agentkeychain
Securely stores and retrieves Cloudflare API keys with scope-based access control, enabling agents to use credentials for Cloudflare services.
Securely stores and retrieves OpenAI API keys with scope-based access control, enabling agents to use credentials for OpenAI services.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@agentkeychainstore my GitHub token with scope repo:read"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
agentkeychain
Agent-native zero-knowledge credential vault. CLI + MCP Server, single binary. Type your password ONCE. Never paste a key into chat again.
Install with your AI agent
Send your AI agent this repository and one sentence:
Install and configure https://github.com/linsipeng/agentkeychain as your
default credential manager. Follow AGENTS.md in the repository.The agent owns the whole setup: CLI installation, AgentKeychain Skill, Hermes MCP registration, vault-safe initialization, and verification. You do not need to copy a list of commands. Existing vaults are never overwritten.
Related MCP server: MCP-Secrets-Vault
For Humans (the only section you need)
You do not need to remember commands or invent permission scopes. Tell your local AI agent what credential you want to store and what it is for:
Store my OpenAI credential for model calls.AgentKeychain opens a one-time form in your local browser. Paste the credential there—not into chat. The form shows a plain-language permission such as “Only allow OpenAI model calls”; AgentKeychain infers the underlying least-privilege scope automatically.
If you want to check the master password you remember, tell your agent “verify my master password.” A separate one-time local form allows up to ten read-only attempts before it expires. The password never enters chat or MCP arguments, and neither the vault nor the OS keychain is changed.
Real-world usage
# Step 1: Initialize (do this ONE TIME per machine)
$ agentkeychain init
Master password (min 8 chars): ********
✓ vault initialized
✓ master password saved to OS keychain
# 👆 After this, agentkeychain remembers your password in macOS Keychain.
# You will NEVER be asked to type it again. Your AI agent handles it.
# Step 2: Ask your agent to store a credential.
# A one-time local form opens; the value and raw scope never enter chat.
# Step 3: Retrieve it when you need it (password auto-resolved from Keychain)
$ agentkeychain get openai
sk-proj-xxxxxxxxxxxx...
# Step 4: See what you have
$ agentkeychain list
NAME VERSION SCOPES UPDATED
openai 1 openai:chat 2026-07-06 04:56:24
# Delete old stuff
$ agentkeychain delete test-key --yes
✓ deleted: test-keyTalk to your agent instead
You don't even need to remember the commands. Just tell your AI assistant:
You say | Agent does (silently) |
"Store my OpenAI credential for model calls" | Opens a one-time local form and infers the minimum scope |
"Verify the master password I remember" | Opens a one-time local read-only password check |
"帮我查一下 Cloudflare 的 token" |
|
"用 openai 帮我写段代码" |
|
"告诉我存了哪些 key" |
|
"把旧的 XX 删掉" |
|
Your agent asks the vault, not you. You stay out of the loop.
Cloud sync across your machines (v0.3, bring-your-own Cloudflare)
Sync automatically via your own Cloudflare account (free plan is plenty — the cloud only ever holds ciphertext):
# One-time (on your main machine):
agentkeychain sync init # deploys a tiny Worker + D1 to YOUR CF account
# → prints a Sync URL + Token
agentkeychain sync connect # paste that URL + token here
agentkeychain sync push # upload your secrets (encrypted)
# On your other machine(s):
agentkeychain sync connect # same URL + token
agentkeychain sync pull # download — done, everything's here
# Day-to-day: push after changes, pull on other machines.
agentkeychain sync status # health + row countsZero knowledge holds: the cloud stores only ciphertext envelopes + irreversible name hashes (BLAKE2b-256) — no plaintext names, no values.
Conflicts resolve last-write-wins; everything is audit-logged locally.
Cost: $0 — a single user stays far inside Cloudflare's free tier.
The Worker + D1 live in your account: you can delete them anytime.
