v2.6.0

[2.6.0] — 2026-07-20

Added

• Proactive, adaptive request pacing (in-memory token bucket). Sustained
  load now glides within TeslaFi's advertised rate limit instead of bursting
  to the wall and stalling. A token bucket is derived from the rate_limit
  object TeslaFi embeds in responses — capacity ≈ the advertised limit and
  refill ≈ limit / reset_in_seconds tokens per second (e.g. a 15-token bucket
  refilling one token every ~4 s for the documented 15-requests-per-60 s
  budget). Each outbound request spends one token. This intentionally allows an
  initial burst up to the advertised budget — so short, typical
  interactions (a handful of tool calls) stay fast — and then converges to the
  steady refill rate under sustained use, pacing toward the window reset
  rather than slamming into remaining: 0. The newly-captured advertised
  limit (previously ignored) drives the sizing.

Changed

• Request spacing is now composed, not just fixed. The pre-send wait is the
  maximum of three signals: the existing 1 s floor between consecutive
  sends, the new token-bucket wait, and the existing remaining: 0
  reset-window backstop (capped at 60 s). All three remain in force — the 1 s
  floor, the remaining === 0 hard wait, and the HTTP 429 Retry-After retry
  handling are unchanged backstops beneath the token bucket.

Notes

• Internal-only. No tool, parameter, or output shape changed — this is
  purely internal pacing. State is in-memory per process, with zero disk
  writes and no new network calls.
• Self-tuning with a safe fallback. Capacity/refill are re-derived from the
  last-observed advertised limit+window. Before any rate_limit has been
  observed, or when the advertised limit/window is missing, zero, non-finite, or
  otherwise nonsensical, pacing falls back to the previous 1 s-floor-only
  behavior (never assuming a limit, never dividing by zero). Any single computed
  wait is capped at one window and never more than 60 s so a garbled advertised
  value cannot hang a request.

---

Install: download the attached teslafi-mcp.mcpb and double-click to install into Claude Desktop, or see the README for manual setup.
SHA-1 of the bundle: ec1ed409b99eea4e2a8c3be04d1cfca6c0655efb

v2.5.0

[2.5.0] — 2026-07-20

Fixed

• Frequently-visited untagged locations were invisible to
  tesla_get_locations. The location aggregation only counted endpoints with
  a non-empty locationName/startingLocationName, and only TeslaFi
  dashboard-tagged locations ever carry one. Any place the owner visited — even
  daily — that they had not separately geofenced+named inside TeslaFi Locations
  settings was silently dropped, with no count and no ranking. (A car-nav
  favorite is not the same thing as a TeslaFi-tagged Location and does not carry
  a locationName.) The aggregation now falls back to the stable per-address
  key (locationKey for the end endpoint, startingLocationKey for the start
  endpoint) when the name is empty, deriving a readable display name from the
  address (address / startingAddress) — so a place you visit often is
  surfaced with the correct visit count even when you never tagged it. Named and
  inferred groups are namespaced so they never collide, and an endpoint is
  skipped only when its name and key and address are all empty. Existing
  visits / as_start / as_end counting and the sort by visits descending
  are unchanged.

Added

• named flag on every location entry. LocationSummary gains a boolean
  named field: true for owner-tagged TeslaFi locations (Home, Work, named
  Superchargers), false for addresses inferred from repeated raw visits. It is
  included in both the structured/JSON output and the markdown view (inferred
  entries are marked "(inferred from address)" with an explanatory note); JSON
  output stays pure and parseable.

Notes

• tesla_get_charges was checked for the same gap. Its per-charge summary
  already falls back to the address (locationName || address || "Unknown"),
  so a charge at an untagged address is still labeled — no change was needed
  there; a regression test now pins that behavior.
• Behavior change (new inferred entries appear and every entry carries a new
  named field) but backward-compatible: all previously-returned tagged
  entries are still present, and the new field is purely additive. No tool
  parameters changed.

