Skip to main content
ChargerDojo runs a server that coding agents can call directly, at https://mcp.chargerdojo.com. It speaks the Model Context Protocol, the standard your agent already uses to reach tools outside itself. Connect it and you can ask OCPI questions in the editor you are already in: which versions exist, how each one encodes the credentials token, what changed between 2.2.1 and 2.3.0, which suites and journeys you can run. That half needs no account, no key and no signup. Answers come from the same registry the simulator obeys, so they match what the site does rather than what a model remembers. Add your cdojo_ API key and the same agent also reads your organisation’s registered endpoints and run history, a verdict, and your remaining quota. An agent that reads pages rather than calling tools can start from chargerdojo.com/llms.txt: every public page in one line each, and facts about OCPI and ChargerDojo it can quote without fetching a page. llms-full.txt adds a note per page.

Connect it

Each of these was run against the live endpoint.
Cursor reads ~/.cursor/mcp.json; opencode reads ~/.config/opencode/opencode.json. Any other client that speaks Streamable HTTP takes the URL and nothing else.
opencode’s own published examples show mcp.servers.<name>. That shape does not connect. The one above does.

Ask it something

Nothing else to set up. Ask your agent a version question and it will reach for the server:
What changed between OCPI 2.2.1 and 2.3.0?
That last line is the habit worth noticing. The server separates what the specification says from what ChargerDojo implements, every time, because conflating the two is how an integrator ends up planning against a module nobody has.

Add your key

Nine tools are open to anyone. The other thirteen read or act on one organisation, so they need an API key, the same cdojo_ key the REST API uses. Send it as a bearer token:
The remove is not optional if you already ran the command above. The name is the identity, so a second add says MCP server chargerdojo already exists and changes nothing: your header never lands, and the account tools never appear. Clients that take a config file instead carry the same header in a headers map beside the URL. Keys are created in the app, under Settings, then Credentials, and need a paid plan. See Create an API key.
A key grants access to its organisation. There are no per-tool scopes: an agent holding your key reads every target and every report the organisation owns, and can spend run quota. Give an agent a key you are willing to revoke.
Runs an agent starts are recorded in your organisation’s audit log as agent-started, naming the tool.

Working with ChargerDojo

Your agent receives this guidance from the server when it connects, so you do not need to paste it into a prompt. It is here so you know what your agent was told.
  • Protocol questions need no account. chargerdojo_get_ocpi_version, chargerdojo_compare_ocpi_versions, chargerdojo_list_checks and the other open tools answer from the registry the simulator obeys. A module list is what ChargerDojo tests, not the whole specification, and a version comparison covers only the rules it names.
  • Start a test from what your connection can do. chargerdojo_describe_access says what your plan and role allow. chargerdojo_list_targets gives the id of an endpoint someone registered in the app, and chargerdojo_start_conformance_run or chargerdojo_run_journey starts a run on it. Retrying a start with the same retryKey returns the run it already started.
  • Read a verdict the strict way. A pass is a completed run that graded at least one check and failed none. A run that graded nothing is no answer, never a pass. chargerdojo_get_run_checks lists the checks that did not pass. It quotes the recorded clauses and identifies rules without one. Text the partner sent arrives fenced as data.
  • See what changed. chargerdojo_compare_run compares a run with the previous comparable run, and chargerdojo_get_schedule reads the recurring run on an endpoint.
  • Gate CI on the same verdict. The GitHub Action and the REST recipe below exit 0 for a pass on the checks that ran, 1 for a failed required check and 2 for no answer.
  • Tell the team. When the tools cannot do what you asked, or you think a verdict is wrong, your agent offers to send your words to the ChargerDojo team with chargerdojo_send_feedback, and sends only once you agree.
A pass is ChargerDojo’s verdict on the checks that ran. It is not an OCPI certification.

What each plan unlocks

The open tools are free for everyone. An API key, and with it every account tool, comes with a paid plan; what a run may reach follows the same plan as in the app. Ask your agent to call chargerdojo_describe_access for your own plan and role. Plans and pricing has every row.

Every tool

Open to anyone, no account: With an API key: The ids chain: list_targets gives you a targetId, which the start tools take, and every run tool takes the reportId that list_runs or a start tool hands back. Your agent can read all of this from the server itself, which stays right when this table drifts.

What it will not do

Can an agent start a conformance run against my partner?

Only against an endpoint someone in your organisation registered in ChargerDojo, and only on a plan that tests your own endpoints, Integrate and up. chargerdojo_start_conformance_run and chargerdojo_run_journey take a registered target’s id. Neither can register an endpoint, change its address or read its credentials. A run against your own endpoint sends graded traffic to it and spends your run quota, the same as starting it in the app. Against the sandbox endpoints neither tool needs a plan feature, though calling it still needs an API key. Pass a retryKey when your agent may retry a call: the same key returns the run it already started instead of starting a second one. A key that started a run on one endpoint, journey or module is refused for another, so use a new key for each run you mean to start.

How much of my rate can an agent use?

An agent gets half your rate, not all of it. Tools that start work draw on half your plan’s runs per minute and half its in-flight cap, so a looping agent leaves your CI pipeline the other half.

