PromptLatch

Redact secrets before prompts reach an LLM provider.

PromptLatch is a local proxy and a Python redaction library. Use the proxy for coding agents and clients that support a custom base URL. Use the library to redact prompt text, message arrays, SDK parameters, and structured request payloads before your application calls an LLM SDK.

Scanning and monitoring stay local. PromptLatch sends no telemetry. Server mode downloads public LiteLLM pricing metadata for cost estimates; disable this with observability.pricing_enabled: false.

Website: https://bvolpato.github.io/promptlatch/

Agent integration prompt: PROMPT.md

Choose a mode

Coverage

Default rules cover provider keys, personal access tokens, passwords, JWTs, signed URLs, URL credentials, PEM/PGP private keys, and common secret fields such as api_key, token, authorization, password, signed_url, and credentials.

Detection uses deterministic provider rules plus your exact-tail or regex rules. Entropy-only matching is disabled to avoid unpredictable false positives.

Security boundary

• Request bodies and query parameters are scanned before they leave your machine.
• Audit logs record redaction counts and rule names without storing secret values.
• Local request history stores redacted bodies and metrics. Redacted prompts can still contain private information; use metadata-only storage when needed.
• When request history is enabled, it can also retain a small allowlist of upstream error fields after full redaction. Full upstream response bodies and raw headers are not stored.
• PromptLatch strips cookies and secret-named client headers. Provider credentials are added from config or dedicated X-Target-* headers after that filtering step.
• Unknown private token formats need a custom exact-tail or regex rule.

See SECURITY.md for deployment defaults and remaining limits.

Install

Python package: promptlatch on PyPI.

Homebrew:

brew tap bvolpato/tap
brew install promptlatch
promptlatch version

uv:

uv tool install promptlatch
promptlatch doctor

Python library:

uv add promptlatch
# or
python -m pip install promptlatch

Source:

git clone https://github.com/bvolpato/promptlatch.git
cd promptlatch
uv sync --extra dev --locked
uv run promptlatch doctor

Install this checkout and the codex-latch / claude-latch commands:

./install.sh

Requires uv. Installs into uv's tool bin directory without editing Codex, Claude Code, shell startup files, or proxy configuration. See coding-agent shims for server setup.

ASGI servers can load promptlatch.asgi:app directly. Importing CLI or proxy helpers does not load user config until a command or app requests it.

Upgrading from PromptCloak

Version 0.2 renamed package, command, environment variables, config directory, container image, and Helm chart. Old Python imports, environment variables, and default config path remain compatible through 0.2.x and emit migration warnings.

Remove old uv tool so stale promptcloak command cannot shadow new install:

uv tool uninstall promptcloak

Move local config before switching services:

mv ~/.config/promptcloak ~/.config/promptlatch

Existing Helm releases can upgrade in place without changing immutable selectors or rotating chart-managed proxy key:

helm upgrade <existing-release-name> ./charts/promptlatch \
  --set migration.preserveLegacyNames=true

Keep migration flag on later upgrades for that release. Fresh installs should omit it and use release name promptlatch. Existing external Secrets may keep PROMPTCLOAK_SERVER_API_KEY for this upgrade, then rename key to PROMPTLATCH_SERVER_API_KEY.

Run proxy

Configure an upstream, keep its key in your shell, and start PromptLatch:

promptlatch init --target-base-url https://api.openai.com/v1
export OPENAI_API_KEY="<openai-upstream-key>"
promptlatch serve

Point clients at http://127.0.0.1:8000/v1. For example:

curl http://127.0.0.1:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-6-sol",
    "messages": [{
      "role": "user",
      "content": "Here is my .env: OPENAI_API_KEY=<api-key-like-value>"
    }]
  }'

Provider receives OPENAI_API_KEY=[REDACTED_SECRET] in request content.

