About AXIOM
AXIOM is a transactional write gate for coding agents, available as an MCP server and CLI. It controls repository changes through plans, manifest checks, hash-confirmed two-phase application, and a journal.
Key features
- 01Compile plans into canonical manifests
- 02Run set-level checks before writes
- 03Hash-gated two-phase apply
- 04Journal changes and rollbacks
- 05stdio and Streamable HTTP transports
Use cases
- 01Review multi-file changes before applying
- 02Check repository changes for conflicts
- 03Track or roll back applied changes
MCP support
AXIOM runs as an MCP server over stdio or Streamable HTTP, exposing tools for planning, checks, writes, apply, status, and rollback. The server can expose the core or full toolset.
Configuration
Local server config is the same for every client. Copy it into your client’s MCP config file.
npx -y @codai/axiom-mcpAXIOM_SIGNING_KEYSecretOptional Ed25519 private key (base64 PKCS#8 or raw seed) used by axiom_sign; never stored under a root.Local servers need a runtime on your machine (e.g. Node.js 18+). Config file locations differ by client; see your client’s docs.
Tools
18Saved tool list · updated 2w ago
| Parameter | Type | Required | Description |
|---|---|---|---|
| plan | object | Yes | Plan document (apiVersion axiom.dev/v2, kind Plan) |
Details
From the project README@codai/axiom-mcp
MCP server (stdio or Streamable HTTP) and CLI for AXIOM v2 — the transactional write gate for coding agents:
Plan → canonical ManifestBundle → set-level checks → hash-gated two-phase apply → journal.
dist/cli.js (thin entry) + dist/cli-main.js (lazy-loaded engines; the MCP SDK v2 and zod are
bundled into the mcp-lazy / http-lazy chunks); no runtime dependencies.
Install & run
npx @codai/axiom-mcp mcp --root /abs/path/to/repo # stdio MCP server
npx @codai/axiom-mcp mcp --root /abs/path/to/repo --http 127.0.0.1:3411 # Streamable HTTP at /mcp
npx @codai/axiom-mcp --help # CLI verbs
--root may repeat. Every tool root argument must equal or lie inside one of them; with exactly one
root it is the default. There is no env-var or cwd fallback (ERR_ROOT_REQUIRED / ERR_ROOT_NOT_ALLOWED).
Protocol revisions — --wire 2026|2025|2026-only
Built on MCP TypeScript SDK v2 (@modelcontextprotocol/server 2.0.0). The server speaks the
2026-07-28 revision (no initialize, per-request _meta envelope, server/discover,
ttlMs/cacheScope on list results) and the 2025-era revisions (initialize handshake,
HTTP sessions) from the same entry — the SDK pins each stdio connection, or routes each HTTP
request, to the era the client opened with. --wire 2026 (default) and --wire 2025 both serve
both; --wire 2026-only refuses 2025 openings with the unsupported-protocol-version error. Clients
on SDK v1 need no change. Details: docs/reference/mcp-tools.md.
VS Code — .vscode/mcp.json
{
"servers": {
"axiom": {
"type": "stdio",
"command": "npx",
"args": ["-y", "@codai/axiom-mcp", "mcp", "--root", "${workspaceFolder}"]
}
}
}
Streamable HTTP — --http <host:port>
axiom mcp --root /abs/repo --http 127.0.0.1:3411 # loopback, no token needed
axiom mcp --root /abs/repo --http 0 # random port; URL logged at info level
AXIOM_HTTP_TOKEN=$(openssl rand -hex 32) axiom mcp --root /abs/repo --http 0.0.0.0:3411 --log-level info
- Endpoints:
POST /mcp(2026-07-28 requests are served statelessly; a 2025initializeopens a session and returnsMcp-Session-Id, which every later 2025 request must send),GET /mcp(standalone SSE stream, one per 2025 session),DELETE /mcp(close the session),GET /health→{ ok, name, version }(unauthenticated). Anything else is404JSON. - Loopback by default. A non-loopback host refuses to start unless a bearer token is present in
the env var named by
--http-token-env <NAME>(defaultAXIOM_HTTP_TOKEN, ≥ 16 chars). Clients sendAuthorization: Bearer <token>; the compare is constant-time. A token is optional on loopback. - DNS-rebinding protection is on for loopback binds (
Hostmust be<host>:<port>,localhost:<port>or127.0.0.1:<port>; otherwise403). Request bodies over 4 MiB are413. - 2025 sessions idle for 30 minutes are evicted; each session (and each 2026 request) has its own
server instance from one factory (roots, discovered sub-roots and guard settings are shared). The
transport lives in
dist/http-lazy.js, loaded only with--http, and is plainnode:http— no express/hono at runtime. - VS Code:
{ "type": "http", "url": "http://127.0.0.1:3411/mcp" }; add"headers": { "Authorization": "Bearer ${input:axiom-token}" }when a token is set.
