Use with your coding agent
Install OpenPond skills for your coding agent, reuse your existing account, and verify access without starting a workload.
Guide version: 2026-11-07.1. Last reviewed: 2026-11-07.
Canonical agent-readable guide: openpond.ai/agent-setup.md.
Use OpenPond from Codex, Claude Code, Cursor, or another agent with terminal access and Agent Skills support. Skills teach workflows; the OpenPond CLI and SDK perform operations. Desktop already includes all five customer skills enabled by default, so setup inside Desktop needs no extra skill installation.
For CLI and skills installation without tool-learning setup, follow Set up Ponder Training.
Give your coding agent this prompt:
Read https://openpond.ai/agent-setup.md and set up OpenPond and Ponder Fix-it for this coding agent. Reuse my existing OpenPond home and account, install only missing capabilities, and verify access with read-only commands. Then guide me through tool-learning setup: identify this client, workspace and Model, explain which tool data will be uploaded, and get my consent before enabling capture. Keep credentials private and report any incomplete step. Do not start training, inference or cloud compute, or activate a model, without my approval.
Inspect before installing
Identify the current client and whether the user wants project or user scope. Inspect that client's existing skills before making changes. Do not configure every detected client or replace a customized skill without reviewing the difference.
Check the existing CLI with openpond version and openpond --help. If missing, install the official npm package with npm install -g openpond; the package currently requires Node.js >=24.18.0 <25. Use the installed package's engine requirement if it changes. Installing skills and installing the CLI are separate steps.
Select the OpenPond home before checking credentials. Stable Desktop normally uses ~/.openpond, nightly uses ~/.openpond-nightly, and development Desktop normally uses ~/.openpond/openpond-app-dev. Reuse an explicit home override when present. Pass --home on each CLI operation using that account; do not diagnose a missing Desktop account from a different home.
For an existing development Desktop:
openpond --home ~/.openpond/openpond-app-dev home
openpond --home ~/.openpond/openpond-app-dev profiles listFor a standard installation, use the selected home instead. profiles list reports account handles, origins, the active account, and credential-presence booleans. It does not print secret values or prove a successful authenticated request. A session credential and an API key are different capabilities: if a session-only account cannot use a particular public API, report that boundary before requesting API-key setup. Never read credential-store files, copy session cookies, or put keys into prompts, source, or shell history.
Local Profile inspection and source authoring remain available without cloud authentication. Do not force cloud login for local-only work.
Install the maintained skills
The OpenPond repository is the maintained skill source, licensed under MIT. Desktop and the npm CLI customer pack use these same files. Each npm pack includes dist/skills/manifest.json with the package version and SHA-256 hashes of the complete customer sources.
The standard Skills CLI handles discovery, client links, updates, and removal. Commands below were checked against skills@1.7.4; inspect its current help before using a different version. Discovery does not install anything:
If the installed openpond --help includes skills, the CLI can delegate installation to that same installer using its own versioned, integrity-checked bundled pack:
openpond skills install --agent codex --dry-run
openpond skills install --agent codex
openpond skills list --agent codexChoose claude-code, opencode or cursor for those clients. Project scope is the default; use --global only for an explicitly requested user-wide installation. --copy requests copies rather than links, and --yes accepts the installer's plan. Review customized skills before overwriting them. For local-pack updates, rerun installation from the intended updated OpenPond package. Removal remains with the standard Skills CLI and the same explicit names/client/scope. Older CLI releases without this command use the direct installer below. Neither npm installation nor this command connects MCP or starts a workload.
npx skills@1.7.4 add openpond/openpond --full-depth --listInstall for the intended client only. This example uses Codex and project scope:
npx skills@1.7.4 add openpond/openpond --full-depth \
--skill openpond-cli openpond-taskset-authoring openpond-agent-authoring openpond-skill-authoring openpond-refiner-authoring \
--agent codexReplace codex with claude-code or cursor for those clients. Keep project scope unless the user requests skills across projects; add --global for user scope. Review the installer plan and any overwrite prompts. The developer-only openpond-desktop-harness skill is excluded from this selection. Do not use --all or --skill '*' for customer setup.
