ambara 0.6.0

Reviewable project context for AI coding agents

Ambara

AI coding agents forget everything between sessions. Ambara fixes that.

Start Here

Want to see the product before wiring it into a real project?

pip install ambara
ambara demo
cd ambara-demo
./run-proof.sh

That local proof gives you a knowledge repo, a wired project repo, a fixture signal source, a static HTML view, and a runnable script that proves startup, reads, writes, review branches, capture, and value receipts without needing a network connection.

When you are ready to connect a real project:

Get Started

pip install ambara
cd your-project
ambara onboard

Restart your coding agent. Done.

What You Say

You don't learn an API. You just talk to your agent like you already do:

"Save where I left off."
→ Saved. Next session picks it up.

"We decided to use gRPC for internal services. Record that."
→ Saved as a decision. Goes through the reviewed path.

"What did we decide about the auth migration?"
→ Agent checks the context repo and tells you.

"Update the login flow to use the new token endpoint."
→ Agent checks for relevant context before working.

What It Does

Ambara gives your coding agents shared context that survives across sessions, repos, and people. You record a decision or hand off work, and the next agent starts with that context before it touches code.

Ambara also shows value as restraint: startup/context packets include a compact receipt showing the small Ambara payload, full bodies deferred to explicit reads, and estimated tokens avoided versus a full eligible context dump. A local value ledger can be enabled for longer-running proof with ambara value.

Under the hood, context lives in a git repository of Markdown files with YAML frontmatter. Ambara indexes, ranks, and serves it through MCP (the protocol that lets AI coding tools connect to external tools) or through a plain context file any agent can read. You don't need to know any of that to use it.

Who It's For

Use case	What Ambara gives you
Solo developers	Continuity across repos, days, and multiple agents. Future you and the next agent see the handoffs and decisions that matter.
Small teams	Shared work context and decisions without turning every useful note into a meeting.
Growing orgs	Reviewable agent guidance through review branches and requests, scoped by project, with lifecycle controls.

Why Not Built-In Agent Memory?

Built-in agent memory is useful for personal preferences and tool behavior. Ambara is for project facts that need provenance: decisions, workflows, handoffs, constraints, and status that should survive tool changes and team changes.

Ambara keeps those facts in git, makes durable updates reviewable, and requires explicit writes. It is not a replacement for a model or client's private memory; it is the shared project context layer beside it.

How You Verify It

Ambara is strongest when the human can inspect what happened directly.

• ambara demo gives you a no-network proof path before real setup
• ambara startup can write a plain context file that any agent can read
• ambara html renders a static context browser for direct human inspection
• the knowledge repo stays as normal Markdown in git, so review and history stay visible outside any one agent client

If Ambara only creates a review branch and does not open a review request, it now says that explicitly. It does not claim an opened review request when only a branch or compare link exists.

Setup

Ambara requires Python 3.12+.

Solo / Local

pip install ambara
cd your-project
ambara onboard --yes

Creates a local context repo, adds agent instructions to your CLAUDE.md, configures the MCP server, and enables handoffs. Restart your coding agent and you're ready.

Team

ambara onboard --remote <git-remote-url>

Team members connect with the same command using the same remote URL.

Example remotes:

• GitHub: [e-mail]:your-org/team-context.git
• GitLab: [e-mail]:your-group/team-context.git

Verify

ambara doctor --check-startup

For troubleshooting, see Install And Troubleshooting.

Try The Local Demo

ambara demo
cd ambara-demo
./run-proof.sh

Creates a no-network demo under ./ambara-demo with a local knowledge repo, a wired project repo, a runnable proof script, a value ledger path, a fixture project signal, and a static HTML inspection path.

The proof script validates the knowledge repo, checks startup, compiles compact context, proves restricted entries stay out of startup and static HTML, reads a selected source body, records active context, proposes durable context on a review branch, imports docs through a review branch, queues and flushes ambient capture, checks local posture, and prints value accounting.

CI Capture

Ambara also supports commit-driven capture in CI. The shared capture entrypoint bootstraps the local knowledge repo from the declared remote, applies forge-aware authenticated remotes when a token is provided, fetches the main branch, and then evaluates commit capture with the same content-policy gate used by interactive writes.

For GitHub, use the bundled composite action.