Verify redaction

A model reply cannot verify the forwarded request. Send a fixture token to an echo endpoint and inspect the echoed body:

FAKE_GEMINI_KEY="AI""zaSyFixtureToken000000000000000000000"

curl --compressed -fsS http://127.0.0.1:8000/post \
  -H "X-Target-Base-URL: https://postman-echo.com" \
  -H "Content-Type: application/json" \
  --data "$(jq -nc --arg key "$FAKE_GEMINI_KEY" \
    '{messages:[{role:"user",content:("GEMINI_API_KEY=" + $key)}]}')" \
  | jq -r '.data.messages[0].content'

Expected output:

GEMINI_API_KEY=[REDACTED_SECRET]

Audit logs omit matched values and include counts and rule names. Homebrew users can run PromptLatch as a background service after configuring its environment:

brew services start bvolpato/tap/promptlatch

Use as a Python library

Use PromptLatch without starting the proxy. Redact each prompt or request value in the same process, immediately before passing it to your SDK or HTTP client. Install the SDK you use separately; PromptLatch has no OpenAI, LiteLLM, LangChain, Anthropic, or LlamaIndex SDK dependency.

Choose a helper for the shape of the value you send:

All four helpers are exported from promptlatch. They return the safe value. They do not send the request or change an SDK client for you. Pass the returned value to the SDK call; never pass the original value after redaction. If redaction raises an error, do not retry by sending the original input.

Redact a standalone prompt

Use redact_text when the prompt is assembled as a string, before passing it as input, prompt, or another SDK field:

from openai import OpenAI
from promptlatch import redact_text

client = OpenAI()
prompt = "Inspect this config: OPENAI_API_KEY=example-secret-value-123456"

response = client.responses.create(
    model="gpt-6-sol",
    input=redact_text(prompt),
)

Redact a message array

Use redact_messages for chat-style messages arrays. It supports mapping-based messages, (role, content) tuples, Pydantic/OpenAI message models, and LangChain message objects:

from openai import OpenAI
from promptlatch import redact_messages

client = OpenAI()
messages = [
    {"role": "system", "content": "Summarize the user's log."},
    {"role": "user", "content": "Authorization: Bearer FixtureToken000000000000000000000"},
]

safe_messages = redact_messages(messages)
response = client.chat.completions.create(
    model="gpt-6-sol",
    messages=safe_messages,
)

For non-Pydantic message objects, PromptLatch copies the object and redacts its content field. For Pydantic messages, it recursively redacts fields, including tool calls, and preserves the model type. Inputs are left unchanged.

Redact SDK parameters or structured payloads

Use redact_params when assembling SDK keyword arguments. It returns a dictionary you can unpack directly into the SDK call. A messages argument is treated as a message array; other values, including input, tools, and nested parameter objects, are scanned recursively.

from openai import OpenAI
from promptlatch import redact_params

client = OpenAI()
safe_params = redact_params(
    model="gpt-6-sol",
    input="Summarize this: GITHUB_TOKEN=example-token-value-123456",
)
response = client.responses.create(**safe_params)

For raw JSON-compatible bodies, use redact_payload and pass its result to the transport. It recursively scans strings in mappings and lists without reshaping the request schema:

import os
import httpx
from promptlatch import redact_payload

payload = {
    "model": "gpt-6-sol",
    "input": [
        {
            "role": "user",
            "content": [
                {
                    "type": "input_text",
                    "text": "OPENAI_API_KEY=example-secret-value-123456",
                }
            ],
        }
    ],
    "metadata": {"job": "support-triage"},
}
response = httpx.post(
    "https://api.openai.com/v1/responses",
    headers={"Authorization": f"Bearer {os.environ['OPENAI_API_KEY']}"},
    json=redact_payload(payload),
)

Use with other SDKs

Call the same helpers at each SDK boundary. Keep provider authentication in the SDK's normal environment/configuration; PromptLatch redacts prompt and payload content, not credentials held in the SDK's transport configuration.

LiteLLM:

from litellm import completion
from promptlatch import redact_params

response = completion(
    **redact_params(
        model="openai/gpt-6-sol",
        messages=[{"role": "user", "content": "GEMINI_API_KEY=example-key-value-123456"}],
    )
)

Anthropic:

from anthropic import Anthropic
from promptlatch import redact_messages

client = Anthropic()
response = client.messages.create(
    model="claude-opus-4-8",
    max_tokens=1024,
    messages=redact_messages([{"role": "user", "content": "token=example-token-value-123456"}]),
)

LangChain:

from langchain_openai import ChatOpenAI
from promptlatch import redact_messages

llm = ChatOpenAI(model="gpt-6-sol")
response = llm.invoke(
    redact_messages(
        [
            ("human", "Here is my token: example-token-value-123456"),
        ]
    )
)

