Aga MCP Server README
The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Aga MCP Server listing page.
AGA - Attested Governance Artifacts
Verifiable decision records for AI agents: each recorded tool-call decision is a signed, hash-chained receipt, exported in evidence bundles a reviewer can verify offline against the published format. Verification establishes the integrity of the receipts present, not that every action was recorded.
Status: published reference implementation, before independent pilot validation. The gateway emits classical Ed25519-SHA256-JCS bundles. The published
@attested-intelligence/aga-verify@2.2.2CLI checks that classical profile; on a v2/hybrid bundle it reports FAILED because it does not implement that profile. The package also exposes an ML-DSA-65 + Ed25519 composite as a library profile, andaga-proxy verifycan check it. Reference verifiers have their own unsupported-profile behavior. These are different components, not one interchangeable verifier. Build provenance concerns the published build; it is not runtime correctness or an external security audit.
Runtime status. Since 3.5.0, a measurement requested after the active artifact's TTL expires moves it to TERMINATE;
delegate_to_subagentalso refuses after expiry. Nothing checks the TTL on a schedule, and the exported bundle does not record that transition. Do not downgrade to deprecated 3.3.3 as the evaluation path. Since 3.6.0,aga-proxyhonorsAGA_GATEWAY_KEY/AGA_GATEWAY_KEY_FILE; the stdio upstream can inherit those variables. Read the known issues andTHREAT_BOUNDARY.mdbefore any runtime evaluation.
Runtime examples below identify the observed 3.6.2 package, not a newly approved production deployment. Review the known issues first; an isolated synthetic evaluation is required before considering a pilot. Prefer the static verifier path for the first check.
A Python companion SDK (aga-governance) is documented in the Python SDK section below.
Verify this yourself (don't take our word)
You do not have to take any of this on faith. The repo ships the reference verifier, the canonical vectors, and sample bundles, so you can check one offline right now, with no network and no callback to us:
The published @attested-intelligence/aga-verify CLI agrees on the tested classical corpus as the harness supplies it. npm run conformance:cross-stack (first: npm run build && npm --prefix independent-verifier run build) proves that six v1 verifier configurations, spanning three independent toolchains (JavaScript, Go, and Python, including a pure-stdlib, no-third-party-crypto path), agree on the 54 object-level cases. The five file-parsing verifiers also agree on the 7 raw-byte/file-parse cases (61 total). The in-server engine is library-only, receiving parsed objects rather than raw file bytes, so it does not run the file-parse cases; six configurations do not agree on all 61 and this no longer claims they do. npm run conformance:cross-stack-v2 proves two genuinely independent-language oracles (@noble/JS and CIRCL/Go) agree on the v2 composite corpus. For a source-and-build reproduction (build the package yourself, reproduce the published tarball byte-for-byte, re-run every gate), see the REVIEWER_GUIDE.md (a command-by-command self-service path), REPRODUCIBILITY.md, and the step-by-step SKEPTICAL_AUDITOR.md. This release carries SLSA build provenance, checkable with npm audit signatures.
What This Does
This is built for teams shipping agentic-AI products into financial services and insurance, at the moment a customer's vendor-risk, model-risk, or internal-audit review asks what your agent did and how anyone would know.
Covered tool calls routed through aga-proxy are evaluated against its configured policy. Each recorded decision (PERMITTED or DENIED) takes the form of a signed, hash-linked governance receipt; known issue 7 below describes calls refused without a receipt. aga-proxy also signs the SHA-256 of its policy's canonical JSON into every receipt; see KNOWN_LIMITATIONS.md for what that field binds. Receipts are collected into evidence bundles that anyone holding the published format and the public key can verify offline, with no callback to us.
Record. Prove. Verify.
Scope: a verified bundle proves the integrity of the receipts present: each is authentic, correctly ordered, Merkle-included, and (when a key is pinned) provenance-bound. It does not prove non-omission (that every action the agent took was logged); completeness is bounded by the tamper-evidence of the interception point, which is outside the bundle. See KNOWN_LIMITATIONS.md for the full honest boundary, and THREAT_BOUNDARY.md for the per-field detail.
Use with Claude Desktop
Add to your Claude Desktop MCP config (claude_desktop_config.json):
Claude can then seal artifacts, measure integrity, generate evidence bundles, and verify them offline through natural language.