---

Install: download the attached teslafi-mcp.mcpb and double-click to install into Claude Desktop, or see the README for manual setup.
SHA-1 of the bundle: 4ab9ed3d4c8692d7a825b28c0052e34afbe222f4

v2.4.0

[2.4.0] — 2026-07-20

Added

• Opt-in update check (off by default). A new TESLAFI_CHECK_UPDATES
  environment variable (truthy only for true/1/yes, case-insensitive —
  the same rule as TESLAFI_ENABLE_COMMANDS) enables a lightweight check for a
  newer GitHub release. When on, the server makes a single unauthenticated
  GET to api.github.com/repos/Zenotech-bv/teslafi-mcp/releases/latest,
  compares the release tag to the running version using a numeric
  major.minor.patch comparison (so 10.0.0 > 9.9.9; a leading v and any
  pre-release suffix are ignored), and — only when the release is strictly newer
  — appends a single one-line notice to data-tool markdown output, at most once
  per session. The notice never appears in response_format: "json" output, so
  JSON stays pure and parseable.
• enable_update_check extension config. manifest.json gains a plain
  boolean user_config.enable_update_check (title "Check for updates", default
  false), wired into server.mcp_config.env as TESLAFI_CHECK_UPDATES.
• Single VERSION source of truth. The running version is now a single
  exported constant consumed by both the McpServer constructor and the update
  comparison, kept in lockstep with package.json and manifest.json by a test.

Security / privacy

• The update check is a standalone request that never touches the TeslaFi
  request pipeline: it sends no TeslaFi token, no Authorization
  header, and no TeslaFi data, and reuses the same maxRedirects: 0 /
  proxy: false hardening. It is fire-and-forget (every failure — network
  error, timeout, 404 for a private repo, 403 rate limit, malformed JSON — is
  swallowed), in-memory only (zero disk writes), and off by default: when
  disabled, no GitHub request is ever made and the only network destination
  remains www.teslafi.com. A private repository returns 404 unauthenticated,
  so the notice only appears once the repository is public.

Not breaking: no tool parameters changed and all existing behavior is preserved.

---

Install: download the attached teslafi-mcp.mcpb and double-click to install into Claude Desktop, or see the README for manual setup.
SHA-1 of the bundle: fc08991c9dc455a823179966a14deac119448d45

v2.3.0

[2.3.0] — 2026-07-20

Added

• Selectable units on every data-returning tool. tesla_get_current_data,
  tesla_get_drives, tesla_get_charges, tesla_get_trips_for_month,
  tesla_analyze_home_work_trips, and tesla_get_locations take an optional
  units argument — "metric", "imperial", or "both" — and every response
  now carries a top-level units field stating the applied system.
• Effective-unit resolution. When units is omitted it is resolved in
  order: the explicit argument, then the new TESLAFI_UNITS env var
  (metric/imperial, case-insensitive; other values ignored), then the
  TeslaFi account measure (remembered in memory from the most recent feed
  response), then metric. The OS locale is not visible to the server, so
  TESLAFI_UNITS is the documented way to pin a default.
• TESLAFI_UNITS configuration. Documented in the README config table and
  added to manifest.json user_config as an "Automatic / Metric / Imperial"
  select (default Automatic), wired into server.mcp_config.env.
• Explicit unit fields everywhere. Every physical quantity in the
  normalized/summary/structured output carries its unit as a field-name suffix
  (_km/_mi, _kmh/_mph, _c/_f, _bar/_psi, _kwh, _m/_ft,
  _pct); with units: "both" both suffixed fields are emitted. Every
  markdown renderer prints the unit next to every number. Live-feed values
  (Tesla-native regardless of account) are converted to the chosen system,
  while history records (which already carry both systems) have the pre-existing
  field selected — never re-converted — to avoid double conversion/rounding
  drift. Percentages, kWh, and currency are never converted, only labeled.
• History coverage metadata. tesla_get_drives, tesla_get_charges, and
  tesla_get_trips_for_month now return date_range ({ earliest, latest },
  computed from the full fetched dataset before pagination — no extra network
  call) and coverage_note, so a client cannot mislabel a partial history
  fetch as a "lifetime" figure. The markdown reports counts honestly too
  (e.g. 2,526 drives on record (since 2023-12-18); showing 1–50).