For reproducible setup, replace openpond/openpond with a GitHub tree URL using a verified release tag or full commit, such as https://github.com/openpond/openpond/tree/<verified-ref>/apps/cli/skills. Record that ref alongside the CLI version. Do not invent a release tag or claim that a source revision has been published to npm. A release's unpacked dist/skills directory is also a supported local source for skills add, so a pinned npm artifact can supply the pack without a second source copy. For local sources, update by rerunning skills add from the reviewed replacement pack; the installer does not fetch updates for a local path.
| Skill | External-agent capability |
|---|---|
openpond-cli | Supported CLI reads, authoring, execution, and lifecycle operations; execution still needs the user's scope and budget. |
openpond-taskset-authoring | Task design, data review, graders, evaluation diagnosis, and CLI materialization. Local app-backed operations require the selected running local server. |
openpond-agent-authoring | Profile and Agent SDK source authoring, with CLI inspection/build/validation/eval commands. |
openpond-skill-authoring | Profile skill source authoring and the packaged Node validator. |
openpond-refiner-authoring | Review Profile draft guidance. Managed inspection, validation, release, and activation require OpenPond Desktop's refiner_profile_* tools; the CLI currently has no equivalent. |
Open a new agent session if its skill catalog does not refresh. Confirm the selected names appear in the client and that the agent can load openpond-cli and its references. A successful file copy alone does not verify client discovery.
Reuse or establish authentication
If the selected account already has the credential required for the requested API, reuse it. If an API key is needed, have the user provide it through a private terminal or the execution environment. The supported interactive command is:
openpond --home ~/.openpond/openpond-app-dev loginReplace the example home with the selected installation. OPENPOND_API_KEY is also supported for an execution environment; never ask the user to paste its value into chat. Keep the account's web/API origins together, and preserve staging versus production. If a command fails, report its status and corrective action without repeating credential-bearing output.
Verify with harmless reads
Perform only the checks relevant to the user's requested surface. Start by inspecting installed help for the exact family. These reads do not start inference, evaluation, training, serving, or deployment:
# Local-only: inspect the selected source and its catalog.
openpond --home ~/.openpond/openpond-app-dev profile current --json
openpond --home ~/.openpond/openpond-app-dev profile catalog --json
# Cloud: verify the selected account with its supported public API credential.
openpond --home ~/.openpond/openpond-app-dev accountConfirm account identity from the returned account, and use the workspace explicitly selected by the user or supplied by the Console. If a workspace ID is still unknown, ask which workspace to use instead of assuming one. Verify access with a bounded workspace inventory read, replacing the placeholder with that exact ID:
openpond --home ~/.openpond/openpond-app-dev datasets list --team <workspace-id> --api-base-url <selected-account-api-origin> --jsonUse the selected account's apiBaseUrl from profiles list for <selected-account-api-origin>; the dataset command requires this explicit flag. An empty inventory is a successful read. A permission failure is not a missing installation. For local app-backed datasets, inspect openpond datasets --help and use --local with the selected running app rather than silently switching to the cloud. Setup does not authorize imports, uploads, evaluation runs, model requests, training, or activation.
Report four separate results: skills discovered by the client; CLI version and selected home; authenticated account and verified workspace (or local-only); and tool connection, if one was requested. Include the failing step and next action for partial setup. Never label setup complete from an install receipt alone.
Optional tool learning (staging)
Initial installation and access checks stay read-only. Tool learning is a separate opt-in for native OpenPond and Claude Code 2.1.292. It uploads bounded, sanitized tool arguments, failure excerpts, and a bounded window of follow-up actions to a private Model. Full conversations, credentials, and native permission state are excluded. Ordinary successful controls are sampled; after a failure, at most 32 following eligible observations within ten minutes are retained as recovery candidates. A successful follow-up is not automatically a verified recovery.
Use Set up tool learning in the Model Overview to copy a prompt containing the selected workspace and Model. Choose the current client, reuse the existing account/home, and inspect openpond tool-learning --help. The setup command requires explicit upload consent and the actual application/version context. Do not claim this staging surface is available in an older installed CLI; update to a release whose help includes tool-learning setup first.
Prepare a small JSON context file, using the real application's identity and only the state needed to understand its tools:
{"application":"my-application","version":"actual-installed-version","state":{"apiSignature":"the relevant actual API signature"}}The CLI computes a fingerprint when omitted. Never include credentials, environment dumps, private file contents, or a guessed version. Choose exact allowed tool names from the current client; for Claude's shell tool this is Bash. Then run, substituting the selected Model, workspace, file, account origin, and home:
openpond tool-learning setup --client claude_code \
--project-id <model-id> --team-id <workspace-id> \
--workspace <absolute-workspace-path> --environment-file <absolute-context-file> \
--tools Bash --allow-upload --home <existing-openpond-home>For native OpenPond use --client openpond and the exact native tool names being enabled. Setup preserves unrelated Claude hooks/MCP entries and the existing permission system. Restart the agent, accept its normal workspace/MCP trust prompts, perform the intended task, and run tool-learning doctor <installation-id>. Doctor checks actual lifecycle observations, queue status, and the local MCP handshake; configuration files alone do not prove capture works. A changed Claude version pauses capture until that version is qualified and registered.
