Get help with Talon, report issues, or reach out to us. Most issues land in the docs or the FAQ below. For everything else, open a ticket from the CLI with talon support new — your session metadata attaches automatically, and you can follow it with talon support list and talon support view. Reply within one business day during alpha. No account yet? Use the contact form.
Four commands, under two minutes — from pipx install to your first Claude Code session with Talon active.
How Talon works and Research cover the scoring mechanism, components, and corpus methodology.
Run talon support new in the project you hit the problem in — session details attach automatically. No account yet? Use the contact form.
Comprehensive answers to common questions about Talon's capabilities, security model, and deployment. If yours isn't covered, use the contact form.
Talon is a context intelligence and steering engine that sits between your codebase and your AI coding agent. It predicts what each task needs, serves the model exactly the right context, and steers the agent clear of the loops that burn turns and tokens — so it starts each turn with precise structural knowledge instead of discovering files by trial and error. Measured under a sealed benchmark: 0.41–0.63× cost on Python and 0.605× on Java at resolve parity, and 2× fewer broken tests — full per-task tables, including where we lose, on /research.
Claude Code today (both Max subscriptions and direct API usage). Integration is via the Model Context Protocol (MCP), so any MCP-compatible AI coding tool can work with Talon. Additional integrations are on the roadmap, and the protocol-based architecture means new ones don't require core changes.
Four file types today: Python (.py), TypeScript and TSX (.ts, .tsx), and Java (.java). The language-aware parser uses tree-sitter AST analysis to extract functions, classes, imports, call chains, and parameter relationships. Multi-language repositories are supported — the graph spans language boundaries through import and call-chain analysis. Other languages, including plain JavaScript, are not parsed yet; they land as the parser adapters are written.
By solving discovery — it puts the right context in front of the agent instead of letting it search:
1. It predicts what the task needs. The agent starts with the relevant files — including the structural context that grep and embeddings miss — rather than scanning whole directories to find them.
2. It serves that context sized to the task. Enough to start informed, not so much that the window fills with things the task never needed.
3. It steers clear of waste. Redundant re-reads, scattered edits, retry storms and rediscovery loops are caught in real time, so they don't inflate the window even when the starting context was good.
On large codebases, sealed benchmarks show 35–50% less agent spend at resolve parity, with 2× fewer broken tests; per-task results on /research.
Graph queries are served from a multi-layer cache, and the cache pre-warms data for files the task context suggests will be needed, so the first query is not slower than the ones after it. Indexing a new repository is a local parse plus an upload, so it scales with repository size; incremental re-indexing after a few file changes is much faster than a full pass. We have not published calibrated latency or indexing figures, so we are not quoting any here.
Yes. The graph database and caching layers handle repositories with tens of thousands of files and hundreds of thousands of symbols. Talon has been tested across 8 open-source Python projects spanning ~1.28M LOC total:
Our multi-prompt benchmark corpus draws 6 tasks from this set (SQLAlchemy, Airflow, Django, Celery); the rest are scale tests. For very large monorepos, the scoring system becomes more valuable — as the number of candidate files increases, the importance of accurate relevance filtering grows proportionally. Per-task results on /research.
Two different answers, and the distinction matters. The indexer runs locally: it performs AST analysis and call-graph extraction on-device and transmits only structural metadata — function names, file paths, and call relationships — over TLS, where it is stored encrypted and used to serve graph queries. It never uploads file contents, variable values, string literals, or comments. The proxy is a different matter: because Talon sits in the request path between your agent and the model, your prompts and the model's responses — including any source, diffs, or context they contain — transit our servers so Talon can inject context and measure cost. That traffic is processed only to deliver the service and is never used to train any model. The full picture, including hosting and retention, is on /privacy.
By default, anonymous usage telemetry: query counts, response latency, and error rates. This contains no code content, file paths, symbol names, or repository structure. Telemetry can be completely disabled via a single environment variable — TALON_TELEMETRY=0 — after which zero data is transmitted. We use aggregate telemetry to improve scoring accuracy and query performance — not to analyse your code. Full retention and processing detail on /privacy.
Talon is free for early users during the private alpha — no credit card required. Pricing tiers are being finalised for launch, with an early-user discount window; enterprise and self-host options are on the roadmap.
On large codebases, sealed benchmarks show 35–50% less agent spend at resolve parity, with sealed cells published for Python, Java and TypeScript. Net saving depends on session shape and codebase size, and below the size floor the saving thins or inverts — we will say so before you pay. Pricing is not yet set, so we are not publishing a net-of-fees figure. The research page has the full per-task tables, including the cell where we cost more, so you can judge the shape against workloads like yours.
If you have an alpha account, run talon support new — that's the canonical channel, and your session metadata attaches automatically. If you don't have an account yet, use the contact form with topic technical. The public issue tracker ships post-alpha. Bug reports with reproduction steps and environment details get the fastest response.
Five checks, in order:
1. Verify Talon is active. Check for graph query responses in your agent's tool-use log. If you don't see any Talon graph tool calls, the MCP server isn't connected — re-run talon init and restart Claude Code.
2. Re-index your repository. Run talon index from the terminal to trigger a fresh analysis (talon init only writes the MCP config — it does not build the graph). The indexer runs locally; only structural metadata is sent over TLS. A stale index misses recently added files and symbols, which reduces accuracy of context selection.
3. Check your session score. Run talon score to see this session's Talon score — a normalised health metric answering "is Talon controlling this session effectively?". It's computed from observable signals: tool calls per prompt, proportion of prompts routing through Talon's graph tools, context accumulation per prompt, and how often the agent falls back to direct file reads after graph context has already been served.
Score bands (calibrated against the alpha corpus):
4. Consider task complexity. Very small tasks may not see large reductions because the graph overhead doesn't fully amortise. Talon delivers the strongest results on moderate-to-complex tasks where context accumulation is most severe — see /research for per-task patterns including the two regressions where vanilla Claude Code already runs near-optimally.
If issues persist after the checks above, run talon support new — session details attach automatically and we'll investigate. No account yet? Use the contact form.
One number, and anything you want to add. This is the rating we read every morning during alpha — the same scale as talon feedback in the CLI, and it lands in the same place. Email is optional; leave it if you want a reply.
Thanks — that's recorded. If you left an email we may follow up.
Need a reply on something specific, or hit a bug? The contact form opens a ticket — this box is just the rating.
Run talon support new from your project — reply within one business day during alpha. Alpha bug reports jump the queue. No account yet? The contact form is the way in.