If your agent cannot connect

  • The URL is https://mcp.chargerdojo.com, with no path and no version segment. docs.chargerdojo.com runs a different server, for these pages.
  • The endpoint answers the current revision plus 2025-11-25 and 2025-06-18 from the same URL. Anything older is refused deliberately: those revisions may legally send JSON-RPC batch arrays, which this endpoint does not accept.
  • If the account tools are missing, ask the agent to call chargerdojo_describe_access. If it says you are calling anonymously, the key did not arrive. The usual cause is the one above, a second add refused because the name was taken. Otherwise check the header reached the server rather than that it sits in the config file: some clients send only the headers they were given at add time.
  • If it says your organisation’s plan no longer includes API access, the key is fine and the plan lapsed. The open tools keep working, and each account tool is refused naming the plan that includes it.
  • A revoked or unknown key gets a 401 and the message Invalid or revoked API key. It never quietly drops you to the open tools, because running an agent against the wrong account is worse than an error.
  • A key whose creator has left the organisation gets a 403 with the code ORGANISATION_ACCESS_REVOKED. Ask an admin for a new key.
  • Each key can make 120 calls a minute. Past that, calls get a 429 with Retry-After until the minute is up, and other keys behind the same network are unaffected.
  • Each of these refusals arrives as a JSON-RPC error your client can show, not as a dropped connection.
  • The server is listed in the official MCP registry as com.chargerdojo/ocpi, if your client discovers servers that way.

REST automation

The same API key acts within its organisation, with the same targets and entitlements. All request and response bodies are JSON. Success has { "success": true, "data": ... }; errors have { "success": false, "message": "..." }. Store the plaintext key in a secret manager. A key cannot be retrieved again; revoke and replace it if leaked. For a CI pipeline, use long-polls with wait=55 rather than busy-polling.

Endpoints

List targets

Returns a page of your organisation’s testing targets, whatever their status (pending, registered, failed). The default page size is 50; set limit up to 100 and increase skip by that limit for each subsequent page. External targets come before sandboxes, so keep paging until you find the sandbox with the role and OCPI version you need, or a page contains fewer rows than your limit. Sandbox CPO, Sandbox eMSP and Sandbox PTP endpoints ("kind": "sandbox") are already registered. Use the matching target’s _id to test against our reference implementation. Runs require status: "registered".

Register your own endpoint as a target

ourRole/theirRole accept CPO, EMSP, HUB, NAP, NSP, OTHER or SCSP. Choose partnerEnvironment explicitly: test permits deliberate writes and commands; production permits GET-only conformance runs. The example uses a test system. ocpiVersion accepts 2.1.1, 2.2, 2.2.1 or 2.3.0. Then run the OCPI credentials handshake:
A 502 response carries the handshake failure detail in message. Fix the endpoint and re-register. Status must be registered before runs.

Start a conformance run

Optional body field module restricts the run to one suite module (e.g. "credentials"). wait (seconds, max 55) long-polls: if the run finishes inside the window you get 200 with the verdict; otherwise 202 with ids to keep polling. Terminal (200):
Still running (202):

Poll until terminal

Long-polls up to 55s and returns the current report; loop while data.status is queued, running, recovering or recovery_required. Recovery can need operator action; the job timeout still bounds the wait. The terminal report includes per-check evidence in data.suites[], including any recorded exchanges, alongside data.summary.

Exit contract

A pass requires a completed run, at least one graded check (passed + failed > 0) and zero failed checks. Warnings remain nonfatal because declining a recommendation breaks no requirement. A run that graded nothing proves nothing and exits 2, for no answer. Graded failures exit 1, for a broken requirement. An unfinished run with no failures also exits 2. Read data.suites[] for the evidence. A completed run with zero failures can still have incomplete scope. data.withheldModules names modules excluded by the plan, and data.summary.unobservable counts checks that nothing could grade. The recipe returns success only for the checks that ran, listing both limits. An absent unobservable means zero for legacy reports; malformed counters or module lists are no answer, not a pass.

Share the report

Successful share creation returns:
Anyone with the link can read the report (owner identity is stripped):
The recipe prints a shared report for completed or failed runs with valid counters, including graded failures. Runs that stop without results return no answer; inspect run history for their evidence.

Run history

Cursor-paginated, newest first (meta.nextCursor when more pages exist). List rows omit the heavy suites payload. Fetch a run by id for detail.

What changed since the last run

Compares a conformance run with the previous comparable run against the same endpoint, OCPI version and module. data.regressions lists the checks that passed then and fail now, and data.recoveries the reverse. data.comparable is false when there is nothing to compare with, and data.incomparableReason says why; that is never a pass.

Read a connection’s schedule

Returns the recurring run set on a connection: its cadence, whether it is paused and why (pausedReason), the next run (dueAt) and what the last scheduled run found (lastOutcome). A connection with no schedule answers 404. Schedules are set in the app.

Usage and quota headroom

Returns your plan tier and current usage against its ceilings: active connections, API keys, in-flight runs and the per-minute run rate, runs started this month, and the report-retention window in days. Useful before a matrix build fans out: if runs.inflight is at runs.inflightCap, stagger the jobs instead of burning retries on 429s.

Error semantics

GitHub Actions recipe

timeout-minutes is the backstop for polling. The 45-minute example gives a direct OCPI 2.2.1 CPO run time to finish and leaves room for queueing. POST requests retry only an explicit 429, with at most three retries and 120 cumulative seconds of server-provided integer Retry-After delays. Missing, malformed or over-budget delays stop the job. Other HTTP or transport failures do not retry a POST: the server may already have accepted work, so inspect run history before starting again. Sharing must succeed with a nonempty token before the recipe prints its link. Run the script with Bash, curl and jq installed. In GitLab CI or Jenkins, use Bash, configure the same job timeout and adapt the secret injection to that system.