Changed

• BREAKING — output field names are now unit-suffixed. Compact drive/charge
  summaries, trip objects, and the live-feed structured output changed shape:
  physical fields are renamed to unit-suffixed names for the applied system
  (e.g. the old drive distance_km/distance_mi pair is now emitted per the
  chosen units; tesla_get_current_data structured output is now a normalized
  object with suffixed fields plus passthrough of non-physical fields, rather
  than the raw feed record). include_raw still embeds the untouched raw
  record. Clients that read fixed field names must update to the suffixed names.

---

Install: download the attached teslafi-mcp.mcpb and double-click to install into Claude Desktop, or see the README for manual setup.
SHA-1 of the bundle: ff7273d1f4de37abc964fe61c80bbe2f9d226c9a

v2.2.0

[2.2.0] — 2026-07-19

Added

• Command allowlist aligned with TeslaFi's official API documentation.
  The allowlist (now 24 commands) matches the command list on TeslaFi's
  logged-in API settings page, upgrading the provenance from community
  documentation (ha-teslafi) to the official source.
• Six new commands plus a quota probe: wake (resume TeslaFi polling:
  idle timer 0, polling on), sleep (optional time in minutes),
  enableLogging, disableLogging, seat_heater (heater: 0/1/2/4/5 —
  there is no seat 3; level: 0–3), preset (number, optional
  statement), and command_count, which returns the current command/wake
  quota counters and is treated as read-only (retried like a read, never like
  a mutating command).
• no_wake parameter on tesla_send_command, mapping to TeslaFi's
  noWake=true global modifier: the command is not sent at all when the
  vehicle is sleeping. Mutually exclusive with wake_seconds (a clear local
  error is raised when both are set). noWake also joined the reserved
  parameter keys.
• Local parameter value validation. Out-of-range values are rejected
  before any network request (so they never cost a quota request):
  seat_heater heater must be one of {0, 1, 2, 4, 5} and level 0–3,
  charge_limit_soc 50–100, charging_amps 1–80, set_temps temp
  15–28 °C, sleep time 1–1440 minutes, preset number a small positive
  integer.
• Monthly quota surfacing. TeslaFi allows 500 command + 50 wake requests
  per month (reset on the 1st; a command while the car is asleep costs
  1 command + 1 wake). When a response embeds tesla_request_counter, the
  tesla_send_command and tesla_get_current_data markdown output now ends
  with a Quota: N/500 commands, M/50 wakes used this month line, and the
  counter rides along in structured output. The counter is preserved across
  { response: ... } envelope unwrapping, and its absence never causes a
  failure.
• Quota tracking with high-usage warnings. The server keeps the
  last-known tesla_request_counter (commands/wakes + timestamp) in memory,
  updated from every command and data response that carries one. Once
  usage reaches 80% of an allowance (≥ 400/500 commands or ≥ 40/50 wakes),
  every tesla_send_command markdown result is prefixed with a warning line
  (e.g. WARNING: 43/50 monthly wake requests used). Absent or malformed
  counters never cause a failure. All state is in-memory, per process.
• Proactive rate-limit pacing. The rate_limit object embedded in data
  responses (remaining, reset_in_seconds) is remembered; when the last
  response reported remaining: 0, the next outbound request waits out the
  remainder of the reported reset window (capped at 60 s) instead of firing a
  guaranteed HTTP 429. The existing 1 s minimum spacing and all existing
  retry/429 handling are unchanged.
• Quota-exhausted errors are terminal and actionable. TeslaFi errors
  indicating the monthly command or wake allowance is exceeded (matched
  permissively — the exact payload is undocumented) are never retried and are
  surfaced with an actionable message: the command was NOT sent, the last
  known usage when available, allowances reset on the 1st of the month, and
  extra requests are purchasable from TeslaFi support. Auth and
  malformed-request errors are never classified this way.
• Asleep feedback. When a command is not sent because the car was asleep
  under noWake (matched permissively on TeslaFi's unsent/asleep phrasing),
  the tool output states the command was NOT sent and explains the choices:
  retry with wake_seconds (costs 1 of the 50 monthly wakes) or wait until
  the car is awake.