Inspect redaction results

scan_text, scan_messages, scan_params, and scan_payload return a RedactionResult with .value and .stats. Use .value for the SDK call, and use .stats.redactions or .stats.rule_hits for counts by rule. Stats do not contain matched secret values.

from promptlatch import scan_text

result = scan_text("OPENAI_API_KEY=example-secret-value-123456")
print(result.stats.redactions)
safe_text = result.value

Configure custom rules

The default helpers use the default RedactionConfig. For per-application rules, create a PromptLatch instance. Exact rules of 16 characters or fewer match a secret tail inside a longer value. Prefer storing only that tail in config.

from promptlatch import PromptLatch
from promptlatch.config import RedactionConfig, RuleConfig

latch = PromptLatch(
    RedactionConfig(
        rules=[
            RuleConfig(type="exact", value="abcd1234", name="internal-token"),
        ]
    )
)
safe_prompt = latch.text("internal token: private-abcd1234")

Redaction copies mappings, lists, tuples, and Pydantic models rather than mutating caller-owned values. A detected secret in a Pydantic model's extra field name or mapping-key collision raises ValueError; fail closed and do not send the original input. Unknown private token formats need a custom exact-tail or regex rule.

Configuration

Default config: ~/.config/promptlatch/config.yaml

server:
  host: 127.0.0.1
  port: 8000
  api_key: null
  max_request_body_bytes: 134217728

target:
  default_base_url: https://api.openai.com/v1
  api_key: ${OPENAI_API_KEY}
  api_key_header: authorization
  forward_client_authorization: false
  timeout_seconds: 180
  allowed_base_urls: []
  block_private_targets: true

redaction:
  enabled: true
  engine: detect-secrets
  redact_mode: full
  encrypted: false
  max_extra_rules: 20
  max_extra_rule_chars: 1024
  allow_extra_regex_rules: false
  rules:
    - type: exact
      value: abcd1234
      name: tail-only-example
    - type: regex
      value: sk-[A-Za-z0-9_-]{20,}
      name: openai-style-token

observability:
  enabled: true
  max_database_bytes: 1073741824
  max_body_bytes: 262144
  store_request_bodies: true
  pricing_enabled: true
  dashboard_enabled: false

Store only key tails in exact rules. Full masking is default; partial masking is available through redact_mode: partial.

The default request limit is 128 MiB, including inline base64 images and conversation history. Override it with server.max_request_body_bytes or PROMPTLATCH_MAX_REQUEST_BODY_BYTES. Existing configs retain their explicit limit. This limit is separate from the 256 KiB history capture per request and the 1 GiB SQLite storage cap.

Local monitoring

Server mode records request metrics automatically. Python library calls do not create monitoring databases. Enable the optional dashboard on the same server:

promptlatch serve --dashboard --config ~/.config/promptlatch/config.yaml

Open http://127.0.0.1:8000/. The command prints the path of a private dashboard token file. Enter that file's token in the app. This token is separate from the proxy API key, is not printed in logs, and is never added to a URL. Set observability.dashboard_enabled: true to enable it through config. The dashboard shares the proxy's bind address and port. Keep the default loopback binding for local use; remote access requires a protected transport such as an HTTPS reverse proxy or SSH tunnel.

Signing in creates a browser session that expires 30 days after authentication. The session uses an HttpOnly cookie and survives page reloads, browser restarts, and server restarts. Activity does not extend its expiry. The app does not retain the file token in browser storage. Lock removes the browser session cookie. Changing the file token and restarting the server invalidates all sessions; otherwise, a copied session cookie remains valid until its expiry. Cookies belong to one hostname and browser profile. Keep using the same address; localhost and 127.0.0.1 have separate cookies. Signing in clears a stale root-path dashboard cookie that could otherwise collide with the current cookie.

When enabled, / serves the app and /_promptlatch/ is reserved for its assets and authenticated monitoring API. These requests are not proxied or included in request history. Provider routes continue on their existing paths. The monitoring namespace remains local when the dashboard is disabled, so stale browser tabs cannot forward their access token upstream. With the dashboard enabled, forwarded responses receive a browser sandbox policy that prevents upstream HTML or scripts from accessing its origin. API response bodies are unchanged.