Persist the signing key (do this first)
By default the gateway signs with an ephemeral key that rotates on every restart. That is fine for a first look, but evidence-bundle provenance cannot be pinned across restarts (and the server warns about it on stderr). Set one stable 64-hex Ed25519 seed so provenance stays pinnable:
Since 3.6.0 this applies to both binaries.
aga-proxyreads the same two variables through the same resolver and prints the active public key at startup so you can pin it out of band;--ephemeralmakes a throwaway key a stated choice. In 3.5.0 and earlieraga-proxyignored both variables silently; a key you set had no effect and no warning was printed, so evidence from such a proxy is integrity-verifiable but not provenance-pinnable across restarts. SeeDEPLOYMENT.md§2.
Provide it via AGA_GATEWAY_KEY, or AGA_GATEWAY_KEY_FILE (a path to the seed). In Claude Desktop, add an env block:
Keep the seed secret and out of version control; see DEPLOYMENT.md for key handling. A seed in an agent client's environment is not a separate trust domain, and the stdio upstream can inherit the key-related variables (known issue 3). Same-key restarts do not preserve the in-memory ledger: export and verify before stopping.
MCP Tools (15)
| Category | Tools |
|---|---|
| Identity | get_server_info, get_portal_state |
| Lifecycle | init_chain, attest_subject, revoke_artifact |
| Measurement & decision | measure_integrity, measure_behavior, verify_chain |
| Evidence | generate_evidence_bundle, verify_bundle_offline |
| Privacy | request_claim, list_claims |
| Delegation | delegate_to_subagent |
| Audit | get_receipts, get_chain_events |
measure_behavioris detective-only by default: it observes tool-usage patterns and records a signed, provable drift finding, but does not block. Enforcement (drift → quarantine) is opt-in viaenforce=trueand off by default. Hard governance decisions (PERMITTED/DENIED) are made by the portal/PEP, not the behavioral monitor.
Quick Start: verify a bundle offline
A bundle this package emits (via the generate_evidence_bundle MCP tool) is a canonical SEP bundle. Verify it offline, with no network and no callback to us:
The published @attested-intelligence/aga-verify CLI is the shipped path (the older forgeable 1.0.0 is deprecated); the reference verify-sep.mjs provides another implementation from a repo clone; verdict agreement is scoped to tested cases, not all input bytes. Without --pubkey you get an integrity-only result (issuerVerified=false); supply a nonempty expected key from a separate trusted channel to authenticate that signing key; the mapping to an organization depends on that channel. A trailing --pubkey without a value falls back to integrity-only success in 2.2.2. See THREAT_BOUNDARY.md §3.7. A hosted browser verifier is linked under Links.
The reference §6 algorithm is implemented in three languages: JavaScript (aga-receipt-spec/verify/verify-sep.mjs), Go (verify.go, stdlib crypto/ed25519), and Python (verify.py, pure-stdlib RFC-8032 Ed25519). A cross-stack harness (npm run conformance:cross-stack; first: npm run build && npm --prefix independent-verifier run build) proves all three, plus the in-server engine and aga-verify, agree on the published canonical cases as the harness feeds them (object cases are re-serialized; raw-byte cases are separate). Outside that corpus, the implementations differ, including some parser and pin semantics. The v2 composite profile (ML-DSA-65+Ed25519-SHA256-JCS) is held to the same bar by a second harness (npm run conformance:cross-stack-v2): a @noble/JavaScript engine and a CIRCL/Go oracle, two genuinely independent toolchains, render identical verdicts on the pinned v2 corpus, and the reference v1 verifier (verify-sep.mjs/verify.py/verify.go) returns UNSUPPORTED_PROFILE (exit 3) on a v2 bundle, signalling "profile not implemented" rather than a misleading "invalid". (The published aga-verify CLI does not implement this profile trichotomy: on a v2 bundle it returns FAILED (exit 1). Use exit 3 as the unsupported-profile signal only with the reference verifiers.)
