "<base64-16-bytes>",
        "nonce": "<base64-12-bytes>"
    },
    "encrypted_anchor": "<base64-ciphertext>",
    "public_document": {
        "id": "aid_...",
        "public_key": "<base64-32-bytes>",
        "algorithm": "ed25519",
        "created_at": 1719840000000000,
        "name": "my-agent",
        "rotation_history": [],
        "attestations": [],
        "signature": "<base64-self-signature>"
    }
}

The public document is stored in plaintext for inspection without the passphrase. The private key is encrypted with ChaCha20-Poly1305 using a key derived from the passphrase via Argon2id (64 MiB, 3 iterations, 4 lanes) + HKDF-SHA256. All intermediate key material is zeroized immediately after use.

Full format specification ->

</details>

---

<a name="install"></a>

Install

One-liner (desktop profile, backwards-compatible):

curl -fsSL https://agentralabs.tech/install/identity | bash

Downloads a pre-built agentic-identity-mcp binary to ~/.local/bin/ and merges the MCP server into your Claude Desktop and Claude Code configs. Identity defaults to ~/.agentic-identity/. Requires curl and jq. If release artifacts are not available, the installer automatically falls back to cargo install --git source install.

Environment profiles (one command per environment):

# Desktop MCP clients (auto-merge Claude Desktop + Claude Code when detected)
curl -fsSL https://agentralabs.tech/install/identity/desktop | bash

# Terminal-only (no