The dashboard opens a compact request inspector with a session tree in the sidebar. Sessions show status, model, combined request time, estimated API cost, and message previews. Combined request time sums retained request durations, including overlaps. Token totals show a + when some retained requests have unknown usage. Requests and captured conversation history appear newest first. Search and status/provider/model filters apply to loaded sessions. Each selected request has a #request=<request-id> deep link. Copy link copies that address; opening it restores the request after sign-in, including older session and iteration pages. Evicted requests show an unavailable notice. Session and turn rows include the beginning of the user request. Each iteration also previews the latest user text and previous assistant answer from its captured request history. Previews are redacted, capped at 200 characters each, and evicted with request details; they are not part of long-lived usage statistics. A positive body capture limit allows previews even when the body is truncated. Setting max_body_bytes: 0 disables both. Incremental requests may omit the previous answer from their context. Expand a session to browse user turns and request iterations, then select an iteration to inspect its conversation history, raw payload, metrics, and redactions in the main pane. User-turn groups are inferred from submitted history; tool-result iterations stay under the preceding user turn. Incremental Responses requests expose retained parent links when available. Older records without turn metadata remain ungrouped iterations. Stats is global, independent of the selected request and sidebar filters. Active requests enter statistics when they finish. It shows token totals, latency, estimated cost, and redaction counts by rule. All time uses preserved lifetime totals; daily, model, and rule breakdowns cover retained aggregate history. The request tree defaults to inference calls. Select All HTTP traffic to inspect ancillary GETs and other proxy calls. Their records remain retained; deep links to those requests open the matching traffic view automatically. Conversation omits entries without readable content, including encrypted-only reasoning and tool metadata declarations. Payload preserves the captured data, with JSON formatting, syntax highlighting, line wrapping, and a copy control. The capture badge distinguishes complete, partial, and unavailable request bodies.

Counts describe matches, not unique leaked credentials. Request details contain the redacted payload, capped at 256 KiB by default, without secret samples, raw headers, or response bodies. Bodies are omitted when redaction is disabled, partial masking is selected, or content is binary or encoded.

Request details can retain allowlisted upstream error event, type, code, message, and param fields. These fields are always fully redacted, regardless of request redaction settings. message is capped at 2 KiB; event at 128 bytes; and type, code, and param at 256 bytes each, with all error fields under 4 KiB total. Capture is disabled when store_request_bodies: false or max_body_bytes: 0. Error fields are stored with request details, evicted with them, and excluded from long-lived aggregates. Full upstream response bodies and raw headers are never retained. Transport failures retain the exception type, safe reason, first failed transport phase and its elapsed milliseconds, socket cause and errno when available, request body size, and observed response byte count. Selected provider request IDs and Cloudflare ray IDs are bounded, validated, and fully redacted before storage. Raw exception text, URLs, and trace payloads are never retained or included in proxy-generated 502 responses. Recognized LLM streams that end without a completion event show stream_incomplete when no more specific provider or transport error is available. Older records cannot recover these details.

Native Responses requests preserve recognized encrypted reasoning and compaction envelopes unchanged. Typed image parts also preserve canonical PNG, JPEG, GIF, and WebP data URLs with matching binary signatures. Regular URLs remain scanned. Readable summaries, prompts, tool output text, and other fields still pass through redaction. The stored troubleshooting copy can apply additional credential masking; it is never reused as the forwarded request. An invalid_encrypted_content error identifies rejected encrypted conversation state, rather than a connection timeout. Existing corrupted history cannot be repaired from a truncated or redacted capture.

Streams close upstream connections on client disconnect, including cancellation during cleanup. Native response bytes and upstream errors remain unchanged; PromptLatch does not replay partially delivered generations or invent a successful completion event. target.timeout_seconds limits upstream read inactivity, not the total generation time.