Check-name mapping across implementations
The JS reference verifier and the Python SDK (aga-governance) decompose the same seven-check verification differently. Overall verdicts and exit codes agree on all 61 conformance-corpus cases as the cross-stack harness feeds them (object-level cases re-serialized, so float spellings arrive as integers; measured on aga-governance 0.3.2 on 2026-09-25) and on the 10 cells re-proven 2026-07-01 (pristine and tampered bundles with unpinned, correct and wrong keys). On the literal file bytes of the corpus's float-spelled leaf_index case (0.0), aga-governance 0.3.2 reports FAILED where the JS reference, aga-verify, Go and Python reference verifiers report VERIFIED; see https://attestedintelligence.com/spec. The sub-check that reports a given tamper can differ:
| JS reference check | Python result field | What it covers |
|---|---|---|
structural | algorithm_valid + parts of bundle_consistent | algorithm id, key well-formedness, receipt/proof counts |
receipt_signatures | receipt_signatures_valid | Ed25519 over canonical receipt bytes |
chain_and_ordering | chain_integrity_valid | prev-leaf linkage, canonical non-decreasing timestamps (ids are not ordering fields and are not checked) |
merkle_and_bijection | merkle_proofs_valid | leaf recompute, single-root walk, index bijection |
signed_checkpoint | checkpoint_valid | gateway-signed root + count + chain-head binding |
envelope_consistency | envelope_consistent | envelope gateway_id, generated_at, merkle_root vs signed content (bundle_id, schema_version, the envelope policy_reference and offline_capable are unsigned and unchecked) |
gateway_key_match (with --pubkey) | gateway_key_match / provenance | pinned issuer key |
Known decomposition difference: the JS reference recomputes every Merkle leaf from full receipt content, so a receipt-signature tamper also fails merkle_and_bijection; the Python verifier surfaces the same tamper in receipt_signatures_valid, chain_integrity_valid, and bundle_consistent while its merkle_proofs_valid can remain true. Neither is looser: the bundle fails in both stacks, exit 1. A --pubkey KEY that is not 64 lowercase hex characters is a usage error (exit 2) in the JS reference, aga-verify and the Python SDK, and a 64-hex pin that is not a valid curve point is honored, fails to match, and fails the bundle (exit 1). Written as --pubkey=KEY, the pin is ignored by the JS reference, aga-verify, verify.go and v2/verify-v2.go, and by verify.py when it follows the bundle path (integrity only, exit 0), and read by the Python SDK. Other verifiers differ as well. The in-server engine (the package's ./verify export, which verify_bundle_offline calls) treats a pin that is not a well-formed key for the bundle's profile (for a v1 bundle, a small-order point or a non-canonical encoding) as no pin, and returns VERIFIED with pinned: false. v2/verify-v2.go does the same and prints integrity only; no key pinned (exit 0). A 64-hex value that is not a curve point counts as well-formed, so both take it as a pin and the bundle fails. The Go and Python reference verifiers in aga-receipt-spec/verify/ treat a pin that is not 64 lowercase hex the same way and print integrity only; no key pinned (exit 0). Read pinned before taking a VERIFIED as provenance; in CI, pass the key after a space and check that the output says provenance verified. A --pubkey given with no value is also treated as no pin (exit 0, integrity only) by aga-verify, verify-sep.mjs, verify.py, verify.go and v2/verify-v2.go, and is a usage error in the Python SDK. Other differences concern the bundle rather than the pin, and https://attestedintelligence.com/security lists the ones measured, including which algorithm labels each verifier leaves unchecked; outside the conformance corpus the verifiers differ in both directions. Two examples, where the failing side fails closed: a proof leaf_index spelled as an integral float (1.0) reports FAILED in aga-governance 0.3.2 and VERIFIED in the others, and object keys outside the Basic Multilingual Plane (possible only in a non-string field value, which no shipped producer emits) sort differently in verify.py, verify.go, v2/verify-v2.go and aga-governance than in the JavaScript verifiers, so such a bundle reports VERIFIED in JavaScript and FAILED in Go and Python. These wait for the next reviewed release.
How It Works
MCP Governance Proxy
Run AGA as a proxy in front of an MCP server that it starts as a stdio child process (the default stdio transport), or one it reaches with a plain JSON-RPC POST (--upstream-url; no Streamable HTTP session or SSE handling). The proxy's agent port speaks newline-delimited JSON-RPC 2.0 over raw TCP, not stdio or Streamable HTTP. A stdio MCP client needs a relay you provide (a few lines that pipe stdin to the port and the port to stdout); none ships. A scripted client can speak that framing directly. Every tools/call request with a non-empty string tool name and arguments the proxy can canonicalize is evaluated against the policy and produces a signed receipt, except the calls that known issue 7 below describes as refused without one. Other methods that are not benign are forwarded with a signed passthrough receipt and are not policy-evaluated, and benign protocol methods (initialize, initialized, ping, tools/list, prompts/list, resources/list, resources/templates/list, logging/setLevel, completion/complete and notifications/*) produce no receipt (THREAT_BOUNDARY.md section 3 item 2). Read the known issues below before you expose the port.
permissive records each tools/call it evaluates (known issue 7 describes the exceptions) and denies nothing on policy
grounds. standard and restrictive allow only generic example tool names, so they deny every tool this example server
exposes; to permit some of your server's tools and deny the rest, pass a --policy file that names them.
