AIPocket

How much prepaid credit is left at each of your LLM API providers — in one place, from one command.

A single static binary. No runtime, one dependency, and it never stores your API keys.

$ aipocket

  PROVIDER               BALANCE  STATE   NOTE
  ──────────────────  ──────────  ──────  ────────────────────────────────────────────
  Anthropic                    —  manual  no balance API; key valid
  DeepSeek             18.42 USD  ok
  fal.ai               24.50 USD  ok
  Groq                 25.00 USD  manual  no balance API; key valid; user-maintained figure (2026-08-01, 17 days ago)
  Moonshot AI (Kimi)   49.59 USD  ok
  OpenAI                       —  manual  no balance API; key valid
  OpenRouter           37.65 USD  ok      topped up 50.00 / spent 12.35
  SiliconFlow          12.80 USD  ok      inferred field
  ──────────────────  ──────────  ──────  ────────────────────────────────────────────
  verified            130.16 USD   documented fields only
  inferred             12.80 USD   read from undocumented response shapes
  manual               25.00 USD   your own figures, unverifiable

  outside the verified total: siliconflow (inferred field), groq
    (user-maintained)

  11 providers have no credential configured: cerebras, deepinfra, entrim,
    gemini, mistral, nebius, neuralwatt, novita, replicate, together, xai
    (--all)

The figures are illustrative; the layout is not — it is printed by the same renderer the binary uses. Providers with no credential configured collapse into that last line rather than filling the table with rows saying a variable is not set; --all brings them back.

Why this exists

Most "LLM cost dashboards" track spend that passes through them. That is a different thing from how much money is left in your account, and it is the number you actually want before kicking off a long job.

The awkward truth is that only some providers expose a balance at all. AIPocket's design commitment is to be honest about which is which.

There are three totals, not one. A figure read from a documented field, a figure inferred from an undocumented response shape, and a figure you typed into your config are three different kinds of claim. Summing them would give a guess the same authority as a fact, so verified contains only the first kind. Non-USD balances are reported but never converted — AIPocket has no business guessing an exchange rate.

The tool also refuses to be confidently wrong. A response that no longer matches the manifest, a total - used that comes out negative, a NaN where money should be: each becomes a visible error, never a green 0.00.

confidence	meaning
official	the provider documents both the endpoint and the field
undocumented	the endpoint exists, the response shape was inferred — verify it once against the console
no-api	the provider publishes no balance endpoint; AIPocket checks the key works, the figure is yours to maintain

Providers

Twenty, and the honest summary of the state of LLM billing APIs is the confidence column: five providers publish a balance endpoint you can trust, two answer one whose response shape had to be inferred, and thirteen publish nothing at all — for those AIPocket checks the key and reports whatever figure you choose to keep.

provider	id	confidence	reports
Anthropic	anthropic	no-api	key check only
Cerebras	cerebras	no-api	key check only
DeepInfra	deepinfra	no-api	key check only
DeepSeek	deepseek	official	balance
Entrim.ai	entrim	no-api	key check only
fal.ai	fal	official	balance
Google Gemini	gemini	no-api	key check only
Groq	groq	no-api	key check only
Mistral AI	mistral	no-api	key check only
Moonshot AI (Kimi)	moonshot	official	balance
Nebius Token Factory	nebius	no-api	key check only
Neuralwatt	neuralwatt	undocumented	balance (inferred)
Novita AI	novita	official	balance
OpenAI	openai	no-api	key check only
OpenRouter	openrouter	official	balance
Replicate	replicate	no-api	key check only
SiliconFlow	siliconflow	undocumented	balance (inferred)
Straitly	straitly	no-api	key check only
Together AI	together	no-api	key check only
xAI	xai	no-api	key check only

A provider that ships a balance API gains one here; that is data, not a release. Adding one is a single YAML file — see Adding a provider.

Install

# Go
go install github.com/naturalmoods/aipocket/cmd/aipocket@latest

# or download a binary for linux / macOS / Windows / FreeBSD
# from the releases page — no runtime required

Either way, aipocket version reports the release you actually have: a go install build takes it from Go's own build info, so it is never just "dev". That is the one question worth being able to answer after a security advisory.

Use

aipocket                        # all configured providers
aipocket openrouter deepseek    # just these
aipocket --all                  # include the ones with no credential configured
aipocket --json                 # machine-readable
aipocket providers              # what's known, and where each key is read from
aipocket audit                  # every host that would be contacted, before you run anything
aipocket probe entrim           # look for an undocumented balance endpoint
aipocket probe entrim --dry-run # ...or just list the requests it would make
aipocket config add             # pick a provider, say where its key comes from

With no configuration at all, AIPocket reads each provider's conventional environment variable (OPENROUTER_API_KEY, DEEPSEEK_API_KEY, …). That is enough for CI and for a first try.

Configuration

aipocket config path prints the location (~/.config/aipocket/config.yaml on Unix, %AppData%\aipocket\config.yaml on Windows).