Changed

• BREAKING — car commands no longer wake a sleeping car by default. For
  the car-affecting commands (honk, lights, climate, seat/wheel heaters,
  presets, charging, charge port, locks, sentry), passing neither
  wake_seconds nor no_wake now sends noWake=true automatically: if
  the car is asleep the command is not sent and no wake request is spent.
  Pass wake_seconds to wake the car, or no_wake: false to restore
  TeslaFi's own default (wake + pause up to 15 s). Account-side commands
  (wake, wake_up, sleep, enableLogging, disableLogging,
  command_count) are exempt and never take noWake; no_wake: true on
  them is now rejected locally.
• BREAKING: wake_seconds is now capped at 60 (was 120), matching
  the official documented maximum for the &wake=X modifier. Calls passing
  61–120 are rejected at both the schema and the service layer.
• Documentation (README, tool description) now cites TeslaFi's official API
  documentation as the allowlist source and documents the command quotas,
  the seat-heater position map, and the conditioning prerequisites for
  seat_heater, steering_wheel_heater, and preset.

Removed

• BREAKING: honk_horn — not an official TeslaFi command. Use honk.
• BREAKING: driver_temp / passenger_temp on set_temps — the
  official documentation supports only temp (Celsius, one decimal allowed).

---

Install: download the attached teslafi-mcp.mcpb and double-click to install into Claude Desktop, or see the README for manual setup.
SHA-1 of the bundle: 93a52ff08ebedaf0aa1203d4f6c30a77997b8bea

v2.1.0

[2.1.0] — 2026-07-19

Added

• Local command allowlist with per-command parameter validation.
  tesla_send_command validates against a fixed allowlist of 18 TeslaFi
  commands. Unknown commands, parameter keys not valid for the given command,
  and the reserved keys token / command / wake are rejected by the server
  itself, before any network request. The allowlist lookup uses an own-property
  check, so prototype-inherited names (constructor, __proto__, ...) are
  rejected like any other unknown command. The allowlist is sourced from
  community documentation (the ha-teslafi Home Assistant integration) plus
  commands verified against a live TeslaFi account, since TeslaFi's own API
  docs are login-gated.
• Opt-in gating for vehicle commands. The TESLAFI_ENABLE_COMMANDS
  environment variable (truthy values: true, 1, yes; case-insensitive)
  controls whether tesla_send_command does anything. See the BREAKING notes
  below.
• Retry with backoff and request spacing. Outbound TeslaFi requests are
  spaced at least 1 s apart. Transient failures are retried up to 2 times with
  exponential backoff and jitter. An HTTP 429 is retried only when TeslaFi
  supplies a Retry-After of at most 60 s; a 429 without that header, or
  asking for a longer wait, fails immediately with a rate-limit error. Reads
  are retried freely on network errors and 5xx; commands are only retried for
  failures where the request provably never reached TeslaFi (DNS/connection
  errors, 429 rejections with an honored Retry-After) — never for mid-flight
  errors, so a command cannot run twice.
• Redirects are never followed and proxies are never used. Every outbound
  request sets maxRedirects: 0 and proxy: false, enforcing that the only
  network destination is www.teslafi.com over HTTPS: a server-issued
  redirect surfaces as an error instead of re-sending the token-bearing URL to
  another host or over plain HTTP, and HTTP_PROXY/HTTPS_PROXY environment
  variables are ignored instead of routing the request through a proxy.
• In-memory TTL cache for history. history.php responses (drives and
  charges) are cached for 60 seconds by default, configurable via
  TESLAFI_CACHE_TTL_SECONDS (0 disables). Keyed by endpoint + date range;
  failures are not cached; nothing is ever written to disk.
• include_raw parameter (default false) on tesla_get_drives,
  tesla_get_charges, and tesla_get_trips_for_month, to opt in to embedding
  the full raw TeslaFi records.
• Bounded analysis output. tesla_analyze_home_work_trips caps the
  embedded matched-trip list at 500 entries and reports trips_truncated;
  summary counts still cover the entire month.
• CI and tests. A GitHub Actions workflow builds and tests on Node 20 and
  22, plus a build-only check on Node 18 (the oldest supported runtime; the
  vitest toolchain requires Node >= 20), and a vitest-based npm test script.