Exporting the evidence bundle from a running proxy
The proxy records receipts in its own process and keeps the SEP ledger in memory. To make that live ledger reachable from a separate shell, aga-proxy start opens a loopback-only control channel: an HTTP listener bound to 127.0.0.1 (never a routable interface), on its own port (default 18801, override with --control-port), distinct from the agent-facing proxy port (18800). It exposes only read routes (/export, /status, /receipts); nothing on it mutates policy or state. It does not check a request's Host or Origin header, so a web page in a browser on the same host can read its responses through DNS rebinding unless the browser blocks it (known issue 12). The proxy writes the chosen control port to ~/.aga-proxy/control.json alongside proxy.pid.
A separate aga-proxy export invocation reads that file and fetches the same signed bundle the running proxy would emit:
Export and verify before you stop the proxy: aga-proxy stop ends the process without exporting, and the in-memory chain
goes with it (known issue 9 covers export time and bounding the chain).
If no proxy is running, aga-proxy export prints no running proxy found; start it first, or export from within the session and exits non-zero; it never emits an empty or placeholder bundle. Within the MCP server session you can also call the generate_evidence_bundle tool and save the returned JSON.
In-memory ledger: the exported bundle is the durable cryptographic record, but the live in-process chain does not survive a proxy restart. This flow makes the live ledger reachable from another process; it does not add cross-restart persistence, which needs the persistent (SQLite) backend and remains roadmap (see KNOWN_LIMITATIONS.md).
The proxy intercepts tools/call requests, evaluates them against the loaded policy (a JSON file or a built-in profile; the SHA-256 of its canonical JSON is signed into every receipt), and generates a signed SEP receipt for every decision (except the calls that known issue 7 below describes as refused without one). Permitted calls are forwarded to the downstream server; denied calls return an MCP error and never reach it. Every decision is hash-linked and checkpoint-bound into a tamper-evident bundle. (Methods other than tools/call aren't policy-evaluated, but non-benign ones are recorded as signed passthrough receipts for auditability, and a library caller can pass a method denylist (denyMethods) to reject them; the aga-proxy CLI has no flag for it; see THREAT_BOUNDARY.md §3.2.)
Three built-in policy profiles:
- permissive -
audit_only: denies nothing on policy grounds and records eachtools/callit evaluates (default); the fail-closed refusals below and known issue 7 still apply - standard - an allowlist of ten generic example tool names (
filesystem_read,shell_execute,web_searchand others) with rate limits, and substring denials on two of them; every other tool is denied, so a real server's tools need a--policyfile - restrictive - an allowlist of three generic example tool names with lower rate limits and a path prefix on one; every other tool is denied
Because the default (permissive) is audit-only, starting with an audit_only policy prints a loud stderr banner stating that every call is permitted and recorded and no call is denied in that mode. No call is denied on policy grounds, but the proxy still refuses, fail-closed, a tools/call with no tool name or with arguments it cannot canonicalize (nested past 100 levels, for example), and signs a DENIED receipt for each; a name of 0, false, null or an empty string counts as no name. Policy denial needs --profile standard or restrictive, or a --policy file in allowlist or denylist mode. In denylist mode a policy denies each tool it lists as an object whose allowed value is missing or falsy (false, 0, null or an empty string); any other allowed value allows the tool, including the string "false", and so does a listed entry that is itself false, 0, null or an empty string rather than an object. In 3.6.0 through 3.6.2 it also denies an unlisted tool named like a built-in object property, such as constructor. It applies the rate limits of the listed tools it allows; path and pattern rules apply in allowlist mode only and check only top-level string arguments (known issue 10). A --policy file in audit_only mode permits every call; one with any other mode, or none, denies every tools/call, and in 3.6.0 through 3.6.2 one in allowlist or denylist mode whose constraints member is missing or null refuses, with no receipt and no response, every tools/call that has a tool name and arguments the proxy can canonicalize (known issue 7). An unrecognized --profile value is a hard error (exit 2 listing the valid names), never a silent fallback to permissive.