For GitLab CI, install Ambara in the job and run the same entrypoint directly:

ambara_capture:
  image: python:3.12
  variables:
    AMBARA_KNOWLEDGE_REMOTE: https://gitlab.com/your-org/team-context.git
    AMBARA_KNOWLEDGE_TOKEN: $AMBARA_KNOWLEDGE_TOKEN
    AMBARA_MAIN_BRANCH: main
  script:
    - pip install ambara
    - python -m ambara.action.capture

This is for explicit ambara: commit messages and related workflow capture. It is not a generic background collector.

What a Saved Entry Looks Like

When you say "record this as a decision," Ambara creates a Markdown file:

---
type: decision
title: Use gRPC for internal services
status: active
projects: [api-platform]
created: 2026-01-15T10:00:00Z
---

REST endpoints between auth and billing are creating versioning friction.
Use gRPC for all new internal service communication. REST stays for public APIs.

Decisions and workflows go through reviewed durable write paths. Handoffs are short-lived and expire automatically.

Privacy

Ambara does not log what agents read, track who's online, or create activity feeds. It only stores what you explicitly ask it to save. There is no collector for query text, read patterns, editor activity, or participation data. The system can't track what it was never built to capture.

Value receipts and the optional local value ledger keep the same boundary: they record aggregate token estimates and omission/suppression counts only. They do not store query text, response text, body snippets, caller identity, or durable per-entry read receipts. Reported savings are Ambara app-payload estimates, not total client/session/cache billing.

Security, privacy, and secret scanning are release gates for this product, not optional hardening. See Security Policy.

What Ambara Does Not Claim

Ambara is not a secrets manager, observability product, agent runtime, or total LLM cost meter. It reports Ambara app-payload estimates only. It does not prove who read what, infer presence, bypass git review, or replace human review for durable project knowledge.

Claude Code setup and the plain file adapter are first-class for this release. Other MCP clients should work through the protocol, but client-specific support is only claimed after that client has been smoked.

---

Reference

Everything below here is reference material. You don't need it to get started.

How It Works

              ┌──────────────────┐
              │  Context Repo    │  Reviewed Markdown + YAML frontmatter
              │  (normal git)    │  in git — reviews, history, provenance
              └────────┬─────────┘
                       │
              ┌────────▼─────────┐
              │  Ambara Server   │  Indexes, ranks, filters,
              │                  │  enforces privacy + secret scanning
              └────────┬─────────┘
                       │
          ┌────────────┼────────────┐
          │            │            │
    ┌─────▼─────┐ ┌───▼───┐ ┌─────▼──────┐
    │Claude Code│ │Cursor │ │ Any MCP    │
    │           │ │       │ │ client     │
    └───────────┘ └───────┘ └────────────┘

When a developer starts a task, the agent calls startup() to get relevant context, then works normally. When the developer says "save this" or "record that decision," the agent handles the filing.

What Agents See

Tool	Purpose
startup()	Get relevant context before working (recommended entry point)
context()	Relevant decisions, project state, and project signals
sync()	Refresh local index from the context repo
read() / read_many()	Full text for entries the agent directly relies on
query()	Deep-dive on a specific topic
search()	Keyword search across all entries
record()	Save a handoff, project update, or active note
propose()	Save a decision or workflow for team review via a review branch/request
hygiene()	Lifecycle advisories for stale context
list_entries()	Browse entries with optional filters
expire()	Mark an entry as expired
feedback(entry_id, signal)	Mark context as useful, noisy, stale, duplicate, or pinned locally
explain()	Print what Ambara does and does not track
status()	Show connection health and sync state
validate()	Check entries for schema or constraint violations

Startup packets include a compact value receipt. Compiler packets include cost_accounting detail for debugging and proof. Full bodies stay deferred until read() or read_many() is explicitly needed.

Client Support

Ambara has first-class generated setup for Claude Code and a plain file adapter for agents that cannot use MCP. Other MCP clients can connect to ambara serve with the generated config, but client-specific docs should be treated as protocol-compatible until manually smoked.

Support Matrix

Surface	Current truth
Review backends	GitHub pull requests and GitLab merge requests are first-class.
Other git hosts	Compatible through branch-only fallback until a forge-specific review adapter exists.
CI capture	GitHub Actions and GitLab CI are first-class. Other CI runners can call python -m ambara.action.capture directly.
Agent setup	Claude Code and the plain file adapter are first-class. Other MCP clients are protocol-compatible until manually smoked.
Signals	GitHub review signals, GitLab review signals, and Jira issues are supported.

