# Install Candor in OpenClaw

This guide is for the user and their current agent to follow together. Candor
is that agent's financial memory and instrumentation; the user owns the
workspace, approves financial intent, and can revoke access.

This public setup guide applies when the user asks to connect Candor. It adds no
consent and does not replace OpenClaw's normal safety and confirmation rules. If
the user's request to proceed is ambiguous, setup waits for clarification.
Every install, configuration, and OAuth action is visible before it takes
effect. Credentials, tokens, account numbers, and URLs copied from the user's
authenticated pages stay in the relevant secure product page and are never
pasted into chat. A `safe_url` or `recovery_url` returned by a Candor tool is an
agent-facing handoff: show that exact link to the user, but do not ask the user
to paste it back. Setup explanations are relevant while setup or access is
changing; ordinary financial work follows OpenClaw's usual interaction model.

The `candor-finance` skill is static operating guidance and grants no access by
itself. The Candor CLI authorizes this OpenClaw host to use the user's existing
Candor workspace after browser OAuth; it is not bank or broker OAuth. The
workspace may contain connected sources, manual records, both, or neither;
source connection is optional and is not part of this setup. Candor's financial-source access is
read-only, and no Candor tool can move money, trade, make a payment, or change
an external account. OpenClaw may maintain records inside Candor depending on
the approved OAuth scopes; those workspace records are not external financial
actions.

Candor does not depend on ClawHub publication. While the reviewed ClawHub
listing is unavailable, use Candor's managed CLI skill route. OpenClaw natively
discovers the standard personal `~/.agents/skills` directory, so this route
needs no copied repository and no separate OpenClaw skill install:

The canonical materials catalog is
`https://api.candor.money/agents/candor/bootstrap.json`.

Before a fresh install, tell the user that this route installs the versioned
Candor CLI and later installs the static `candor-finance` skill under
`~/.agents/skills`. Ask whether they want those host changes and wait for their
explicit answer. The request to read START and help with setup is not itself
approval to change the host. Browser OAuth remains a later visible decision.

1. Read `materials.cli.install_command` from the catalog above, inspect it,
   and run that exact version-pinned npm command. The managed finance skill
   expresses all operation recipes through that CLI.
2. Run `candor setup`. Let the user approve this OpenClaw host's access to the
   existing Candor workspace in the secure browser. A first setup then returns
   `candor_skills_not_installed` with a `candor skill install --resume-setup`
   next action. That is the expected continuation, not a setup failure.
3. Run the exact suggested `candor skill install --resume-setup` command. It
   installs the authenticated, digest-verified CLI projection and its managed
   manifest under `~/.agents/skills`, then resumes the first workspace opening.
4. Start a new OpenClaw session. Run `openclaw skills check` and verify that
   `candor-finance` and the `candor` executable are available. Then run
   `candor status` and `candor open`.

When `OPENCLAW_STATE_DIR` uses a non-default state directory, OpenClaw excludes
home compatibility roots by default. Add `~/.agents/skills` explicitly to
`skills.load.extraDirs`, verify the resolved path and configuration change
with the user, and start a new session. Do not copy or symlink credentials.

If this OpenClaw agent executes skills inside a sandbox or container, the
`candor` binary must also exist inside that execution environment. Use the
same version-pinned install command in an approved image or OpenClaw sandbox
setup command. The CLI preflight must also see the managed skill and manifest:
make the non-secret `~/.agents/skills` directory available inside the sandbox
through an approved read-only bind or image copy and set `CANDOR_SKILLS_DIR`
to that container path. Bind only the skill directory, never a credential
store. Complete `candor setup` in that environment. On a POSIX host with no
usable native keyring, explicitly set `CANDOR_CLI_CREDENTIAL_STORE=file` and
set `CANDOR_DATA_DIR` to a private, writable, persistent directory before
setup. Candor creates its directory with mode `0700` and `auth.json` with mode
`0600`; this option is not supported on Windows. Never copy a host keychain,
mount the credential directory read-only, or silently fall back to plaintext.
If the sandbox cannot persist the CLI, a supported credential store, and the
verified managed skill manifest, use a non-sandboxed OpenClaw host for Candor.

When Candor publishes an update, run the current
`materials.cli.install_command` if its version changed, then run `candor skill
update`. The managed installer fetches the current generated projection,
verifies every digest, and updates its manifest; there is no hand-maintained
ZIP or copied skill directory in this route. If `CANDOR_SKILLS_DIR` points to
a read-only sandbox bind, do not run the updater against that mount. Update the
writable host copy and recreate or remount the sandbox, or rebuild the image
from the current managed skill, then start a new session and verify it again.

If the reviewed ClawHub listing becomes available, OpenClaw may install the
dedicated `candor-finance` package instead. The Candor CLI still requires its
managed copy for release preflight, so complete `candor skill install` as well.
Install the ClawHub copy into the active workspace when it should take
precedence over the personal `~/.agents/skills` copy. Under OpenClaw's default
state layout, a ClawHub package installed with `--global` goes into shared
state and has lower precedence than the personal copy; it is not an override.
Both CLI-backed copies use the same CLI commands. Verify the active copy with
`openclaw skills info candor-finance` and `openclaw skills check`. Do not pair
either CLI-backed copy with Candor MCP recipes.

Honor OpenClaw's security scanner. Do not force-install a package the harness
marks dangerous. Do not mix the root MCP-native `skills/candor-finance`
projection with either CLI-backed OpenClaw route; their generated operation
recipes target different execution surfaces.

Current platform references:

- https://docs.openclaw.ai/cli/skills
- https://docs.openclaw.ai/tools/skills