Verification (canonical SEP 3.0; normative §6 algorithm in aga-receipt-spec/verify/verify-sep.mjs)
- Structural floor - Bundle declares Ed25519-SHA256-JCS, public key well-formed (all small-order encodings + non-canonical
y ≥ prejected),receipts.length > 0, proof count = receipt count - Receipt Signatures - Ed25519 over JCS-profile canonical JSON, sorted-key (signature field excluded)
- Chain + ordering - Each receipt's
previous_receipt_hash= leaf of the preceding receipt; non-decreasing timestamps - Merkle Proofs - Recompute every leaf from receipt content, walk siblings/directions to one root, leaf indices form the complete
0..N-1bijection - Signed checkpoint - Verify the gateway-signed checkpoint binding
merkle_root,leaf_count, and chain head (this makes the no-prefix construction truncation-safe) - Provenance (when a key is pinned) -
public_key == expected key; otherwise integrity-only is reported
Cryptographic Primitives
| Primitive | Purpose |
|---|---|
| Ed25519 | Receipt signatures |
| SHA-256 | Hash chaining, Merkle trees, leaf computation |
| JCS-profile (sorted-key canonical JSON) | Deterministic signing (canon is byte-compatible with the reference verifier) |
| Merkle Trees | Binding all receipts to a single verifiable root |
Live Gateway
A demo gateway is deployed on Cloudflare Workers (a separate deployment that may track its own version; treat it as a convenience mirror, and always verify what it returns offline against a pinned key, not as the canonical artifact):
Python SDK
Status, checked against PyPI on 2026-09-28. The current release is
aga-governance0.3.2. Version 0.3.1 fixed the depth-bomb crash; 0.3.0 was yanked for that crash. Both 0.2.6 files are now yanked as well, but that does not repair installed copies. Use the reviewed current version when evaluating untrusted bundles and retain its documented parser and verdict limitations. The JavaScript reference verifier andaga-verifydo not have that Python depth-bomb crash.
Test Suite
Automated tests across TypeScript and Python, plus a conformance corpus:
- TypeScript MCP server: 428 automated tests (vitest), including provable-denial and behavioral-monitor regressions
- SEP conformance corpus:
npm run test:conformance(valid → VERIFIED, negatives → FAILED) - Python companion SDK: the separately-published
aga-governancePyPI package (install + smoke-checked here; its full pytest suite runs from the source tree). The smoke check imports the package and prints its version. It does not exercise the verifier.
Benchmarks
Receipt-format determinism is reproducible here: npm test runs the cross-language vectors, and npm run conformance:cross-stack (first: npm run build && npm --prefix independent-verifier run build) shows the six v1 verifier configurations (across three independent toolchains: JS, Go, Python) agree on the 54 object-level cases of the canonical 61-case corpus. The remaining 7 are raw-byte/file-parse cases run by the five file-parsing verifiers, since the in-server engine never receives raw bytes. npm run conformance:cross-stack-v2 shows the two independent-language v2 oracles agree on the composite corpus.
Project Structure
Links
- Website
- Technology
- Live Verifier
- Trust and Scope
- Diligence Materials
- MCP Server (npm)
- Python SDK (PyPI)
- Changelog
- Threat boundary
- Deployment guide
Known issues in 3.6.0 to 3.6.2 and the published verifiers
3.6.1 and 3.6.2 change only the documentation and the version number; the runtime is 3.6.0's. Items 1 to 4 were reproduced on 2026-09-23 on @attested-intelligence/aga-mcp-server 3.6.0 installed from npm, and
concern the aga-proxy gateway. Item 5, added 2026-09-25, concerns the verifiers and was reproduced on 2026-09-25
on the current releases. Item 6, also added 2026-09-25, concerns aga-proxy with an HTTP upstream and was
reproduced on 2026-09-25 on 3.6.2. Item 7, also added 2026-09-25, concerns tools/call messages that aga-proxy refuses
without a receipt and was reproduced on 2026-09-25 on 3.6.2 (its oversized-message case was added and reproduced on
2026-09-26). Items 8 to 12, added 2026-09-26, concern non-ASCII text,
the cost of exporting evidence, what policy constraints check, oversized tool results and the control channel; each
was reproduced on 2026-09-26 on 3.6.2, as were the memory and Windows port cases added to item 1 that day. The same
list is kept at https://attestedintelligence.com/security.
- The agent port listens on every network interface, with no authentication. Anyone who can reach the host on that port can send governed calls through the proxy. The proxy also sets no limit on the number of connections, so although each connection's unfinished message is capped (item 7), the memory they hold together is not: on 3.6.2, ten connections that each sent 7.5 MiB without ending a message raised the proxy's working set from 68 MiB to 392 MiB while they stayed open, and a governed call on another connection was still answered. On Windows, a process running under the same user account as the proxy can bind 127.0.0.1 on the same port while the proxy listens, and a local client that connects to 127.0.0.1 then reaches that process instead of the proxy, with no policy check and no receipt (measured on 3.6.2 with both processes under one account). Block inbound traffic to the port in the host firewall, or admit only the agent with network policy; on Windows, run nothing untrusted under the proxy's account, and check while the proxy runs that its process is the only listener on the port. The control port is bound to loopback; see item 12.