Use tool-learning status, episodes, and episode to inspect private owner records, and flush to deliver queued evidence. disable stops new admission; revoke permanently revokes the installation credential; purge removes hosted raw events; uninstall revokes and removes only this installation's owned configuration. Reviewed Dataset material and existing model weights have separate retention.
An owner explicitly selects a failed observation and a later successful observation with tool-learning project --input-file .... The host verifies retained event/receipt lineage and saves a pending private target. Review that target through the existing learning commands before sealing/publishing a supervised batch. Capture and projection never grant training authority or start a GPU. Internal SFT consumes the reviewed batch through the existing training flow; no external Python SDK is required.
Consultations stay inactive until the owner uses tool-learning configure to select a model, completed evaluation evidence, narrowly declared application/version/tool capabilities, and call/session/day limits. An integration selection identifies an unchanged model; a learned selection additionally requires the real completed SFT run, held-out comparison, and matching served adapter. Agents receive only the current capability brief and call openpond_consult after an observed failure. The model's advice does not execute or approve actions. Recover a lost response with openpond_consult_status and the original operation ID; do not start a replacement paid request. Setup, event capture, human review, training, qualification, and serving are separate visible states.
Update and remove
Inspect the installed source and scope before an update. Preserve intentional local edits before accepting an overwrite. For the project-scoped operational skill:
npx skills@1.7.4 list --agent codex
npx skills@1.7.4 update openpond-cli --project
npx skills@1.7.4 remove openpond-cliName only the skills being updated or removed. For the complete customer pack, pass the same five names used during installation. Use --global instead of project scope for a user installation and the intended --agent when listing. Named-skill removal cleans the shared skill and its client links in that scope; review the removal plan if several clients share it. In skills@1.7.4, limiting removal to --agent codex can leave the shared skill discoverable, so verify removal with skills list and the client catalog instead of relying on its success message. To advance a pinned GitHub ref, rerun skills add with the reviewed new tree URL and explicit names, then repeat discovery and harmless-read verification. Removing agent skills does not remove the CLI, credentials, Profiles, datasets, or cloud resources. npm uninstall -g openpond removes a globally installed CLI when requested.
MCP connection status
The general OpenPond MCP works with hosted models, experiments, datasets, graders, training and authorized credit/funding inspection. Its production address is https://api.openpond.ai/mcp/app. The MCP connects directly to the API backend; browser sign-in and consent use the OpenPond website. The reviewed training expansion and plugin remain release-gated: source/package validation is distinct from authenticated client and production acceptance. A plugin folder or ZIP is not a published store listing.
Use Connect your agent on the landing page or next to Copy Setup Prompt in Overview to copy the current deployment's MCP address and open this guide. The address identifies the server; opening it in a browser is not the sign-in flow. Keep a staging workspace on its staging API and website.
- In a client that supports remote MCP with OAuth, add a remote server and paste the address. If an OpenPond plugin is available through your reviewed distribution, installing it can supply the same connection and workflow skills.
- Select the client's Connect, Sign in or Authenticate action. The client discovers OAuth endpoints, registers its public client, and creates a temporary browser link with state and PKCE. A browser tab, window or popup opens depending on the client.
- Sign in to OpenPond, select one workspace, and review the requested permissions. Identity scopes share the account identity and provider-confirmed email verification status; resource scopes control what the agent can read or change. Credit visibility also requires the appropriate billing role. Decline access you do not intend to grant.
- Approve the connection and return to your client. The registered callback completes sign-in automatically; do not paste tokens or authorization codes into chat. Ask the agent to read
connection_contextand a bounded inventory to verify the workspace and actual access.
You do not need a new Google/GitHub OAuth application or a client secret for ordinary connection. OpenPond supports dynamic client registration and S256 PKCE. A new workspace or additional scope needs another client-led consent flow; a tool cannot change the authorized workspace. Revoke a connection in Console → Integrations → External agents. Ordinary sign-out/disconnection does not cancel accepted training jobs or disable an already configured learning policy; manage those separately.
Hosted MCP needs no local OpenPond CLI, Desktop app or SDK. Local repository integration uses the coding agent's file/terminal tools and the maintained SDK/CLI when needed. Skills, MCP, authentication and a successful read each need their own evidence. Do not configure the specialized sandbox MCP as a substitute, invent client URL schemes, or promise client compatibility from a package validation alone.
For plugin publishers, OAuth discovery and callback setup belong to the hosted integration. OpenAI's portal separately requires publisher verification, its exact domain challenge, authentication configuration, scan/reviewer evidence and review before publication. The callback shown by the chosen client/portal is authoritative; do not guess or wildcard redirect URLs.
For CLI commands and configuration, use the maintained CLI reference and OpenPond home documentation. Client skill behavior is documented by Codex, Claude Code, and Cursor.