• TTFT: elapsed time from request start to first generated output, rather than upstream response headers. Nonstreaming responses do not provide TTFT.
• TPS after first output: (output_tokens - 1) / (last_output - first_output). Stream chunks can contain multiple tokens, so this is approximate. Missing usage or fewer than two timed output events leaves TPS unknown.
• Usage: provider-reported input, output, and total tokens. Missing usage stays unknown; PromptLatch does not substitute character counts. Reasoning, cached input, and cache-write tokens appear when supplied. Response parsing recognizes JSON and SSE even when the content type is missing or incorrect. Requests show why metrics are unavailable; old captures cannot recover unrecorded usage.
• Cost: estimates from the installed LiteLLM package and its public pricing map for supported exact model matches. Unknown models, usage, or unsupported pricing remain unknown. Estimates are not invoices. The map refreshes in the background daily; no prompts, keys, model selections, or usage are sent to the pricing source. ChatGPT subscription calls use equivalent OpenAI API token rates, labelled API-price equivalent. These estimates compare API usage costs; they are not subscription charges. Non-LLM requests do not count as missing LLM usage or pricing. Retained subscription requests with complete usage can also receive estimates in background batches. Request details and aggregate costs update together; missing usage remains unknown. Historical comparisons use current standard API text-token rates, excluding extra media and hosted-tool charges. They do not reconstruct an earlier bill or service tier.

For reliable grouping, send X-PromptLatch-Thread-ID with a stable conversation identifier. PromptLatch strips this header before forwarding. Thread identifiers are HMAC-hashed locally. Metadata thread/conversation/session IDs and Responses previous_response_id links also group related requests. Without those, matching first-user-message and preceding system content provides best-effort grouping, labelled as inferred. Requests without usable history remain ungrouped.

Storage defaults to $XDG_STATE_HOME/promptlatch/requests.sqlite3, or ~/.local/state/promptlatch/requests.sqlite3 when XDG_STATE_HOME is unset. Set observability.database_path to choose another location. Both the request database and <database_path>.stats.sqlite3 share the configured physical cap, 1 GiB by default, including rollback journals. SQLite page limits reserve journal headroom, leaving roughly half the budget for database pages. Oldest request details are evicted first, then oldest daily/model aggregates if needed. Lifetime request, token, cost, latency, and redaction totals survive eviction. The dashboard shows combined storage and pruning counts. Startup checks existing files and installs page limits before writing. Legacy files without room for safe journal copies are refused without growth and require offline compaction; startup does not create a temporary VACUUM copy. A bounded background writer keeps SQLite work off the request event loop; dropped records are reported when storage cannot keep up. Graceful shutdown drains queued events. Abrupt process termination can lose events that have not reached SQLite.

One server process can write to each history database. Use separate database paths for multiple proxy workers; starting a second writer against the same path fails.

Redaction is not anonymization. Stored prompts can retain source code, personal information, and unrecognized credentials. Set observability.store_request_bodies: false for metadata-only monitoring, or observability.enabled: false to disable recording. These settings affect future requests; existing history remains on disk. Keep the state directory private.

Additional settings: queue_size (default 1024), pricing_refresh_hours (24), and pricing_file for a local LiteLLM-format JSON price map. Environment overrides use PROMPTLATCH_OBSERVABILITY_ followed by the uppercase setting name; the dashboard override is PROMPTLATCH_DASHBOARD_ENABLED.

For containers, mount a persistent writable volume and set XDG_STATE_HOME or PROMPTLATCH_OBSERVABILITY_DATABASE_PATH inside it. Enable the dashboard on that server when needed; it uses the same history and listening port.

Supported routes

PromptLatch forwards any path, with first-class tests for:

• /v1/chat/completions
• /v1/responses
• /v1/completions
• /v1/models
• /v1/messages for Claude-compatible gateways

Tests cover streaming responses, tool payloads, and vision payloads. PromptLatch redacts recursively without reshaping JSON request schemas.

Provider targets

Set default backend in config, or choose one per request:

curl http://127.0.0.1:8000/v1/responses \
  -H "X-Target-Base-URL: https://api.openai.com/v1" \
  -H "X-Target-API-Key: $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"gpt-6-sol","input":"scan this <api-key-like-value>"}'

Set X-Target-API-Key-Header: x-api-key for Anthropic-style upstream authentication.

Configured target keys are bound to target.default_base_url. A dynamic target that requires authentication must receive its key through X-Target-API-Key or X-Target-Authorization; PromptLatch never reuses configured key for another host.

An empty target.allowed_base_urls permits any public target. Add URLs to restrict dynamic routing. Set block_private_targets: false only for trusted local targets.

Per-request rules are exact matches by default. Regex rules remain available in trusted config. Set redaction.allow_extra_regex_rules: true only for authenticated clients you trust.

PromptLatch forwards routes without reshaping provider payloads.

