blindfold

A local secret broker for coding agents. They use secrets they can never read.

Claude Code, Codex, and friends need your database strings, API keys, and platform tokens to do real work — but everything an agent sees lands in context windows, transcripts, and provider logs, and a prompt-injected agent can be induced to leak what it holds. Blindfold keeps the values in an encrypted local vault and lets agents ask for operations instead: run this query, set this Vercel env var, run this Railway deploy. The daemon performs the operation with the real value; the agent gets results with every vault value scrubbed out.

# what the agent sees
DATABASE_URL = «db_url:prod»          # usable, not readable
> db_query("prod", "SELECT count(*) FROM users")  →  48112  ✓

The honest sentence

Operations the daemon performs itself (SQL, provider APIs, secret generation — "Tier 1") can never return a value, by construction. Operations that spawn a process with secrets in its environment (provider CLIs, your dev server — "Tier 2") are a raised bar, not a wall: hostile code in that process could read its own env. Tier 2 is defended in depth — no shell (args array only), binary pinning, injection scoping, policy gating with native approval dialogs, and output redaction — and every use is audited. If you want zero soft boundary, deny Tier 2 in policy and live with typed operations only.

Management operations (registering databases and CLIs, adding secrets, reloading policy, shutting down) require a capability token that lives in a 0600 file and that the MCP surface never sends — so an agent limited to the MCP tools cannot reconfigure the broker around its own constraints. An agent with a shell running as you can read that file, which is the same-user case below.

Out of scope entirely: malware running as your user, root, and compromised upstream CLI releases.

Requirements

• macOS (Keychain + native approval dialogs; Linux runs with BLINDFOLD_MASTER_KEY set and policy-only approvals)

• Bun ≥ 1.1

• Postgres for the DB proxy (covers Neon, Supabase, Railway Postgres, RDS)

Quick start

git clone https://github.com/xxgvqxx/blindfold && cd blindfold
bun install
bun run bf init          # vault + policy + daemon + MCP instructions
bun run bf add api_key:stripe        # hidden prompt — paste the value
bun run bf db add mydb --mode ro     # hidden prompt for the connection string
bun run bf cli preset vercel         # pin the vercel CLI, store its token
bun run bf import .env --project myapp   # import a dotenv, then delete the file

# register with Claude Code (bf init prints the exact command):
claude mcp add --scope user blindfold -- bun run /path/to/blindfold/src/mcp/index.ts

The daemon starts on demand and holds the vault key from the macOS Keychain (BLINDFOLD_MASTER_KEY env var on Linux/CI). State lives in ~/.blindfold/.

What the agent gets (19 MCP tools)

Policy

~/.blindfold/policy.toml — ordered rules, first match wins, sensible defaults underneath (reads allow; prod+rw SQL and prod pushes require approval; Tier 2 requires approval unless the subcommand is registered read-only). remember = "8h" turns one approval into a session grant. Approvals are native macOS dialogs that time out to deny.

[[rule]]                       # trust vercel passthrough on one project for 8h per approval
op = "cli_run"
where = { tool = "vercel", project = "myapp" }
action = "approve"
remember = "8h"

[[rule]]                       # never let the agent touch a prod rw database
op = "db_query"
where = { alias_tag = "prod", alias_mode = "rw" }
action = "deny"

Redaction

Everything that leaves the daemon — query rows, CLI output, error messages — passes one scrubber that knows every vault value plus its base64, URL-encoded, and JSON-escaped forms, and the individual credentials inside connection strings. Replacements are labeled («db_url:prod») so agents can debug without learning anything. Arbitrary transforms by hostile Tier-2 code are not catchable — redaction is the last layer, not the boundary.

Status

v0.1 — working: vault, daemon, policy/approvals/audit, redaction, Postgres proxy, Tier-2 runner with pinning, Vercel + Railway env sync, dotenv import, 19-tool MCP server, bf CLI. 56 tests, including live-Postgres tests that try to escalate a read-only alias six different ways and an end-to-end suite that proves injection + redaction with real processes. Untested against live provider APIs: the Vercel decrypt=true pull and the Railway GraphQL schema — treat env_pull/env_push as beta until exercised. Not yet built: Modal/Supabase typed verbs (use bf cli preset passthrough), Touch ID helper (osascript dialogs today), MySQL, Linux keyring, launchd install.

Pointing your agents at it

Once the MCP server is registered, agents can use the vault — the docs/AGENTS-snippet.md file is what makes them prefer it. Paste it into each project's CLAUDE.md / AGENTS.md, import your .env files with bf import, and delete the plaintext originals. When there is no file to read, the brokered path is the only path.

Contributing

Adapters are the highest-leverage PRs — the interface is four methods in src/daemon/adapters/types.ts (Fly, Render, Neon, Cloudflare all wanted). bun test runs the suite, including an end-to-end daemon test that proves injection + redaction with real processes.

License

MIT