Conformance: packages/conformance runs @modelcontextprotocol/conformance server against this transport
in CI with an expected-failures baseline (packages/conformance/baseline.yml).
Claude Desktop — claude_desktop_config.json
{
"mcpServers": {
"axiom": { "command": "npx", "args": ["-y", "@codai/axiom-mcp", "mcp", "--root", "/abs/path/to/repo"] }
}
}
Tools
Toolsets (3.0, D-43)
axiom mcp serves the core toolset by default — the six tools an agent needs to write
through the gate: axiom_write, axiom_status, axiom_rollback, axiom_check,
axiom_apply, axiom_apply_dry_run. Start the server with --tools full to also expose the
plan / manifest / task plumbing below (every row not marked core). spec/tools.json and
spec/codai-tools.json list every tool with "toolset": "core" | "full". Library users pass
createServer(policy, { toolset: "full" }) / serverFactory(policy, { toolset });
parseToolset(value) maps the flag value (ERR_USAGE otherwise).
| tool | risk | input | output |
|---|---|---|---|
axiom_plan_validate | READ | { plan } | { ok, planDigest?, errors[] } |
axiom_plan_compile | ACT | { plan, store?: inline|cas, root? } | ManifestBundle (writes only under <root>/.axiom/ — CAS blobs and the stored manifest — when a root is given) |
axiom_manifest_verify | READ | { bundle, root? } | { ok, manifestDigest, canonical, signed, missing[], errors[], signatures?: { trustFile, keyids[], findings[], ok, code? } } — signatures only when root has .axiom/trust/keys.json; code is ERR_SIGNATURE_MISSING | ERR_SIGNATURE_INVALID when not ok |
axiom_check core | READ | { bundle, profile?, root? } | CheckReport (verdict: pass|fail|error, preImage: verified|drifted|unverified) |
axiom_check_start | READ | { bundle, profile?, root? } | { taskId, tool, status: "working", pollIntervalMs, ttlMs, elapsedMs } — same evaluation as axiom_check, returned immediately as a task so long guard.external suites (up to 15 min per guard) outlive the client's per-call timeout (S-406 / D-24) |
axiom_task_get | READ | { taskId } | descriptor + result: CheckReport once completed, or error: { code, message } once failed/cancelled; unknown/expired id → ERR_TASK_NOT_FOUND |
axiom_task_cancel | ACT | { taskId } | descriptor; kills every running guard tree, task ends cancelled with ERR_TASK_CANCELLED (idempotent on terminal tasks) |
axiom_plan_begin | ACT | Plan header: { name, intent, profile?, capabilities?, checks?, counter?, metadata? } | { sessionId, artifacts: 0, bytes: 0, limits: { maxArtifacts, maxBytes }, ttlMs } — opens a chunked plan session for Plans whose JSON would exceed the 4 MiB call cap |
axiom_plan_add | ACT | { sessionId, artifacts[] } (each call ≤ 4 MiB) | session descriptor; duplicate path across chunks → ERR_INVALID_PLAN, over budget (2000 artifacts / 64 MiB) or sealed → ERR_PLAN_SESSION_STATE |
axiom_plan_seal | ACT | { sessionId, store?: inline|cas, root? } | ManifestBundle — compiled by the same code path as axiom_plan_compile, so the digest equals a one-shot compile of the assembled Plan (property-tested); the session is consumed |
axiom_apply_dry_run core | READ | { bundle, root, profile? } | ApplyResult{mode:"dry-run", diff} |
axiom_apply core | SENSITIVE | { bundle, root, profile?, confirmDigest, mode?, branch?, commitMessage?, lockTimeoutMs? (0..600000), keepBackups? (0..1000) } | ApplyResult — error.details carries { otherDigest, paths } for ERR_CONFLICT and { holder, waitedMs, queuePosition } for ERR_LOCKED |
axiom_write core | SENSITIVE | { root, files[]: { path, content?, encoding?: utf8|base64, op?: write|delete, mode? }, intent?, profile?, confirmDigest?, lockTimeoutMs?, keepBackups? } | One-tool write (2.5.0): without confirmDigest → status: "preview" (dry-run + checks, diff, manifestDigest, next; nothing outside .axiom/ touched); the same files + confirmDigest → atomic apply like axiom_apply (journal, one-step axiom_rollback). A target changed since the preview → ERR_CONFIRM_DIGEST_MISMATCH. Prefer it for multi-file or sensitive changes |
axiom_rollback core | SENSITIVE | { root, manifestDigest } | { status:"rolled-back", phase, steps } |
