Credential broker for coding agents: paste a key once, approve each directory, keep secrets out of status output.
tokenstash has 22 tools with inconsistent quality. Most tools have descriptions (good), but many lack proper input schemas or have minimal parameter documentation. Tool names are action-oriented (need, ask, answer, forget, rotate) which is positive. However, parameter descriptions are sparse or missing entirely. Output schemas are not documented. Error handling guidance is minimal. The server mixes CLI-oriented tools (init, doctor, run) with MCP-specific tools (mcp, inbox), creating composition confusion. No tool annotations (readOnlyHint, destructiveHint) despite clear risk levels. Several tools (trust, workspaces) are marked for terminal use only, not agent use.
Answer a task from the terminal (paste a secret, approve, or mark done).
File a non-secret task for the human (DNS record, dashboard toggle, OAuth consent screen…).
Show recent audit events (never values).
Bind a secret name to an identity for this project (work vs personal).
Test every stored key against its provider and mark dead ones stale. For a person at a terminal.
Check that everything works.
Write every secret to one passphrase-encrypted file, to move to another machine. For a person at a terminal (an agent with a shell can still run it under a pty; keep that in mind).
No input schemas visible for 8+ tools (check, audit, init, doctor, mcp, open, registry, trust).
Parameter descriptions are minimal or missing. 'names' in need is described as 'Secret names to request' but no format/constraint. 'cmd' in trust/workspaces lacks enum values. LLMs cannot infer valid options.
No output schemas documented. Tools like 'tasks', 'list', 'audit' return structured data but LLMs have no schema to plan downstream calls. Missing return type documentation.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 52 | 2026-07-28+ | v2 |
Delete a secret from the stash.
Merge a bundle from `export` into this machine's stash. For a person at a terminal.
Serve the localhost inbox (started automatically when a task is filed).
Detect agents and wire them up (MCP server + skill, or a `/tokenstash` command with --mode explicit); choose a stash backend.
List stashed secret names and identities. Never values.
Serve the MCP server over stdio (used by agents; configured by `init`).
Request one or more secrets for this project (hit → inject, miss → file a task). Exit 0 injected / 10 pending / 20 denied / 30 expired.
Open the inbox in your browser.
Print the provider registry (names and signup URLs).
Agent-facing: a provider rejected a key tokenstash injected (401/403). Never ask the user in chat; call this, then `need` again.
Replace a key: marks it stale and asks you for the new value now.
Run a command with the project's env file loaded; if it dies on a missing key, file the task, wait, inject, restart.
List open tasks for this project (--all for every project).
Deprecated since 0.2 (directories pair once instead); kept so `trust rm` can tidy old config.
Paired directories and what each may receive; revoke or forget one. For a person at a terminal.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk levels. 'forget' is DESTRUCTIVE, 'need' is READ_ONLY, but no hints in schema.
Mixed CLI and MCP tools in same server. 'init', 'doctor', 'run', 'inbox', 'open' are terminal-oriented; 'need', 'ask', 'answer' are agent-oriented. Composition is unclear, agents cannot safely call init or doctor.