OpenRouter per request:

target:
  allowed_base_urls:
    - https://openrouter.ai/api/v1

curl http://127.0.0.1:8000/v1/chat/completions \
  -H "X-Target-Base-URL: https://openrouter.ai/api/v1" \
  -H "X-Target-API-Key: $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"openai/gpt-oss-120b","messages":[{"role":"user","content":"scan this <api-key-like-value>"}]}'

Anthropic-compatible target:

curl http://127.0.0.1:8000/v1/messages \
  -H "X-Target-Base-URL: https://api.anthropic.com" \
  -H "X-Target-API-Key: $ANTHROPIC_API_KEY" \
  -H "X-Target-API-Key-Header: x-api-key" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{"model":"claude-opus-4-8","max_tokens":256,"messages":[{"role":"user","content":"scan this <api-key-like-value>"}]}'

Local OpenAI-compatible target:

target:
  default_base_url: http://127.0.0.1:11434/v1
  api_key: null
  allowed_base_urls:
    - http://127.0.0.1:11434/v1
  block_private_targets: false

Codex with OpenRouter

OpenRouter accepts native Responses requests, so no compatibility bridge is needed. Keep OpenRouter key in environment and send it through PromptLatch's dedicated target header. Generic client Authorization is not forwarded.

Start PromptLatch:

mkdir -p ~/.config/promptlatch
cp examples/promptlatch-openrouter.config.yaml ~/.config/promptlatch/config.yaml
export OPENROUTER_API_KEY="<openrouter-upstream-key>"
promptlatch serve

The checked-in PromptLatch config restricts dynamic routing to OpenRouter and leaves forward_client_authorization and responses_to_chat disabled.

Install Codex profile:

mkdir -p ~/.codex
cp examples/codex-openrouter-promptlatch.config.toml \
  ~/.codex/openrouter-promptlatch.config.toml

Profile contents:

model = "openai/gpt-oss-120b"
model_provider = "promptlatch-openrouter"

[model_providers.promptlatch-openrouter]
name = "PromptLatch OpenRouter"
base_url = "http://127.0.0.1:8000/v1"
wire_api = "responses"
env_http_headers = { "X-Target-API-Key" = "OPENROUTER_API_KEY" }
http_headers = { "X-Target-Base-URL" = "https://openrouter.ai/api/v1" }
request_max_retries = 0
stream_max_retries = 0

Run interactive Codex:

codex -p openrouter-promptlatch

Non-interactive smoke test:

codex exec -p openrouter-promptlatch --strict-config \
  --sandbox read-only --ephemeral --cd "$PWD" \
  "Reply with exactly: promptlatch-openrouter-ok"

Use any OpenRouter Responses-capable model by changing profile model. Current Codex requests include Responses-only custom tool descriptors, so Codex needs a backend with native Responses support. compat.responses_to_chat remains available for simpler Responses clients limited to text, messages, and standard function tools.

OpenCode

Current stable OpenCode config supports custom Chat Completions providers through @ai-sdk/openai-compatible. Copy checked-in example into project, or merge provider block into existing opencode.json:

cp examples/opencode-openrouter-promptlatch.json opencode.json
export OPENROUTER_API_KEY="<openrouter-upstream-key>"

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "promptlatch-openrouter": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "PromptLatch OpenRouter",
      "options": {
        "baseURL": "http://127.0.0.1:8000/v1",
        "headers": {
          "X-Target-Base-URL": "https://openrouter.ai/api/v1",
          "X-Target-API-Key": "{env:OPENROUTER_API_KEY}"
        }
      },
      "models": {
        "openai/gpt-oss-120b": {
          "name": "gpt-oss via PromptLatch"
        }
      }
    }
  },
  "model": "promptlatch-openrouter/openai/gpt-oss-120b"
}

Run:

opencode run -m promptlatch-openrouter/openai/gpt-oss-120b \
  --format json --dir "$PWD" \
  "Reply with exactly: promptlatch-opencode-ok"

For another Chat Completions target, replace base URL, environment variable, and model ID. Use PROMPTLATCH_TARGET_BASE_URL and PROMPTLATCH_TARGET_API_KEY instead when PromptLatch owns one fixed upstream.