axiom_manifest_diff | READ | { a: bundle|"sha256:…", b } | { added[], removed[], changed[] } |
axiom_axm_parse | READ | { source } (.axm text) | { plan?, diagnostics: [{ severity, code, message, range: { start: {line, column}, end } }] } |
axiom_roots_list | READ | {} | { roots: [{ path, writable, hasGit }] } |
axiom_status core | READ | { root? } | { root, lock: { held, holder?: { pid, host?, digest?, since?, stale }, queue[], intents[] }, interrupted[], corrupt[], lastApplied?, chain: { ok, entries, head, firstBad?, missingJournals } } — same as axiom status --json; never takes the lock |
axiom_repo_snapshot | READ | { root?, include?[], exclude?[], maxFiles? (20000, cap 50000), maxBytes? (64 MiB), respectGitignore? (true), withContentDigest? (true) } | RepoSnapshot { snapshotDigest, body: { files: [{ path, bytes, sha256?, mode, kind }], truncated, counts } } — sorted, no timestamps/absolute paths; .git/, .axiom/ always skipped; symlinks recorded, never followed (docs/guides/snapshot.md) |
Every tool carries MCP annotations (readOnlyHint, destructiveHint, idempotentHint, openWorldHint)
and an outputSchema; structuredContent is the full result, content[0].text a small summary (digest,
verdict, counts, first 20 findings). Errors come back as isError: true with { code, message, path? }
from the closed ERROR_CODES enum — a handler never throws. spec/tools.json is generated from the same
registry (pnpm build:spec) and guarded by a parity test. spec/codai-tools.json is the same registry in
codai's packages/agent-core/spec/tools-v2.json entry shape ({ name, risk, toolset, description, parameters }) —
see docs/integration/codai.md.
Resources: axiom://manifest/{sha}, axiom://report/{sha}, axiom://applied/{sha},
axiom://profile/{name}, axiom://schema/{Plan|Manifest|ManifestBundle|CheckReport|ApplyResult|Profile|Journal|RepoSnapshot},
axiom://emitters (template emitters loaded with --emitter <pkg> — empty by default since 3.0; install
@codai/axiom-emitters-web next to the repo and pass --emitter @codai/axiom-emitters-web, see docs/guides/emitters.md),
axiom://lock ({ roots: [{ root, held, holder?, queue[], intents[] }] } for every allowlisted root; read-only).
Trust model
- Roots are realpath'd at startup, must be directories, and the set is frozen. Requested roots are realpath'd too (case-insensitive containment on Windows); anything outside is a hard error.
axiom_applyrequiresconfirmDigest === bundle.manifestDigest— echo the digest you saw in dry-run. Pre-apply checks run against the profile (defaultdefault, or<root>/.axiom/profiles/<name>.json); a non-passverdict aborts withERR_CHECKS_FAILEDbefore any write.- Payloads over 4 MiB are rejected up front (
ERR_BUNDLE_TOO_LARGE). - Tasks and plan sessions live in the server process (shared by every connection/request the
process serves; never on disk). A restart forgets them; finished tasks stay pollable for 10 min, idle
sessions expire after 30 min; at most 8 tasks run concurrently (
ERR_EBUSYbeyond that). Stopping the server aborts every running task and kills its guard trees. .axiom/lockmakes apply single-writer per root; the journal makes it crash-safe and reversible.- stdout carries only JSON-RPC. Logs are JSON lines on stderr (
--log-level error|warn|info|debug, defaultwarn). - External guards (
guard.external) are off unless the process is started with--allow-guardsand the profile setsfacts.allowGuards: true. Relative commands must live under<root>/scripts/; absolute ones must be listed exactly via--guard-allowlist <abs>(repeatable). Guards are spawned with an args array (never a shell), a scrubbed environment, a wall-clock timeout, and must printGuardOutputJSON — seedocs/guides/checks.md. - Signed manifests (
docs/guides/signing.md): a root can pin Ed25519 public keys in.axiom/trust/keys.json; a profile withmanifest.requireSignedthen refuses unsigned, tampered or untrusted bundles, and withantiRollback: truerefuses anycounter ≤ .axiom/trust/state.json#lastCounter.axiom_applyadvances that state only onstatus: "applied". Private keys never enter the server: signing isaxiom signwithAXIOM_SIGNING_KEYor--key-file.