- Two clients that reuse a JSON-RPC id through one proxy can receive each other's tool results. Workaround, measured on 3.6.0: give each client its own id range, or run one proxy per client. With disjoint ids, every result reached the client that asked for it.
- When the gateway key is supplied through
AGA_GATEWAY_KEY(with or without--ephemeral) orAGA_GATEWAY_KEY_FILE, the stdio upstream inherits that variable (the seed, or the file's path), so the upstream sits inside the key's trust domain. Workaround, measured on 3.6.0: run without either variable. The upstream then sees neither, but the proxy signs with a per-process key that cannot be pinned across restarts. - The
--upstream-urlmode forwards raw JSON-RPC over HTTP POST with only a content-type header. It does not implement the MCP Streamable HTTP transport (the Accept header, the request metadata headers and event-stream handling), so a spec-conformant HTTP MCP server rejects its requests. Workaround: bridge to the server over stdio. - A bundle file can repeat a field name anywhere: in a receipt, in the checkpoint or in the envelope. For example,
a forged
"decision": "PERMITTED"can be placed ahead of the signed"decision": "DENIED", or a forged checkpointleaf_countahead of the signed one. The published verifiers (aga-verify 2.2.2, aga-governance 0.3.2, the verifier in this package, and the reference verifiers inaga-receipt-spec/verify/) and the site's /verify page read the last occurrence, and when it holds the genuine value they report VERIFIED, with provenance when the key is pinned. They do not reject the file, so a tool or a person reading the first occurrence can see a value that was never signed. Measured on the public sample bundle, pinned to the sample key: with the repeated name inserted in a receipt, in the checkpoint or in the envelope, aga-verify 2.2.2, aga-governance 0.3.2, the verifier in this package and /verify report VERIFIED, and a real change of the checkpoint value fails. Workaround: treat the verifier's parsed output as the record's content, and reject or flag files with repeated field names before displaying them. A strict rejection of repeated field names is planned for the reviewed release. - A message can repeat the
"method"member when aga-proxy has an HTTP upstream (--upstream-url). Withtools/callfirst and another method last, aga-proxy reads the last copy, so it never checks the tool call against the policy, and it forwards every method other thantools/callto the HTTP upstream as the exact bytes it received. An upstream whose JSON parser keeps the first copy of a repeated name then runs the tool call, even one the policy denies. The bundle holds no receipt for that call: nothing at all when the last method is one the proxy passes through without a receipt (such asping,initialize, a list method or a notification), and otherwise only a passthrough receipt that names the last method. An upstream that keeps the last copy handles the message as the method the proxy read. The stdio upstream, the default, is not affected: the proxy re-serializes each message before writing it, so the upstream receives one method. Measured on 3.6.2 from npm on 2026-09-25, with the restrictive profile and with the default permissive profile. Workaround: keep the stdio default, or have the HTTP upstream reject any message that repeats a member name. A strict rejection of repeated member names in the proxy is planned for the reviewed release. - aga-proxy does not record every
tools/callit refuses. It signs atools/call's receipt before it forwards the call, so with a stdio upstream a call it cannot record never reaches the tool (for an HTTP upstream, see item 6), but such a call leaves no receipt. These cases were measured on 3.6.2 from npm on 2026-09-25, each with no receipt and no response to the client: atools/callwhose tool name is a non-zero number, an array or an object (a name of0,false,nullor an empty string counts as no name and gets a DENIED receipt and an error); one whose tool name or string id holds an unpaired surrogate (the JSON escape\ud800, for example); under a--policyfile inallowlistordenylistmode whoseconstraintsmember is missing or null, everytools/callwith a tool name and arguments the proxy can canonicalize; and, under an allowlist file, a call it would otherwise permit that carries a string path when that tool'spath_prefixis neither a string nor false, 0 or null. The proxy starts with such a policy file, and it reports each refusal listed above only on its own stderr. A message sent as a JSON-RPC batch array or without"jsonrpc": "2.0"is refused differently: the client gets an error, and there is no receipt. A message of 8,388,608 characters or more (UTF-16 code units, about 8.4 million), not counting the newline that ends it but counting a carriage return before that newline, also gets an error and no receipt, and the proxy then closes the connection, dropping any reply still due on it. The limit counts input not yet split into messages, so a message just under the limit can be refused the same way when the read that completes it also carries enough input after it to pass the limit; whether that happens depends on where the reads fall. On 3.6.2, when a 100-character message, a message 58 characters under the limit and a 100,000-character message were sent in one write, the first was answered, the second got the error and the connection closed; sent without the 100,000-character message, or without the 100-character one, every message was forwarded. These are the cases measured, not a proof that no other input does the same. Workaround: give every policy file aconstraintsobject whosepath_prefixvalues are strings, and have the client time out a call that gets no reply. A DENIED receipt and an error for a malformed tool name, and a check of the policy file at startup, are planned for the reviewed release. - aga-proxy can alter non-ASCII text whose bytes are split between two reads. It decodes each chunk it reads from