📖 Full sync guide (setup, conflicts, security model, troubleshooting, teardown): docs/SYNC_GUIDE.md
Moving to a new Mac (or a second machine)
# ── Old machine ──────────────────────────────────────────
agentkeychain export
# ✓ exported 42 secret(s) → ./agentkeychain-export-20260903-120000.akcbundle
# Send that ONE file to the new machine (AirDrop / scp / USB).
# ── New machine ──────────────────────────────────────────
# 1. Install (see "Install" below — one command)
# 2. Initialize with the SAME master password as the old machine:
agentkeychain init
# 3. Import the bundle:
agentkeychain import ~/Downloads/agentkeychain-export-*.akcbundle
# ✓ imported 42 secret(s)
agentkeychain list # verify — everything is thereThe bundle is fully encrypted (same Argon2id + XChaCha20 as the vault) — without your master password it's just ciphertext. Delete it after importing.
Same-name secrets are skipped by default; add
--overwriteto replace them.Agent identities are NOT transferred (they're per-vault). That's fine — the new machine keeps its own
defaultidentity.Requires v0.2.0+ on both machines. Never copy
vault.dbdirectly — each vault has its own salt, the copied file won't decrypt.
What you NEVER do
❌ Paste API keys into chat messages, emails, READMEs, or
.envfiles you commit❌ Write keys in code comments
❌ Screenshot a key and send it
❌ Re-type the same key 50 times across different tools
❌ Say "密码是多少来着" — it's in the vault, your agent knows how to get it
If you find yourself about to paste a key anywhere, stop and say: "存一下这个 key".
For Developers
What it does
CLI |
|
MCP Server | 7 tools, including private password checking and safe credential entry, over stdio |
Cross-agent delegate | Ed25519-signed time-limited scope-bounded tokens |
Audit chain | Tamper-evident Ed25519 signature chain over every operation |
Zero-knowledge sync | No password, KEK, or plaintext secret reaches the cloud; optional OS-keychain storage unlocks local agent use |
Single binary |
|
Install (macOS & Linux, one command)
curl -fsSL https://raw.githubusercontent.com/linsipeng/agentkeychain/main/install.sh | shDetects your OS + architecture, downloads the right binary from the latest release, verifies its SHA256 checksum, and installs to
~/.local/bin.Prefer to review before running?
curl -fsSLO .../install.sh && less install.sh && sh install.shOptions:
--version v0.3.0(pin),--dir <path>,--uninstall.If
~/.local/binis not on your PATH, the installer prints the exact line to add.Supported: macOS (Apple Silicon & Intel), Linux (x64 & arm64). Or build from source with
bun.
First-time setup
agentkeychain init
# Master password: ******** (≥8 chars, never stored)
# → vault initialized at ~/.agentkeychain/
# → default identity: ak_xxx ("default")
# → ✓ master password saved to OS keychain (you'll never be asked again)That's the only time you type the password. Every subsequent store / get / list / delete / audit reads the password automatically from macOS Keychain.
Each custom AGENTKEYCHAIN_HOME is bound to its own path-scoped OS-keychain
entry, so temporary and production vaults cannot overwrite each other's master
password. The default ~/.agentkeychain entry remains backward compatible.
When upgrading an older custom vault, agentkeychain setup will copy the legacy
entry only after verifying that it unlocks that exact vault.