Entry Types

Type	What it's for
active	Short-lived handoffs. "Left off at the token endpoint." Expires automatically.
project	Current project state. "Auth migration is in flight."
decision	Reviewed team decision. "Use gRPC for internal services." Goes through the reviewed path.
workflow	Recurring team process. "Run validation before review." Goes through the reviewed path.
principle	Durable standard. "Restricted data stays out of context entries."

Secret Scanning

All writes are checked before reaching the context repo:

• Credentials and secrets are blocked (API keys, tokens, passwords)
• PII and personnel content are flagged for review
• Signal text from review signals and issues is sanitized before surfacing

Project Signals

Ambara reads labeled GitHub pull requests, GitLab merge requests, and Jira issues from configured sources and surfaces them as project context: continue, reuse, revise, pause, handoff, or clarify. GitHub, GitLab, and Jira are supported.

AMBARA_PR_SIGNAL_REPOS=your-org/repo1,your-org/repo2
AMBARA_SIGNAL_PROVIDER=gitlab
AMBARA_GITLAB_BASE_URL=https://gitlab.example.com

AMBARA_PR_SIGNAL_REPOS=your-org/repo1,your-org/repo2
AMBARA_SIGNAL_PROVIDER=github

AMBARA_SIGNAL_PROVIDER=jira
AMBARA_JIRA_BASE_URL=https://jira.example.com
AMBARA_JIRA_PROJECTS=PROJ1,PROJ2

Curation

Context needs maintenance. Ambara generates cleanup review branches that expire stale handoffs and flag durable entries for revalidation. Agents also get quiet advisories inside context() so they know not to rely on stale entries.

CLI Reference

Command	Purpose
ambara serve	Start the MCP server over stdio
ambara onboard	Guided first-time setup
ambara demo	Create a local no-network proof project
ambara doctor	Check configuration and connectivity
ambara startup	Write compact startup context as JSON or Markdown
ambara validate	Validate context entries
ambara import-docs	Import ADRs, RFCs, or selected docs
ambara import-context	Import a packaged context pack
ambara install-hook	Install ambient commit capture hooks
ambara posture-check	Check local agent-safe-start guardrails
ambara install-claude-md	Add or update agent instructions in CLAUDE.md
ambara explain	Print the product boundary
ambara value	Show local app-payload savings estimates
ambara feedback	Record local relevance feedback for an entry
ambara html	Generate a static context browser
ambara read-model	Build or query the local SQLite read model

Run ambara --help for the full list.

Configuration

Variable	Default	Purpose
AMBARA_KNOWLEDGE_REMOTE	-	Git remote for the context repo
AMBARA_ACTIVE_REFS_ENABLED	true	Enable session handoffs
AMBARA_ACTIVE_REF_TTL_HOURS	72	How long handoffs stay active
AMBARA_SIGNAL_PROVIDER	github	Signal source: github, gitlab, or jira
AMBARA_PR_SIGNAL_REPOS	-	Repos for GitHub PR or GitLab MR signals
AMBARA_GITLAB_BASE_URL	-	Base URL for GitLab.com or self-hosted GitLab
AMBARA_JIRA_BASE_URL	-	Jira instance URL
AMBARA_JIRA_PROJECTS	-	Jira project keys
AMBARA_CONTEXT_BUDGET_MODE	balanced	quiet, balanced, or deep context packing
AMBARA_VALUE_LEDGER_PATH	-	Enables local JSONL value ledger at this path
AMBARA_VALUE_LEDGER_ENABLED	path-dependent	Force local value ledger on/off

Documentation

Guide	Audience
Install And Troubleshooting	Setup, configuration, and common issues
Release Proof	Local release-candidate verification
v0.6.0 Release Notes	Multi-forge foundation and first-class GitLab release notes
v0.5.1 Release Notes	Trust-and-reliability follow-up release notes
v0.5.0 Release Notes	Release notes, launch copy, and post-publish smoke
Contributing	Development setup, testing, code style
Security	Vulnerability reporting and supported versions

Development

git clone https://github.com/ostinato-forge/ambara.git
cd ambara
python3 -m venv .venv && . .venv/bin/activate
pip install -e ".[dev]"
pytest -q && ruff check . && mypy src

License

Apache 2.0 — see LICENSE.

Contributions are certified with the Developer Certificate of Origin 1.1. Sign commits with git commit -s; see DCO and CONTRIBUTING.md.

Disclaimer

This is a personal project. It does not represent the views, policies, or endorsement of my employer.