the agent's connection, and from a stdio upstream's output, on its own, so a character split between two chunks
becomes one or more replacement characters (U+FFFD). A tool call's arguments can then reach the upstream altered, and
the receipt's arguments hash is the hash of the altered arguments; a large non-ASCII result from a stdio upstream can
reach the agent altered. An HTTP upstream's result is decoded whole. Whether a split happens depends on how the bytes
arrive, so any message with non-ASCII text can be affected, and large ones more often. Measured on 3.6.2 from npm on
2026-09-26: a forced split inside "é" reached the upstream as two replacement characters, and a result of 200,000 "€"
reached the client with 15 replacement characters in it. Workaround: send JSON whose non-ASCII characters are written
as
\uXXXXescapes, so every byte the proxy reads from the agent is ASCII, and have a stdio upstream do the same; a forced split of the escaped message then arrived intact. A fix is planned for the reviewed release. - Exporting the evidence bundle takes time that grows with the square of the number of receipts, and aga-proxy
handles nothing else while it runs: every governed call waits until the export ends. A call forwarded to a stdio
upstream that has not answered when an export starts gets a timeout error if the export ends more than 30 seconds
after the call was forwarded, although the upstream ran it and its receipt says PERMITTED, so an agent that retries
can run the tool twice. Measured on 3.6.2 from npm on
2026-09-26 through the control channel's
GET /export: 2.8 seconds at 1,000 receipts and 17.4 seconds at 2,500, with atools/callsent during the export waiting as long; at 4,000 receipts an export took 43.8 seconds, and a call forwarded just before it, which the upstream answered in 2 seconds, got the timeout error. At 1,000 to 4,000 receipts a compact bundle took about 1.7 to 1.9 KB per receipt, rising with the count. An audit the same day measured 88 to 113 seconds at 5,000 receipts. Verification time grows close to linearly. Workaround: each export covers every receipt since the proxy started and does not shorten the chain, so bound the chain by restarting the proxy on a schedule: pause the agents, export and verify, then restart. A restart begins a new chain that is not linked to the last one and resets the rate-limit counts; with a per-process key (--ephemeral, or neitherAGA_GATEWAY_KEYnorAGA_GATEWAY_KEY_FILEset) it also begins a new signing key, printed at startup, that cannot be pinned across restarts (item 3). Export outside busy periods, and export and verify before any stop, because the live chain is kept in memory and a stop loses receipts not yet exported. A fix that leaves the bundle's bytes unchanged is planned for the reviewed release. - Policy constraints check less than their names suggest. A
path_prefixis checked only when the value under the checked key (path, or the keys a rule lists inpath_keys) is a string, so the same path sent inside an array or an object is not checked.denied_patternsmatch case-sensitively and only in top-level string arguments, so an uppercased command, or one inside an array or a nested object, is not matched. A constraint key the proxy does not recognise, such as a misspelling, is ignored without a warning, and a value of the wrong JSON type is not rejected:allowed: "false", a string, allows the tool; in denylist mode a tool listed asfalse,nullor0instead of an object is allowed; and a non-emptypath_keysstring instead of an array makes the check read each character as a key name, so the intended key goes unchecked. Rate limits count per tool name across the whole proxy, shared by every client, and in allowlist mode the limit is checked before the path and pattern rules, so a call those rules deny still uses up a slot. Measured on 3.6.2 from npm on 2026-09-26 with allowlist and denylist policy files: apath_prefixof/homedenied"/etc/passwd"and forwarded["/etc/passwd"]; a denied pattern ofrm -rfdeniedrm -rf /and forwardedRM -RF /and the same command inside an array; a rule spelleddenied_patterndenied nothing;allowed: "false"forwarded the call in both modes; a denylist entry offalseforwarded the call;path_keys: "path"forwarded/etc/passwdpast a/homeprefix; and, under a limit of 2 a minute, two calls denied by a/homeprefix left a third call, to a path under/home, denied for the rate limit, while after one such denial that call was forwarded. Workaround: treat path and pattern rules as a convenience rather than a boundary, restrict paths in the upstream server itself, and check a policy file's keys against the constraint names indist/proxy/types.d.ts. Checks that fail closed on these inputs, and a check of