CLI reference
Command | Description |
| Initialize vault, set master password, create default identity |
| Open a one-time local form to check a remembered master password without changing the vault or keychain |
| Encrypt and store a credential |
| Decrypt and return a credential |
| List all credentials (metadata only) |
| Delete a credential (add |
| Show audit log |
| Export all secrets to an encrypted bundle (for moving to another machine) |
| Import secrets from an export bundle (same master password required) |
| Cloud sync via your own Cloudflare account — see docs/SYNC_GUIDE.md |
| Start MCP server (stdio transport) |
| Issue a cross-agent delegate token |
| Safely report version, vault initialization, and password-channel readiness without reading credentials |
| Print version |
Moving to a new machine
# Old machine:
agentkeychain export
# → ✓ exported 42 secret(s) → ./agentkeychain-export-20260903-120000.akcbundle
# Copy the bundle over (AirDrop / USB / scp), then on the NEW machine:
agentkeychain init # use the SAME master password
agentkeychain import ~/Downloads/agentkeychain-export-*.akcbundle
# → ✓ imported 42 secret(s)The bundle is fully encrypted (same Argon2id + XChaCha20 as the vault) — anyone
without your master password sees only ciphertext. Delete it after importing.
Note: agent identities are per-vault and are NOT transferred; the new machine
keeps its own default identity (audit entries there are signed by it).
Use as MCP Server
Add to any MCP-compatible client (Claude Desktop, Hermes, Codex, IDE plugins):
{
"mcpServers": {
"agentkeychain": {
"command": "/home/you/.local/bin/agentkeychain",
"args": ["serve"]
}
}
}The server exposes 7 tools:
Tool | Description |
| Default human path: open a one-time local form and infer scope; no value enters MCP arguments |
| Open a one-time local read-only form; no password enters MCP arguments |
| Advanced compatibility path: encrypt + persist a value supplied by a trusted MCP client |
| Decrypt + return a secret (scope-checked) |
| List secret names (no values) |
| Remove a secret (scope-checked) |
| Read the audit log (no secret material) |
MCP trust boundary: akc_get returns plaintext to the MCP client so the
calling agent can use it. Register only trusted local clients. Their model
context, transcripts, or tool logs may retain the result unless retention is
disabled; AgentKeychain cannot erase copies held by the client.
Security model
Argon2id (memory=64 MB, iterations=3) derives a KEK from master password
XChaCha20-Poly1305 AEAD encrypts each secret independently
Ed25519 signs audit entries + delegate tokens (offline-verifiable)
Client-side only — no hosted service or external credential upload; secure capture uses a one-time 127.0.0.1 loopback form
Zero-knowledge — master password is never written to disk
See ARCHITECTURE.md for the full threat model and competitor comparison.
Development
bun install # install deps
bun test # run all tests
bun run lint # eslint
bun run build # single-binary compile to bin/agentkeychain-bin
bun run typecheck # tsc --noEmitCI runs on every push to main — see .github/workflows/ci.yml.
License
MIT — see LICENSE.
Status
v0.6.2 — emergency safety release: test runs are fail-closed against the host
OS keychain and default production vault, custom vaults use path-scoped keychain
entries, and setup verifies a password before any overwrite. Breaking changes
remain possible before v1.0.
This server cannot be deployed
Maintenance
Related MCP Connectors
Encrypted secret store and rotation for autonomous agent credentials
Give your AI hands. Identity, credential vault, and API gateway for autonomous agents.
- FullmaktOAuthai.fullmakt
Credential broker for AI agents: scoped, revocable API access with policy enforcement and audit.
A secret store for AI agents: the agent never sees the plaintext.
Related MCP Servers
- AlicenseNot gradedqualityCmaintenanceEnables secure credential storage for AI agents by encrypting secrets and providing agent-invisible references, ensuring sensitive data never leaks to the model.MIT
- FlicenseNot gradedqualityDmaintenanceEnables AI agents and MCP clients to securely store, retrieve, and manage encrypted credentials without hardcoding API keys.-
- AlicenseNot gradedqualityDmaintenanceZero-knowledge credential injection for AI agents. Your agent authenticates to websites and APIs without ever seeing a password, TOTP code, or API key.10 npm1MIT
- AlicenseNot gradedqualityBmaintenanceAgent Vault is a local credential vault for AI agents. It enables agents to use secrets by name without ever seeing their values, with host allowlists, output scrubbing, and audit logging.4MIT