CLI
axiom mcp [--root <abs>]... [--allow-guards] [--guard-allowlist <abs>]... [--log-level warn]
[--http <host:port>] [--http-token-env AXIOM_HTTP_TOKEN] [--require-human]
axiom init [--root .] [--harness auto|copilot|claude|codex|vscode|cursor|windsurf|gemini|junie|cline] [--daemon]
[--profile-name <name>] [--force] [--upgrade] [--managed --out <dir>] [--json]
(.vscode/mcp.json + Pre/PostToolUse hooks + .axiom/profiles/<name>.json + .axiom/gate-profile.json + .gitignore; idempotent)
axiom doctor [--root .] [--json] (exit 0 = no failing check, 2 = at least one fail; warnings exit 0)
axiom stats [--root .] [--since 7d] [--json] (gate decisions, p50/p95, per harness; silent-gate alarms)
axiom undo --session <agent> [--root .] [--dry-run] [--force] [--since <iso>] (restore one session's writes; exit 1 = conflicts)
axiom import-guard <guard.ps1> [-o profile.json] (PowerShell guard tables → gate profile "commands")
axiom status [--root .] [--json] (lock holder/queue/intents, interrupted journals, last applied, chain)
axiom log [--root .] [--limit 20] [--json] (journal chain, newest first)
axiom compile <plan.json|plan.yaml|plan.axm> [-o out.json] [--store cas --root .] [--allow-net [--net-allow host[,host]]] [--allow-file]
axiom verify <bundle.json> [--root .] (--root: also verify signatures against .axiom/trust/keys.json)
axiom verify --journal --root . (journal hash chain; broken → exit 1 + code ERR_JOURNAL_CHAIN)
axiom verify <bundle.json> --tree <root> [--pre] [--attest out.intoto.json] (tree matches manifest? docs/guides/verify-tree.md)
axiom check <bundle.json> --root . [--profile p] [--json | --format text|json|sarif] [--allow-guards] [--guard-allowlist <abs>]...
axiom apply <bundle.json> --root . [--dry-run] [--profile p] [--confirm <digest>] [--lock-timeout <ms>] [--keep-backups <n>]
[--pr [--branch b] [--message m]] [--allow-guards] [--guard-allowlist <abs>]...
axiom rollback <digest> --root .
axiom gc --root . [--dry-run] [--older-than 30d] [--keep all-manifests|journal] [--undo-ttl 7d] (CAS garbage collection; CLI only, no MCP tool)
axiom diff <a.json> <b.json>
axiom schema <Plan|Manifest|ManifestBundle|CheckReport|ApplyResult|Profile|Journal|RepoSnapshot>
axiom emitters [--json]
axiom keygen [--out <dir>] [--name <label>] (ed25519; private key → <dir>/axiom-signing-<id>.key 0600, public entry → stdout)
axiom sign <bundle.json> [--key-file <path>] [-o out.json] [--root-id <id>] (key from --key-file or $AXIOM_SIGNING_KEY; --root-id = root-bound envelope)
axiom sign <bundle.json> --keyless [--bound --root . | --root-id <id>] [-o out.json]
(Sigstore via ambient OIDC → bundle.keylessSignatures; no OIDC → exit 1 ERR_KEYLESS_UNAVAILABLE)
axiom trust add <pub.json> --root . | remove <keyid> --root . | list --root . | root-id [<id>|--clear] --root .
axiom gate --stdin [--post] [--format <harness>] [--root <dir>] [--profile <file>] [--fail-open] [--no-shell-scan] [--no-root-discovery]
[--track-reads] [--deny-exit 0|2] [--log-level warn]
axiom serve --hook --root <abs>... [--host 127.0.0.1] [--port 47861] [--token-env <NAME>] [--otel]
axiom migrate v1 <manifest.json> [-o plan.json] [--profile default] [--cas <root>] [--content <dir>] [--overwrite]
(v1 manifest → v2 Plan, lazy chunk; exit 1 = migrated with warnings — docs/guides/migrate.md)
axiom snapshot --root . [-o snap.json] [--include <glob>]... [--exclude <glob>]... [--max-files n] [--max-bytes n] [--no-gitignore] [--no-digest]
axiom snapshot-diff <a.json> <b.json> (RepoSnapshot → { added, removed, changed })
Exit codes: 0 ok · 1 verdict fail / apply failed / journal chain broken / keyless unavailable · 2 usage or error
(doctor: 2 = a check failed). Non-mcp verbs print JSON to stdout (init, doctor, status, log, check, emitters: human lines unless --json).