Coding-agent shims

Run Codex and Claude Code through one local PromptLatch server, using their own provider credentials and the official upstream endpoints:

./install.sh
promptlatch serve --config examples/promptlatch-agents.config.yaml

# In another terminal:
codex-latch
claude-latch

Install and sign in to the original CLIs first (codex login, claude auth login). Requires Codex 0.162.1+ and Claude Code 2.1.220+. These versions support the configuration checks and provider settings used to keep requests on the proxy. The shims preserve client credentials and configuration files. The example binds to loopback, enables client-auth forwarding, restricts targets to the three official endpoints below, and enables the inspector at http://127.0.0.1:8000/. Use your existing server instead if it has equivalent authentication and native Responses settings. Do not start a second server on an occupied port.

Codex detects the saved login mode using codex login status; its output is never printed by the shim. Set PROMPTLATCH_CODEX_AUTH=chatgpt or api-key to select explicitly. API-key mode uses OPENAI_API_KEY when set, otherwise Codex's saved API-key login. Claude keeps its normal authentication precedence, including ANTHROPIC_API_KEY, ANTHROPIC_AUTH_TOKEN, and its saved login.

codex-latch exec "Review this repository"
PROMPTLATCH_CODEX_AUTH=api-key codex-latch
PROMPTLATCH_PROXY_URL=http://127.0.0.1:8001 claude-latch
codex-latch --latch-help
claude-latch --latch-help

The shims check proxy readiness and redaction before launching. They stop if the proxy is unavailable, redaction is disabled, or native client authentication cannot be forwarded. This mode requires no server.api_key or target.api_key; keep the listener local. Normal keyed-proxy configuration remains available for other clients. PROMPTLATCH_PROXY_URL accepts loopback addresses only. Use PROMPTLATCH_CODEX_BIN or PROMPTLATCH_CLAUDE_BIN to select a CLI executable explicitly.

Codex's effective provider configuration is checked before launch. Conflicting provider settings, managed routing requirements, and Codex Cloud commands stop with an error. Profiles work unless they redefine the reserved promptlatch-official provider. Claude's host-managed provider setting prevents settings files from replacing the proxy route.

Codex uses a local process and HTTP Responses streaming. Both shims disable request compression so request bodies remain scannable. Claude's nonessential traffic is disabled for the invocation. These wrappers cover model-provider requests, including conversation history and tool outputs sent back to the model. Login, MCP servers, tool subprocesses, plugins, and other network connections are separate. Files and images embedded as binary or encoded content retain the limits in SECURITY.md.

Configuration references: Codex gateway authentication, Claude Code gateways, and Claude Code networking.

Claude Code manual configuration

Claude Code sends Anthropic Messages requests. Configure provider key on PromptLatch, then use separate local bearer token for proxy authentication:

export ANTHROPIC_UPSTREAM_API_KEY="<anthropic-upstream-key>"
export PROMPTLATCH_TARGET_BASE_URL="https://api.anthropic.com"
export PROMPTLATCH_TARGET_API_KEY="$ANTHROPIC_UPSTREAM_API_KEY"
export PROMPTLATCH_TARGET_API_KEY_HEADER="x-api-key"
export PROMPTLATCH_SERVER_API_KEY="<local-proxy-key>"

promptlatch serve

export ANTHROPIC_BASE_URL="http://127.0.0.1:8000"
export ANTHROPIC_AUTH_TOKEN="$PROMPTLATCH_SERVER_API_KEY"
export DISABLE_TELEMETRY=1
export DO_NOT_TRACK=1
claude

PromptLatch validates local bearer token, removes it, then adds upstream x-api-key. It forwards /v1/messages without translating between OpenAI and Anthropic schemas.

Redaction engine

PromptLatch uses bc-detect-secrets, provider token patterns, and user-defined exact-tail or regex matches. It does not load or call a model.

Coverage includes fixture-shaped examples for:

• AI provider keys: OpenAI/Codex, Anthropic, Gemini, OpenRouter, Z.AI, MiniMax, DeepSeek, xAI/Grok, and Fireworks.
• Developer and cloud credentials: GitHub, GitLab, Atlassian, AWS, Cloudflare, Slack, Stripe, Google Cloud, Azure, npm, PyPI, and other common service tokens.
• Structured credentials: JWTs, signed URLs, URL userinfo, PEM keys, encrypted PEM keys, and PGP private keys.
• Labeled values and JSON fields such as password, token, api_key, authorization, credentials, signed_url, and sas_token.
• User-defined exact-tail and regex rules for private formats.