The config file holds no secrets. It holds instructions for obtaining them:

providers:
  openrouter:
    key: command:op read op://Private/OpenRouter/credential   # 1Password
  deepseek:
    key: command:pass show llm/deepseek                       # pass / gopass
  neuralwatt:
    key: file:~/.secrets/neuralwatt                           # 0600 enforced
  entrim:
    key: env:ENTRIM_API_KEY
    balance_url: https://api.entrim.ai/v1/credits             # see `aipocket probe`
  groq:
    manual: 25.00        # no balance API; reported separately, never in the total
    as_of: 2026-08-01    # optional: when you read it. The row shows the age.

aipocket config add writes that block for you to paste: it lists the providers this build knows, asks which of the three forms the key comes from, and checks the answer with the same rule the config loader applies. It prints — it does not edit the file, and it never asks for the key. Re-marshalling the YAML would drop the comments that explain each entry, and a subcommand that took a key would put it in the shell history on its way into the one file that promises to hold none.

Shelling out to whatever secret manager the machine already has is both more portable and more secure than bundling a keyring library. Concretely: under WSL there is no D-Bus session and no running keyring daemon, so a cross-platform keyring dependency simply fails there — while op, pass and bw all work.

Use from Claude, Cursor, or any MCP client

claude mcp add aipocket -- aipocket mcp

or, in an MCP client config:

{ "mcpServers": { "aipocket": { "command": "aipocket", "args": ["mcp"] } } }

Two read-only tools: get_balances and list_providers. The server tells the model, in its initialization instructions, not to present an undocumented or no-api figure as authoritative — the confidence level travels with the number all the way into the conversation.

Providers' own error bodies are not forwarded to the model. "Ignore your previous instructions" is a legal HTTP 402 body, and a tool result is a channel into a model's context; what travels is AIPocket's own account of the failure, which no remote service chose the wording of. The table still shows you what the provider said.

Security

Short version:

• AIPocket is not a secret store. It resolves credentials at runtime and holds them in memory for one run. A literal key in the config file is refused when the file is loaded, and the refusal does not echo it. env: names must be upper case for that reason: gsk_…, nvapi_… and hf_… are all valid POSIX variable names, so accepting one would print a mis-pasted key back out.
• Every resolved secret is registered with a redactor that runs before any message is shortened, so a key echoed in a provider's 401 body cannot survive as a truncated prefix in a terminal, a CI log or an assistant transcript. There are tests for exactly that, including one for a key straddling the truncation boundary.
• Provider manifests are compiled into the binary and are never fetched at runtime. A manifest says where an API key gets sent; a remotely-updatable registry would be a "send your credentials here" primitive.
• https only, redirects are refused, response bodies are capped, every request has a deadline.
• aipocket audit prints the complete network footprint before you trust it with anything — including the requests aipocket probe would make and any HTTPS_PROXY your environment imposes, since a proxy becomes another host that receives the credential.
• Anything displayed is stripped of control characters. A provider error body is bytes of someone else's choosing on their way to your terminal, and an ANSI escape in one can repaint the total printed above it.

The single most valuable thing you can do is not cryptographic: set a spend limit at the provider, and use a read-only or scoped key wherever one is offered. A balance checker holding a full inference key is the actual exposure.

Full threat model and reporting instructions: SECURITY.md.

Adding a provider

One YAML file. No Go.

id: acme
name: Acme AI
status: official
console: https://acme.ai/billing
docs: https://acme.ai/docs/credits
auth: {type: bearer, env: ACME_API_KEY}
balance:
  url: https://api.acme.ai/v1/credits
  amounts:
    - {path: $.credits_remaining, currency: USD}

aipocket probe <provider> will search the conventional paths on a provider's own host and print a manifest block ready to paste. See CONTRIBUTING.md.

Compatibility

From v1.0.0, four things are contracts and will not break within the major version. New fields and new providers may be added; nothing listed here will be renamed, removed, or change meaning.

• The config file. A config that works today keeps working.
• --json. Field names and their meaning are stable, including the split between total_verified_usd, total_inferred_usd and total_manual_usd.
• The MCP tools. get_balances and list_providers, their input schemas, and the confidence value travelling with every figure.
• Exit codes. 0 success, 1 at least one provider errored, 2 usage error — scripts depend on these.

Three things are deliberately not contracts:

• The table. It is for humans and will be reformatted whenever that makes it clearer. Parse --json.
• Which providers exist, and what each one can report. Providers are data. A provider that ships a balance API gains one; a provider that withdraws one loses it, and its status changes accordingly. That is the provider's doing, not an API break — and it is the reason confidence is in the output rather than in the documentation.
• Anything a provider says. provider_message carries a remote service's own words. Its presence and shape are stable; its contents are theirs.

Dependencies

One: gopkg.in/yaml.v3. The JSONPath subset, the HTTP layer and the MCP server are all in-tree — a binary that reads your billing credentials should have a dependency tree you can audit in an afternoon.

Licence

MIT.