Changed

• BREAKING: tesla_get_drives and tesla_get_charges structured output
  now returns compact per-record summaries (drives: id, date,
  start_location, end_location, distance_km, distance_mi; charges:
  id, date, location, energy_added_kwh) instead of full raw TeslaFi
  records. Pass include_raw: true to get each raw record embedded under a
  raw key.
• BREAKING: the MCP server name changed from tesla-mcp-server to
  teslafi-mcp. Update any client configuration that references the server by
  name.
• BREAKING: tesla_send_command is now disabled by default. Set
  TESLAFI_ENABLE_COMMANDS=true (or enable the "Enable vehicle commands"
  extension setting) to use it; commands must additionally be enabled
  per-command on TeslaFi's Commands tab. When disabled, the tool returns an
  instructive error without contacting TeslaFi.

Fixed

• Parameter injection / token override. User-supplied command parameters
  could override the token and command query parameters sent to TeslaFi.
  Reserved keys are now rejected outright, and the token is applied after user
  parameters so it can never be overridden.
• Token leak via fatal-error logging. Axios errors embed the full request
  URL — including the token= query parameter — and were previously logged
  verbatim on fatal errors. All errors are now converted to sanitized
  messages: the token value, token= query fragments, and TeslaFi request
  URLs are redacted before anything is logged or returned.
• Malformed responses silently treated as empty. HTML error pages and
  plain-text bodies from history.php were parsed as an empty result set,
  making outages look like "no drives found". Malformed payloads now raise a
  descriptive error, distinguished from genuinely empty results. The same
  guard applies to tesla_send_command: a plain-text feed.php body (e.g. a
  login page) is surfaced as an error — with a warning that the command may
  still have executed — instead of being rendered as a successful result.
• Wake timeout mismatch. The HTTP timeout did not account for TeslaFi's
  wake-and-wait window, so commands using a wake wait could time out before
  TeslaFi answered. wake_seconds now extends the request timeout to cover
  the wait.
• Understated dependency floors. @modelcontextprotocol/sdk was declared
  as ^1.6.1 although the server uses registerTool with ZodObject input
  schemas, which requires SDK >= 1.23.0 (which in turn requires zod >= 3.25).
  The declared ranges now match the real minimums (^1.23.0 / ^3.25.0).
• Misleading startup log. The startup message misrepresented the server's
  configuration. It now reports the active token source variable
  (TESLA_API_TOKEN or TESLAFI_API_TOKEN) and whether vehicle commands are
  enabled — never the token value itself.

Removed

• Personal data scrubbed. Real tokens, real location names, and
  machine-specific paths were removed from code, docs, and examples;
  placeholders such as "Home" and "Acme HQ" are used throughout.
• Personal helper scripts moved out of the repository. The generic
  field-inspection helpers (scripts/inspect-teslafi.mjs,
  scripts/debug-history.mjs) remain, and both redact the token in their
  output.

---

Install: download the attached teslafi-mcp.mcpb and double-click to install into Claude Desktop, or see the README for manual setup.
SHA-1 of the bundle: 5b0c35af62927f7e01ed417d36a7033a16c6f7be