JSON, query parameters, and URL-encoded form fields are scanned structurally. Other unencoded text bodies are scanned without changing unrelated bytes. Multipart uploads are rejected with 415 while redaction is enabled: raw byte scanning cannot reliably protect attachment contents or form fields. Compressed request bodies are also rejected; decompress them before sending.

Every scan runs locally without an LLM. Entropy-only matching is disabled; use custom rules for opaque internal formats.

Scan results are cached per string in a bounded, in-memory record. Repeated text, such as conversation history, is scanned once per process. The record is keyed by digest and holds redacted output only, never the original input. One redactor is safe to share between threads.

Encrypt rules at rest

uv run promptlatch encrypt-rules

This creates ~/.config/promptlatch/key with mode 0600, encrypts redaction.rules with AES-GCM, writes redaction.encrypted_rules, and clears plain rules.

You can also provide key material through:

export PROMPTLATCH_CONFIG_KEY="base64-url-safe-32-byte-key"

Docker

Published image:

# ~/.config/promptlatch/provider.env, mode 0600
PROMPTLATCH_TARGET_BASE_URL=https://api.openai.com/v1
PROMPTLATCH_TARGET_API_KEY=<openai-upstream-key>

docker run -d --name promptlatch --rm \
  -p 127.0.0.1:8000:8000 \
  --env-file "$HOME/.config/promptlatch/provider.env" \
  ghcr.io/bvolpato/promptlatch:0.2.7

curl --retry 10 --retry-connrefused --retry-delay 1 \
  -fsS http://127.0.0.1:8000/healthz
docker stop promptlatch

Build current checkout:

docker build -t promptlatch:local .

Compose:

export OPENAI_API_KEY="<openai-upstream-key>"
docker compose up --build

Helm

Local chart:

kubectl create secret generic promptlatch-env \
  --from-env-file="$HOME/.config/promptlatch/kubernetes.env"

helm install promptlatch ./charts/promptlatch \
  --set env.PROMPTLATCH_TARGET_DEFAULT_BASE_URL=https://api.openai.com/v1 \
  --set existingSecret=promptlatch-env

kubectl wait deployment/promptlatch --for=condition=Available --timeout=90s
export PROMPTLATCH_SERVER_API_KEY="$(
  kubectl get secret promptlatch-env \
    -o jsonpath='{.data.PROMPTLATCH_SERVER_API_KEY}' | base64 --decode
)"
kubectl port-forward svc/promptlatch 8000:8000

In another shell:

curl -fsS http://127.0.0.1:8000/healthz
helm uninstall promptlatch

Release asset:

helm pull https://github.com/bvolpato/promptlatch/releases/download/v0.2.7/promptlatch-0.2.7.tgz
helm install promptlatch ./promptlatch-0.2.7.tgz \
  --set env.PROMPTLATCH_TARGET_DEFAULT_BASE_URL=https://api.openai.com/v1 \
  --set existingSecret=promptlatch-env

kubernetes.env must contain PROMPTLATCH_TARGET_API_KEY and PROMPTLATCH_SERVER_API_KEY; keep file outside repository with mode 0600. Without existingSecret, chart generates proxy key and stores secretEnv values in chart-managed Secret. Send Authorization: Bearer $PROMPTLATCH_SERVER_API_KEY on proxied requests. Health probes remain unauthenticated.

Services route only to pods from their own release. Connected Helm upgrades preserve the existing Deployment selector. For offline upgrades of charts through 0.2.2, set migration.preserveSelector=true and retain it on later offline upgrades. Omit it on fresh installs. Old Deployments keep their broad immutable selectors; replacing them is required for controller isolation, even though Service routing is isolated.

Emergency request tracing

promptlatch serve --debug-requests logs raw request bodies before redaction. Restrict it to local fixture data and cases where an echo target is insufficient. Auth, target-key, and redaction-rule headers are masked; body text is visible.

Uvicorn access logs are disabled in both normal and debug mode to avoid logging query-string secrets.

Development

uv sync --extra dev
uv run scripts/audit_secrets.py
uv run pytest
uv run ruff check .
uv build
uv run promptlatch scan 'OPENAI_API_KEY=<api-key-like-value>'

Fixtures are split in source so no real or contiguous fake keys are committed. Release and test commands live in CONTRIBUTING.md. Report security problems through the private path in SECURITY.md, without posting real secrets.