the policy file at startup, are planned for the reviewed release. - A stdio upstream's response whose JSON line is 8,388,608 characters or more (UTF-16 code units, counting JSON escaping and a carriage return before the newline, but not the newline itself) is dropped. The call already has a PERMITTED receipt, and the agent gets a timeout error after 30 seconds, so an agent that retries can run the tool twice; the proxy reports the drop only on its own stderr. The limit counts upstream output not yet split into lines, so a response just under it can be dropped when the read that completes it also carries enough output after it to pass the limit; every response after it that begins in that read, possibly to other clients' calls, is dropped with it. Whether that happens depends on where the reads fall. Measured on 3.6.2 from npm on 2026-09-26: a 9,000,000-character result was dropped and the agent got the timeout after 30.0 seconds, while a 1,000,000-character result came back in 31 milliseconds. Through a running proxy, a response line of 8,388,607 characters was returned and one of 8,388,608 was dropped; and when the upstream answered three calls in one write (100 characters and 58 under the limit for one client, then 100 for the other), the first was answered and the other two, one of them the other client's, timed out, while the same answers without the leading 100-character answer were both returned. These are the cases measured. An HTTP upstream's result is read whole and is not dropped this way; on Linux, one at the same bound closes the agent's connection instead (item 13). Workaround: keep tool results well under the bound, for example by reading large files in parts, and where results are large run one proxy per client. An error returned at once is planned for the reviewed release.
- The control channel does not check a request's Host or Origin header. It listens on 127.0.0.1 (port 18801 by
default) so that a separate
aga-proxy exportcan fetch the live bundle (routes/export,/statusand/receipts). Measured on 3.6.2 from npm on 2026-09-26:GET /receiptsandGET /exportsent with the Host and Origin of another site returned 200, and both carried a denied call's argument path in its denial reason. A web page open in a browser on the same host can therefore read the live receipts and the evidence bundle through DNS rebinding, unless the browser blocks a public site's requests to the loopback address; any local user on the host can read them as well. A page can also start an export with a plainGET /exportwithout rebinding, unless the browser blocks it, and each export holds up governed calls while it runs (item 9). The CLI has no option meant to turn the channel off; a--control-portof 70000, or any number above 65535, leaves it unstarted while governance runs, but then no command can export the running proxy's receipts. Workaround: do not browse the web on the host while the proxy runs, or run the proxy on a host where no one does, and on a host shared with other users treat the live receipts as readable by all of them. A Host and Origin check is planned for the reviewed release. - On Linux, an HTTP upstream's (
--upstream-url) response whose JSON line, as the proxy serializes it, is 8,388,608 characters or more (UTF-16 code units, counting JSON escaping but not the newline) closes the agent's connection. The proxy writes the whole line to the agent's socket in one call. A Linux kernel with its default socket buffers takes only part of a write that size; the rest goes out as the client reads, but the whole line counts as waiting until it has all gone, so the proxy's guard against a client that does not read its responses sees more than 8,388,608 characters waiting, the newline included, and destroys the socket. The call already has a PERMITTED receipt and the upstream has run the tool; the agent's connection closes part-way through the reply with no error message, so an agent that reconnects and retries can run the tool twice; every other call in flight on that connection is lost with it; and the proxy reports nothing. Measured on 3.6.2 from npm on 2026-09-26 on a Linux 6.18 host with Node 24 and the default socket buffers (net.ipv4.tcp_wmem4096 16384 4194304): result lines of 4,000,000, 8,000,000, 8,388,606 and 8,388,607 characters were returned, and lines of 8,388,608, 8,388,609, 9,000,000, 10,000,000, 12,000,000, 16,000,000, 20,000,000 and 32,000,000 characters closed the connection after 3.1 to 5.0 MB of the reply had arrived; a client that read nothing for 4 seconds got a 4,000,000-character result and lost a 9,000,000-character one after 2.7 MB. On Windows 11 the same proxy returned a 20,000,000-character result, and a 9,000,000-character one to a client that read nothing for 4 seconds, because that kernel took each write whole. A stdio upstream's response of this size is dropped instead (item 11). Workaround: keep tool results well under the bound, for example by reading large files in parts. An error returned at once is planned for the reviewed release.
No fixed version is named until one is published.
Security
See SECURITY.md for vulnerability reporting.
Contributing
See CONTRIBUTING.md for development setup and guidelines.
License
MIT. The aga-receipt-spec/ directory carries its own
Apache-2.0 license (see aga-receipt-spec/LICENSE).
Attested Intelligence Holdings LLC