ref sources ({ type: "ref", uri, digest }) are offline by default: a digest already in
<root>/.axiom/cas resolves without network, anything else is ERR_NET_DISABLED. --allow-net
fetches https: only (no redirects, 30 s timeout, 32 MiB cap), optionally restricted to
--net-allow hosts (*.example.com wildcards), verifies the pinned digest and stores the blob in
the CAS — a mismatch stores nothing (ERR_DIGEST_MISMATCH). apply never fetches. The MCP
axiom_plan_compile tool has no network switch. See docs/reference/plan-format.md
and docs/concepts/cas.md.
Hook mode — axiom gate --stdin
A PreToolUse hook for Claude Code, Copilot CLI and VS Code agent hooks. It reads one harness
payload from stdin (both {tool_name, tool_input, cwd} and {toolName, toolArgs, cwd} casings;
toolArgs may be a JSON string), extracts the write target(s) of Write|Edit|MultiEdit|NotebookEdit,
create_file|replace_string_in_file|insert_edit_into_file|apply_patch|multi_replace_string_in_file|edit_notebook_file
and generic write|edit, scans shell tools (Bash, run_in_terminal, …) for write
primitives (>, >>, tee, rm, mv, cp, sed -i, git checkout|reset|clean, PowerShell
Set-Content/Remove-Item, … — a heuristic, documented in docs/getting-started/hooks.md), and runs only the
fast predicates: containment + RelPath rules (.., CON, NTFS ADS → ERR_CONTAINMENT /
ERR_PATH_*), path.deny, path.allow, content.noSecrets and content.maxBytes on the new
content when the payload carries it. Fail-closed (D-18): a non-answer is a deny.
| outcome | exit | stdout | stderr |
|---|---|---|---|
| allow / non-write, non-shell tool | 0 | — | — |
| deny | 0 (2 with --deny-exit 2) | one object: hookSpecificOutput{…} (Claude) + flat permissionDecision/permissionDecisionReason (Copilot) + axiom{verdict, code, path, toolClass, standard:"owasp-acs/0.1"} | AXIOM GATE DENY <code>: <reason> (<relpath>) |
| write tool with no recognised path key | as deny | deny JSON | AXIOM GATE DENY ERR_UNSUPPORTED_OP: … |
| malformed payload, stdin timeout (2 s), internal error | as deny (fail closed) | deny JSON | AXIOM GATE DENY ERR_INTERNAL: … (fail-closed; pass --fail-open to allow) |
same, with --fail-open | 0 | — | AXIOM GATE WARN: … — failing open (--fail-open) |
A deny exits 0 since 2.4.1 (D-35): the npm axiom.ps1 shim on Windows rewrites 2 to 1, and
VS Code treats 1 as a warning. Every JSON-aware harness honours the document on exit 0; use
--deny-exit 2 for exit-code-only harnesses (Windsurf). Read-only tools are answered before the
gate chunk loads. A target in a sibling worktree of the same repository (same git common dir) is
judged against that worktree's root and gate profile instead of ERR_CONTAINMENT.
Root = payload cwd, walked up to the nearest .git / repository .axiom/ ancestor (never the
home dir; --no-root-discovery disables), else --root, else the process cwd (the hook is the
one place where cwd is acceptable: the harness spawns the hook in the project directory and owns
that value). Relative targets stay relative to cwd.
Profile = --profile <file> → <root>/.axiom/gate-profile.json → ~/.axiom/gate-profile.json →
built-in { deny: [".git/**", ".axiom/**", "**/*.lock", "pnpm-lock.yaml", ".env", ".env.*", "**/node_modules/**"], noSecrets: true, pii: false }.
Schema: { deny: string[], allow?: string[], noSecrets: boolean, pii: boolean, maxBytes?: number } (strict).
pii: true additionally scans for personal data (cnp, email, phoneRo, card) — off by default
since S-414 (maintainer e-mails in pyproject.toml denied real edits; see docs/guides/checks.md).
gate is a separate lazy chunk (dist/gate-lazy.js, no MCP SDK): in-process p95 ≈ 5 ms per payload,
end-to-end ≈ 150–200 ms including node startup; check-gate-latency guards p95 ≤ 250 ms.
Wiring for each harness is in docs/getting-started/hooks.md.
Comments
New to MCP?
From what the protocol is to installing your first server, these answers cover most first questions.
An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.